program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3605.1.1"}, {"coremlc-version", "3600.25.1"}})]
{
    func main<ios18>(tensor<fp32, [1, 128, 1501]> mel, tensor<int32, [1]> mel_length) {
            tensor<int32, [3]> var_2238_perm_0 = const()[name = string("op_2238_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")];
            tensor<int32, [1]> input_1_axes_0 = const()[name = string("input_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 128, 1501]> mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_8")];
            tensor<fp16, [1, 1501, 128]> var_2238_cast_fp16 = transpose(perm = var_2238_perm_0, x = mel_to_fp16)[name = string("transpose_484")];
            tensor<fp16, [1, 1, 1501, 128]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = var_2238_cast_fp16)[name = string("input_1_cast_fp16")];
            int32 var_2242 = const()[name = string("op_2242"), val = int32(1)];
            tensor<int32, [1]> n_1 = add(x = mel_length, y = var_2242)[name = string("n_1")];
            string x_1_pad_type_0 = const()[name = string("x_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_1_pad_0 = const()[name = string("x_1_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> x_1_strides_0 = const()[name = string("x_1_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> x_1_dilations_0 = const()[name = string("x_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_1_groups_0 = const()[name = string("x_1_groups_0"), val = int32(1)];
            tensor<fp16, [256, 1, 3, 3]> enc_subsampling_layers_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [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(2432))))[name = string("enc_subsampling_layers_0_weight_to_fp16_quantized")];
            tensor<fp16, [256]> enc_subsampling_layers_0_bias_to_fp16 = const()[name = string("enc_subsampling_layers_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3008)))];
            tensor<fp16, [1, 256, 751, 64]> x_1_cast_fp16 = conv(bias = enc_subsampling_layers_0_bias_to_fp16, dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = enc_subsampling_layers_0_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("x_1_cast_fp16")];
            int32 var_2257 = const()[name = string("op_2257"), val = int32(2)];
            tensor<int32, [1]> var_2258 = add(x = mel_length, y = var_2257)[name = string("op_2258")];
            int32 var_2260 = const()[name = string("op_2260"), val = int32(3)];
            tensor<int32, [1]> var_2261 = sub(x = var_2258, y = var_2260)[name = string("op_2261")];
            string cast_0_to_fp16_dtype_0 = const()[name = string("cast_0_to_fp16_dtype_0"), val = string("fp16")];
            fp16 cast_1_to_fp16 = const()[name = string("cast_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_2261_to_fp16 = cast(dtype = cast_0_to_fp16_dtype_0, x = var_2261)[name = string("cast_7")];
            tensor<fp16, [1]> var_2264_cast_fp16 = floor_div(x = var_2261_to_fp16, y = cast_1_to_fp16)[name = string("op_2264_cast_fp16")];
            fp16 var_2266_promoted_to_fp16 = const()[name = string("op_2266_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> length_1_cast_fp16 = add(x = var_2264_cast_fp16, y = var_2266_promoted_to_fp16)[name = string("length_1_cast_fp16")];
            int32 var_2269 = const()[name = string("op_2269"), val = int32(2)];
            tensor<int32, [1]> var_2270 = add(x = n_1, y = var_2269)[name = string("op_2270")];
            int32 var_2272 = const()[name = string("op_2272"), val = int32(3)];
            tensor<int32, [1]> var_2273 = sub(x = var_2270, y = var_2272)[name = string("op_2273")];
            string cast_2_to_fp16_dtype_0 = const()[name = string("cast_2_to_fp16_dtype_0"), val = string("fp16")];
            fp16 cast_3_to_fp16 = const()[name = string("cast_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_2273_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = var_2273)[name = string("cast_6")];
            tensor<fp16, [1]> var_2276_cast_fp16 = floor_div(x = var_2273_to_fp16, y = cast_3_to_fp16)[name = string("op_2276_cast_fp16")];
            fp16 var_2278_promoted_to_fp16 = const()[name = string("op_2278_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> n_3_cast_fp16 = add(x = var_2276_cast_fp16, y = var_2278_promoted_to_fp16)[name = string("n_3_cast_fp16")];
            tensor<int32, [2]> var_2290 = const()[name = string("op_2290"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_2291_cast_fp16 = reshape(shape = var_2290, x = length_1_cast_fp16)[name = string("op_2291_cast_fp16")];
            tensor<fp16, [1, 751]> var_2287_promoted_to_fp16 = const()[name = string("op_2287_promoted_to_fp16"), val = tensor<fp16, [1, 751]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3584)))];
            tensor<bool, [1, 751]> var_2292_cast_fp16 = less(x = var_2287_promoted_to_fp16, y = var_2291_cast_fp16)[name = string("op_2292_cast_fp16")];
            tensor<int32, [4]> var_2302 = const()[name = string("op_2302"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            string keep_1_to_fp16_dtype_0 = const()[name = string("keep_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 751]> var_2292_cast_fp16_to_fp16 = cast(dtype = keep_1_to_fp16_dtype_0, x = var_2292_cast_fp16)[name = string("cast_5")];
            tensor<fp16, [1, 1, 751, 1]> var_2303_cast_fp16 = reshape(shape = var_2302, x = var_2292_cast_fp16_to_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [1, 256, 751, 64]> input_3_cast_fp16 = mul(x = x_1_cast_fp16, y = var_2303_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1, 256, 751, 64]> input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
            string x_3_pad_type_0 = const()[name = string("x_3_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_3_pad_0 = const()[name = string("x_3_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> x_3_strides_0 = const()[name = string("x_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
            int32 x_3_groups_0 = const()[name = string("x_3_groups_0"), val = int32(256)];
            tensor<int32, [2]> x_3_dilations_0 = const()[name = string("x_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [256, 1, 3, 3]> enc_subsampling_layers_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5184))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7552))))[name = string("enc_subsampling_layers_2_weight_to_fp16_quantized")];
            tensor<fp16, [256]> enc_subsampling_layers_2_bias_to_fp16 = const()[name = string("enc_subsampling_layers_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8128)))];
            tensor<fp16, [1, 256, 376, 32]> x_3_cast_fp16 = conv(bias = enc_subsampling_layers_2_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 = enc_subsampling_layers_2_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_2320_promoted_to_fp16 = const()[name = string("op_2320_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_2321_cast_fp16 = add(x = length_1_cast_fp16, y = var_2320_promoted_to_fp16)[name = string("op_2321_cast_fp16")];
            fp16 var_2323_promoted_to_fp16 = const()[name = string("op_2323_promoted_to_fp16"), val = fp16(0x1.8p+1)];
            tensor<fp16, [1]> var_2324_cast_fp16 = sub(x = var_2321_cast_fp16, y = var_2323_promoted_to_fp16)[name = string("op_2324_cast_fp16")];
            fp16 cast_7_to_fp16 = const()[name = string("cast_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_2327_cast_fp16 = floor_div(x = var_2324_cast_fp16, y = cast_7_to_fp16)[name = string("op_2327_cast_fp16")];
            fp16 var_2329_promoted_to_fp16 = const()[name = string("op_2329_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> length_3_cast_fp16 = add(x = var_2327_cast_fp16, y = var_2329_promoted_to_fp16)[name = string("length_3_cast_fp16")];
            fp16 var_2332_promoted_to_fp16 = const()[name = string("op_2332_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_2333_cast_fp16 = add(x = n_3_cast_fp16, y = var_2332_promoted_to_fp16)[name = string("op_2333_cast_fp16")];
            fp16 var_2335_promoted_to_fp16 = const()[name = string("op_2335_promoted_to_fp16"), val = fp16(0x1.8p+1)];
            tensor<fp16, [1]> var_2336_cast_fp16 = sub(x = var_2333_cast_fp16, y = var_2335_promoted_to_fp16)[name = string("op_2336_cast_fp16")];
            fp16 cast_9_to_fp16 = const()[name = string("cast_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_2339_cast_fp16 = floor_div(x = var_2336_cast_fp16, y = cast_9_to_fp16)[name = string("op_2339_cast_fp16")];
            fp16 var_2341_promoted_to_fp16 = const()[name = string("op_2341_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> n_cast_fp16 = add(x = var_2339_cast_fp16, y = var_2341_promoted_to_fp16)[name = string("n_cast_fp16")];
            tensor<int32, [2]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_2354_cast_fp16 = reshape(shape = var_2353, x = length_3_cast_fp16)[name = string("op_2354_cast_fp16")];
            tensor<fp16, [1, 376]> var_2350_promoted_to_fp16 = const()[name = string("op_2350_promoted_to_fp16"), val = tensor<fp16, [1, 376]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8704)))];
            tensor<bool, [1, 376]> var_2355_cast_fp16 = less(x = var_2350_promoted_to_fp16, y = var_2354_cast_fp16)[name = string("op_2355_cast_fp16")];
            tensor<int32, [4]> var_2365 = const()[name = string("op_2365"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            string keep_3_to_fp16_dtype_0 = const()[name = string("keep_3_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 376]> var_2355_cast_fp16_to_fp16 = cast(dtype = keep_3_to_fp16_dtype_0, x = var_2355_cast_fp16)[name = string("cast_4")];
            tensor<fp16, [1, 1, 376, 1]> var_2366_cast_fp16 = reshape(shape = var_2365, x = var_2355_cast_fp16_to_fp16)[name = string("op_2366_cast_fp16")];
            tensor<fp16, [1, 256, 376, 32]> input_7_cast_fp16 = mul(x = x_3_cast_fp16, y = var_2366_cast_fp16)[name = string("input_7_cast_fp16")];
            string x_5_pad_type_0 = const()[name = string("x_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_5_strides_0 = const()[name = string("x_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_5_pad_0 = const()[name = string("x_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_5_dilations_0 = const()[name = string("x_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_5_groups_0 = const()[name = string("x_5_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1, 1]> enc_subsampling_layers_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9536))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75136))))[name = string("enc_subsampling_layers_3_weight_to_fp16_quantized")];
            tensor<fp16, [256]> enc_subsampling_layers_3_bias_to_fp16 = const()[name = string("enc_subsampling_layers_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75712)))];
            tensor<fp16, [1, 256, 376, 32]> x_5_cast_fp16 = conv(bias = enc_subsampling_layers_3_bias_to_fp16, dilations = x_5_dilations_0, groups = x_5_groups_0, pad = x_5_pad_0, pad_type = x_5_pad_type_0, strides = x_5_strides_0, weight = enc_subsampling_layers_3_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 256, 376, 32]> input_9_cast_fp16 = mul(x = x_5_cast_fp16, y = var_2366_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1, 256, 376, 32]> input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
            string x_7_pad_type_0 = const()[name = string("x_7_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_7_pad_0 = const()[name = string("x_7_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> x_7_strides_0 = const()[name = string("x_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
            int32 x_7_groups_0 = const()[name = string("x_7_groups_0"), val = int32(256)];
            tensor<int32, [2]> x_7_dilations_0 = const()[name = string("x_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [256, 1, 3, 3]> enc_subsampling_layers_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76288))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78656))))[name = string("enc_subsampling_layers_5_weight_to_fp16_quantized")];
            tensor<fp16, [256]> enc_subsampling_layers_5_bias_to_fp16 = const()[name = string("enc_subsampling_layers_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79232)))];
            tensor<fp16, [1, 256, 188, 16]> x_7_cast_fp16 = conv(bias = enc_subsampling_layers_5_bias_to_fp16, dilations = x_7_dilations_0, groups = x_7_groups_0, pad = x_7_pad_0, pad_type = x_7_pad_type_0, strides = x_7_strides_0, weight = enc_subsampling_layers_5_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_2419_promoted_to_fp16 = const()[name = string("op_2419_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_2420_cast_fp16 = add(x = length_3_cast_fp16, y = var_2419_promoted_to_fp16)[name = string("op_2420_cast_fp16")];
            fp16 var_2422_promoted_to_fp16 = const()[name = string("op_2422_promoted_to_fp16"), val = fp16(0x1.8p+1)];
            tensor<fp16, [1]> var_2423_cast_fp16 = sub(x = var_2420_cast_fp16, y = var_2422_promoted_to_fp16)[name = string("op_2423_cast_fp16")];
            fp16 cast_15_to_fp16 = const()[name = string("cast_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_2426_cast_fp16 = floor_div(x = var_2423_cast_fp16, y = cast_15_to_fp16)[name = string("op_2426_cast_fp16")];
            fp16 var_2428_promoted_to_fp16 = const()[name = string("op_2428_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> length_cast_fp16 = add(x = var_2426_cast_fp16, y = var_2428_promoted_to_fp16)[name = string("length_cast_fp16")];
            fp16 var_2431_promoted_to_fp16 = const()[name = string("op_2431_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_2432_cast_fp16 = add(x = n_cast_fp16, y = var_2431_promoted_to_fp16)[name = string("op_2432_cast_fp16")];
            fp16 var_2434_promoted_to_fp16 = const()[name = string("op_2434_promoted_to_fp16"), val = fp16(0x1.8p+1)];
            tensor<fp16, [1]> var_2435_cast_fp16 = sub(x = var_2432_cast_fp16, y = var_2434_promoted_to_fp16)[name = string("op_2435_cast_fp16")];
            fp16 cast_17_to_fp16 = const()[name = string("cast_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1]> var_2438_cast_fp16 = floor_div(x = var_2435_cast_fp16, y = cast_17_to_fp16)[name = string("op_2438_cast_fp16")];
            fp16 var_2440_promoted_to_fp16 = const()[name = string("op_2440_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> tensor_length_cast_fp16 = add(x = var_2438_cast_fp16, y = var_2440_promoted_to_fp16)[name = string("tensor_length_cast_fp16")];
            tensor<int32, [2]> var_2452 = const()[name = string("op_2452"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_2453_cast_fp16 = reshape(shape = var_2452, x = length_cast_fp16)[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 188]> var_2449_promoted_to_fp16 = const()[name = string("op_2449_promoted_to_fp16"), val = tensor<fp16, [1, 188]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79808)))];
            tensor<bool, [1, 188]> var_2454_cast_fp16 = less(x = var_2449_promoted_to_fp16, y = var_2453_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<int32, [4]> var_2464 = const()[name = string("op_2464"), val = tensor<int32, [4]>([1, 1, -1, 1])];
            string keep_7_to_fp16_dtype_0 = const()[name = string("keep_7_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 188]> var_2454_cast_fp16_to_fp16 = cast(dtype = keep_7_to_fp16_dtype_0, x = var_2454_cast_fp16)[name = string("cast_3")];
            tensor<fp16, [1, 1, 188, 1]> var_2465_cast_fp16 = reshape(shape = var_2464, x = var_2454_cast_fp16_to_fp16)[name = string("op_2465_cast_fp16")];
            tensor<fp16, [1, 256, 188, 16]> input_13_cast_fp16 = mul(x = x_7_cast_fp16, y = var_2465_cast_fp16)[name = string("input_13_cast_fp16")];
            string x_9_pad_type_0 = const()[name = string("x_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_9_strides_0 = const()[name = string("x_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_9_dilations_0 = const()[name = string("x_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_9_groups_0 = const()[name = string("x_9_groups_0"), val = int32(1)];
            tensor<fp16, [256, 256, 1, 1]> enc_subsampling_layers_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80256))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145856))))[name = string("enc_subsampling_layers_6_weight_to_fp16_quantized")];
            tensor<fp16, [256]> enc_subsampling_layers_6_bias_to_fp16 = const()[name = string("enc_subsampling_layers_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146432)))];
            tensor<fp16, [1, 256, 188, 16]> x_9_cast_fp16 = conv(bias = enc_subsampling_layers_6_bias_to_fp16, dilations = x_9_dilations_0, groups = x_9_groups_0, pad = x_9_pad_0, pad_type = x_9_pad_type_0, strides = x_9_strides_0, weight = enc_subsampling_layers_6_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1, 256, 188, 16]> input_15_cast_fp16 = mul(x = x_9_cast_fp16, y = var_2465_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [1, 256, 188, 16]> x_11_cast_fp16 = relu(x = input_15_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<int32, [4]> var_2506_perm_0 = const()[name = string("op_2506_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2510 = const()[name = string("op_2510"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 256, 16]> var_2506_cast_fp16 = transpose(perm = var_2506_perm_0, x = x_11_cast_fp16)[name = string("transpose_483")];
            tensor<fp16, [1, 188, 4096]> input_17_cast_fp16 = reshape(shape = var_2510, x = var_2506_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [1024, 4096]> enc_subsampling_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4341376))))[name = string("enc_subsampling_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> enc_subsampling_linear_bias_to_fp16 = const()[name = string("enc_subsampling_linear_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4343488)))];
            tensor<fp16, [1, 188, 1024]> linear_0_cast_fp16 = linear(bias = enc_subsampling_linear_bias_to_fp16, weight = enc_subsampling_linear_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [2]> var_2525 = const()[name = string("op_2525"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [1, 1]> var_2526_cast_fp16 = reshape(shape = var_2525, x = tensor_length_cast_fp16)[name = string("op_2526_cast_fp16")];
            tensor<bool, [1, 188]> var_2527_cast_fp16 = less(x = var_2449_promoted_to_fp16, y = var_2526_cast_fp16)[name = string("op_2527_cast_fp16")];
            tensor<int32, [1]> valid_3_axes_0 = const()[name = string("valid_3_axes_0"), val = tensor<int32, [1]>([-1])];
            string valid_1_to_fp16_dtype_0 = const()[name = string("valid_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 188]> var_2527_cast_fp16_to_fp16 = cast(dtype = valid_1_to_fp16_dtype_0, x = var_2527_cast_fp16)[name = string("cast_2")];
            tensor<fp16, [1, 188, 1]> valid_3_cast_fp16 = expand_dims(axes = valid_3_axes_0, x = var_2527_cast_fp16_to_fp16)[name = string("valid_3_cast_fp16")];
            tensor<int32, [1]> input_21_axes_0 = const()[name = string("input_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> insert_ln_weight_to_fp16 = const()[name = string("insert_ln_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4345600)))];
            tensor<fp16, [1024]> insert_ln_bias_to_fp16 = const()[name = string("insert_ln_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4347712)))];
            fp16 var_2542_to_fp16 = const()[name = string("op_2542_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = insert_ln_bias_to_fp16, epsilon = var_2542_to_fp16, gamma = insert_ln_weight_to_fp16, x = linear_0_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> insert_down_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4349824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5398464))))[name = string("insert_down_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> insert_down_bias_to_fp16 = const()[name = string("insert_down_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5400576)))];
            tensor<fp16, [1, 188, 1024]> linear_1_cast_fp16 = linear(bias = insert_down_bias_to_fp16, weight = insert_down_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_2555_cast_fp16 = mul(x = linear_1_cast_fp16, y = valid_3_cast_fp16)[name = string("op_2555_cast_fp16")];
            tensor<int32, [3]> input_23_perm_0 = const()[name = string("input_23_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string var_2563_pad_type_0 = const()[name = string("op_2563_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> var_2563_pad_0 = const()[name = string("op_2563_pad_0"), val = tensor<int32, [2]>([7, 7])];
            int32 var_2563_groups_0 = const()[name = string("op_2563_groups_0"), val = int32(1024)];
            tensor<int32, [1]> var_2563_strides_0 = const()[name = string("op_2563_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> var_2563_dilations_0 = const()[name = string("op_2563_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 15]> insert_mix_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 15]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5402688))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5418112))))[name = string("insert_mix_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> insert_mix_bias_to_fp16 = const()[name = string("insert_mix_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5420224)))];
            tensor<fp16, [1, 1024, 188]> input_23_cast_fp16 = transpose(perm = input_23_perm_0, x = var_2555_cast_fp16)[name = string("transpose_482")];
            tensor<fp16, [1, 1024, 188]> var_2563_cast_fp16 = conv(bias = insert_mix_bias_to_fp16, dilations = var_2563_dilations_0, groups = var_2563_groups_0, pad = var_2563_pad_0, pad_type = var_2563_pad_type_0, strides = var_2563_strides_0, weight = insert_mix_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("op_2563_cast_fp16")];
            string var_2564_mode_0 = const()[name = string("op_2564_mode_0"), val = string("EXACT")];
            tensor<fp16, [1, 1024, 188]> var_2564_cast_fp16 = gelu(mode = var_2564_mode_0, x = var_2563_cast_fp16)[name = string("op_2564_cast_fp16")];
            tensor<int32, [3]> input_25_perm_0 = const()[name = string("input_25_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> insert_up_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5422336))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6470976))))[name = string("insert_up_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> insert_up_bias_to_fp16 = const()[name = string("insert_up_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6473088)))];
            tensor<fp16, [1, 188, 1024]> input_25_cast_fp16 = transpose(perm = input_25_perm_0, x = var_2564_cast_fp16)[name = string("transpose_481")];
            tensor<fp16, [1, 188, 1024]> linear_2_cast_fp16 = linear(bias = insert_up_bias_to_fp16, weight = insert_up_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> x_13_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_2590_to_fp16 = const()[name = string("op_2590_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 188]> var_2592_cast_fp16 = sub(x = var_2590_to_fp16, y = var_2454_cast_fp16_to_fp16)[name = string("op_2592_cast_fp16")];
            fp16 var_2593_to_fp16 = const()[name = string("op_2593_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 188]> var_2594_cast_fp16 = mul(x = var_2592_cast_fp16, y = var_2593_to_fp16)[name = string("op_2594_cast_fp16")];
            tensor<int32, [4]> var_2599 = const()[name = string("op_2599"), val = tensor<int32, [4]>([1, 1, 1, 188])];
            tensor<fp16, [1, 1, 1, 188]> key_bias_cast_fp16 = reshape(shape = var_2599, x = var_2594_cast_fp16)[name = string("key_bias_cast_fp16")];
            tensor<int32, [1]> valid_axes_0 = const()[name = string("valid_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 188, 1]> valid_cast_fp16 = expand_dims(axes = valid_axes_0, x = var_2454_cast_fp16_to_fp16)[name = string("valid_cast_fp16")];
            fp16 var_2608_promoted_to_fp16 = const()[name = string("op_2608_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1]> var_2609_cast_fp16 = mul(x = length_cast_fp16, y = var_2608_promoted_to_fp16)[name = string("op_2609_cast_fp16")];
            fp16 var_2611_promoted_to_fp16 = const()[name = string("op_2611_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1]> var_2612_cast_fp16 = add(x = var_2609_cast_fp16, y = var_2611_promoted_to_fp16)[name = string("op_2612_cast_fp16")];
            tensor<int32, [3]> var_2616 = const()[name = string("op_2616"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<fp16, [1, 1, 1]> var_2617_cast_fp16 = reshape(shape = var_2616, x = var_2612_cast_fp16)[name = string("op_2617_cast_fp16")];
            tensor<fp16, [1, 375, 1024]> pos_emb_to_fp16 = const()[name = string("pos_emb_to_fp16"), val = tensor<fp16, [1, 375, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6475200)))];
            tensor<fp16, [1, 375, 1024]> input_39_cast_fp16 = mul(x = pos_emb_to_fp16, y = var_2617_cast_fp16)[name = string("input_39_cast_fp16")];
            int32 var_2628 = const()[name = string("op_2628"), val = int32(-1)];
            tensor<int32, [1]> input_29_axes_0 = const()[name = string("input_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7243264)))];
            tensor<fp16, [1024]> enc_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7245376)))];
            fp16 var_2636_to_fp16 = const()[name = string("op_2636_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = enc_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_2636_to_fp16, gamma = enc_layers_0_norm_feed_forward1_weight_to_fp16, x = x_13_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_0_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7247488))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11441856))))[name = string("blocks_0_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [4096]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11450112)))];
            tensor<fp16, [1, 188, 4096]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_0_ff1_1_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_33_cast_fp16 = silu(x = linear_3_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_0_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11458368))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15652736))))[name = string("blocks_0_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15654848)))];
            tensor<fp16, [1, 188, 1024]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_0_ff1_2_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_4_cast_fp16")];
            fp16 var_2656_to_fp16 = const()[name = string("op_2656_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_2657_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_2656_to_fp16)[name = string("op_2657_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_35_cast_fp16 = add(x = x_13_cast_fp16, y = var_2657_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<int32, [1]> input_37_axes_0 = const()[name = string("input_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15656960)))];
            tensor<fp16, [1024]> enc_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15659072)))];
            tensor<fp16, [1, 188, 1024]> input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = enc_layers_0_norm_self_att_bias_to_fp16, epsilon = var_2636_to_fp16, gamma = enc_layers_0_norm_self_att_weight_to_fp16, x = input_35_cast_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_0_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15661184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16709824))))[name = string("blocks_0_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_0_q_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<int32, [4]> var_2663 = const()[name = string("op_2663"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_2664_cast_fp16 = reshape(shape = var_2663, x = linear_5_cast_fp16)[name = string("op_2664_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_0_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16711936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17760576))))[name = string("blocks_0_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_6_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_0_k_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<int32, [4]> var_2668 = const()[name = string("op_2668"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_2669_cast_fp16 = reshape(shape = var_2668, x = linear_6_cast_fp16)[name = string("op_2669_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_0_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17762688))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18811328))))[name = string("blocks_0_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_7_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_0_v_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_2673 = const()[name = string("op_2673"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_2674_cast_fp16 = reshape(shape = var_2673, x = linear_7_cast_fp16)[name = string("op_2674_cast_fp16")];
            tensor<int32, [4]> v_1_perm_0 = const()[name = string("v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_86_to_fp16 = const()[name = string("const_86_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18813440)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_1_cast_fp16 = add(x = var_2664_cast_fp16, y = const_86_to_fp16)[name = string("q_u_1_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_87_to_fp16 = const()[name = string("const_87_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18815552)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_1_cast_fp16 = add(x = var_2664_cast_fp16, y = const_87_to_fp16)[name = string("q_v_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_0_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18817664))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19866304))))[name = string("blocks_0_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_8_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_0_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_2684 = const()[name = string("op_2684"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_1_cast_fp16 = reshape(shape = var_2684, x = linear_8_cast_fp16)[name = string("p_1_cast_fp16")];
            bool input_41_transpose_x_0 = const()[name = string("input_41_transpose_x_0"), val = bool(false)];
            bool input_41_transpose_y_0 = const()[name = string("input_41_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_192_perm_0 = const()[name = string("transpose_192_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_193_perm_0 = const()[name = string("transpose_193_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_193 = transpose(perm = transpose_193_perm_0, x = p_1_cast_fp16)[name = string("transpose_479")];
            tensor<fp16, [1, 8, 188, 128]> transpose_192 = transpose(perm = transpose_192_perm_0, x = q_v_1_cast_fp16)[name = string("transpose_480")];
            tensor<fp16, [1, 8, 188, 375]> input_41_cast_fp16 = matmul(transpose_x = input_41_transpose_x_0, transpose_y = input_41_transpose_y_0, x = transpose_192, y = transpose_193)[name = string("input_41_cast_fp16")];
            tensor<int32, [8]> var_2690_pad_0 = const()[name = string("op_2690_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_2690_mode_0 = const()[name = string("op_2690_mode_0"), val = string("constant")];
            fp16 const_14_to_fp16 = const()[name = string("const_14_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_2690_cast_fp16 = pad(constant_val = const_14_to_fp16, mode = var_2690_mode_0, pad = var_2690_pad_0, x = input_41_cast_fp16)[name = string("op_2690_cast_fp16")];
            tensor<int32, [4]> var_2691 = const()[name = string("op_2691"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_1_cast_fp16 = reshape(shape = var_2691, x = var_2690_cast_fp16)[name = string("scores_1_cast_fp16")];
            tensor<int32, [4]> var_2695_begin_0 = const()[name = string("op_2695_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2695_end_0 = const()[name = string("op_2695_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_2695_end_mask_0 = const()[name = string("op_2695_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_2695_cast_fp16 = slice_by_index(begin = var_2695_begin_0, end = var_2695_end_0, end_mask = var_2695_end_mask_0, x = scores_1_cast_fp16)[name = string("op_2695_cast_fp16")];
            tensor<int32, [4]> var_2696 = const()[name = string("op_2696"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_2697_cast_fp16 = reshape(shape = var_2696, x = var_2695_cast_fp16)[name = string("op_2697_cast_fp16")];
            tensor<int32, [4]> bd_1_begin_0 = const()[name = string("bd_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_1_end_0 = const()[name = string("bd_1_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_1_end_mask_0 = const()[name = string("bd_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_1_cast_fp16 = slice_by_index(begin = bd_1_begin_0, end = bd_1_end_0, end_mask = bd_1_end_mask_0, x = var_2697_cast_fp16)[name = string("bd_1_cast_fp16")];
            bool var_2700_transpose_x_0 = const()[name = string("op_2700_transpose_x_0"), val = bool(false)];
            bool var_2700_transpose_y_0 = const()[name = string("op_2700_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_194_perm_0 = const()[name = string("transpose_194_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_195_perm_0 = const()[name = string("transpose_195_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_195 = transpose(perm = transpose_195_perm_0, x = var_2669_cast_fp16)[name = string("transpose_477")];
            tensor<fp16, [1, 8, 188, 128]> transpose_194 = transpose(perm = transpose_194_perm_0, x = q_u_1_cast_fp16)[name = string("transpose_478")];
            tensor<fp16, [1, 8, 188, 188]> var_2700_cast_fp16 = matmul(transpose_x = var_2700_transpose_x_0, transpose_y = var_2700_transpose_y_0, x = transpose_194, y = transpose_195)[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_2701_cast_fp16 = add(x = var_2700_cast_fp16, y = bd_1_cast_fp16)[name = string("op_2701_cast_fp16")];
            fp16 var_2702_to_fp16 = const()[name = string("op_2702_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_2703_cast_fp16 = mul(x = var_2701_cast_fp16, y = var_2702_to_fp16)[name = string("op_2703_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_3_cast_fp16 = add(x = var_2703_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_3_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_2705_cast_fp16 = softmax(axis = var_2628, x = scores_3_cast_fp16)[name = string("op_2705_cast_fp16")];
            bool out_1_transpose_x_0 = const()[name = string("out_1_transpose_x_0"), val = bool(false)];
            bool out_1_transpose_y_0 = const()[name = string("out_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_2674_cast_fp16)[name = string("transpose_476")];
            tensor<fp16, [1, 8, 188, 128]> out_1_cast_fp16 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = var_2705_cast_fp16, y = v_1_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [4]> var_2707_perm_0 = const()[name = string("op_2707_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2708 = const()[name = string("op_2708"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_2707_cast_fp16 = transpose(perm = var_2707_perm_0, x = out_1_cast_fp16)[name = string("transpose_475")];
            tensor<fp16, [1, 188, 1024]> input_43_cast_fp16 = reshape(shape = var_2708, x = var_2707_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_0_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19868416))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20917056))))[name = string("blocks_0_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_9_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_0_o_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_45_cast_fp16 = add(x = input_35_cast_fp16, y = linear_9_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_0_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20919168)))];
            tensor<fp16, [1024]> enc_layers_0_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20921280)))];
            tensor<fp16, [1, 188, 1024]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = enc_layers_0_norm_conv_bias_to_fp16, epsilon = var_2636_to_fp16, gamma = enc_layers_0_norm_conv_weight_to_fp16, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_0_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20923392))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23020608))))[name = string("blocks_0_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_10_bias_0_to_fp16 = const()[name = string("linear_10_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23024768)))];
            tensor<fp16, [1, 188, 2048]> linear_10_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_0_pw1_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_10_cast_fp16")];
            int32 var_2717_split_num_splits_0 = const()[name = string("op_2717_split_num_splits_0"), val = int32(2)];
            int32 var_2717_split_axis_0 = const()[name = string("op_2717_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_2717_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_2717_split_cast_fp16_1 = split(axis = var_2717_split_axis_0, num_splits = var_2717_split_num_splits_0, x = linear_10_cast_fp16)[name = string("op_2717_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_2717_split_1_sigmoid_cast_fp16 = sigmoid(x = var_2717_split_cast_fp16_1)[name = string("op_2717_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_2717_cast_fp16 = mul(x = var_2717_split_cast_fp16_0, y = var_2717_split_1_sigmoid_cast_fp16)[name = string("op_2717_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_1_cast_fp16 = mul(x = var_2717_cast_fp16, y = valid_cast_fp16)[name = string("h_1_cast_fp16")];
            tensor<int32, [3]> input_51_perm_0 = const()[name = string("input_51_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_53_pad_type_0 = const()[name = string("input_53_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_53_pad_0 = const()[name = string("input_53_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_53_groups_0 = const()[name = string("input_53_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_53_strides_0 = const()[name = string("input_53_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_53_dilations_0 = const()[name = string("input_53_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_38_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23028928))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23038208))))[name = string("const_38_to_fp16_quantized")];
            tensor<fp16, [1024]> const_39_to_fp16 = const()[name = string("const_39_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23040320)))];
            tensor<fp16, [1, 1024, 188]> input_51_cast_fp16 = transpose(perm = input_51_perm_0, x = h_1_cast_fp16)[name = string("transpose_474")];
            tensor<fp16, [1, 1024, 188]> input_55_cast_fp16 = conv(bias = const_39_to_fp16, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = const_38_to_fp16_quantized, x = input_51_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_3_cast_fp16 = silu(x = input_55_cast_fp16)[name = string("h_3_cast_fp16")];
            tensor<int32, [3]> input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_0_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23042432))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24091072))))[name = string("blocks_0_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = h_3_cast_fp16)[name = string("transpose_473")];
            tensor<fp16, [1, 188, 1024]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_0_pw2_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_59_cast_fp16 = add(x = input_45_cast_fp16, y = linear_11_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<int32, [1]> input_61_axes_0 = const()[name = string("input_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24093184)))];
            tensor<fp16, [1024]> enc_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24095296)))];
            tensor<fp16, [1, 188, 1024]> input_61_cast_fp16 = layer_norm(axes = input_61_axes_0, beta = enc_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_2636_to_fp16, gamma = enc_layers_0_norm_feed_forward2_weight_to_fp16, x = input_59_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_0_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24097408))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28291776))))[name = string("blocks_0_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_12_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_0_ff2_1_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_65_cast_fp16 = silu(x = linear_12_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_0_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28300032))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32494400))))[name = string("blocks_0_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_13_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_0_ff2_2_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_13_cast_fp16")];
            fp16 var_2738_to_fp16 = const()[name = string("op_2738_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_2739_cast_fp16 = mul(x = linear_13_cast_fp16, y = var_2738_to_fp16)[name = string("op_2739_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_67_cast_fp16 = add(x = input_59_cast_fp16, y = var_2739_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<int32, [1]> input_69_axes_0 = const()[name = string("input_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_0_norm_out_weight_to_fp16 = const()[name = string("enc_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32496512)))];
            tensor<fp16, [1024]> enc_layers_0_norm_out_bias_to_fp16 = const()[name = string("enc_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32498624)))];
            tensor<fp16, [1, 188, 1024]> input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = enc_layers_0_norm_out_bias_to_fp16, epsilon = var_2636_to_fp16, gamma = enc_layers_0_norm_out_weight_to_fp16, x = input_67_cast_fp16)[name = string("input_69_cast_fp16")];
            int32 var_2752 = const()[name = string("op_2752"), val = int32(-1)];
            tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32500736)))];
            tensor<fp16, [1024]> enc_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32502848)))];
            fp16 var_2760_to_fp16 = const()[name = string("op_2760_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = enc_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_2760_to_fp16, gamma = enc_layers_1_norm_feed_forward1_weight_to_fp16, x = input_69_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_1_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32504960))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36699328))))[name = string("blocks_1_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_14_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_1_ff1_1_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_75_cast_fp16 = silu(x = linear_14_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_1_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36707584))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40901952))))[name = string("blocks_1_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_15_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_1_ff1_2_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_15_cast_fp16")];
            fp16 var_2780_to_fp16 = const()[name = string("op_2780_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_2781_cast_fp16 = mul(x = linear_15_cast_fp16, y = var_2780_to_fp16)[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_77_cast_fp16 = add(x = input_69_cast_fp16, y = var_2781_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<int32, [1]> input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40904064)))];
            tensor<fp16, [1024]> enc_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40906176)))];
            tensor<fp16, [1, 188, 1024]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = enc_layers_1_norm_self_att_bias_to_fp16, epsilon = var_2760_to_fp16, gamma = enc_layers_1_norm_self_att_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_1_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40908288))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41956928))))[name = string("blocks_1_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_16_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_1_q_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_2787 = const()[name = string("op_2787"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_2788_cast_fp16 = reshape(shape = var_2787, x = linear_16_cast_fp16)[name = string("op_2788_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_1_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41959040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43007680))))[name = string("blocks_1_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_17_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_1_k_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<int32, [4]> var_2792 = const()[name = string("op_2792"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_2793_cast_fp16 = reshape(shape = var_2792, x = linear_17_cast_fp16)[name = string("op_2793_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_1_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43009792))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44058432))))[name = string("blocks_1_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_1_v_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<int32, [4]> var_2797 = const()[name = string("op_2797"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_2798_cast_fp16 = reshape(shape = var_2797, x = linear_18_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_88_to_fp16 = const()[name = string("const_88_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44060544)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_3_cast_fp16 = add(x = var_2788_cast_fp16, y = const_88_to_fp16)[name = string("q_u_3_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_89_to_fp16 = const()[name = string("const_89_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44062656)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_3_cast_fp16 = add(x = var_2788_cast_fp16, y = const_89_to_fp16)[name = string("q_v_3_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_1_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44064768))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45113408))))[name = string("blocks_1_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_1_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<int32, [4]> var_2808 = const()[name = string("op_2808"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_3_cast_fp16 = reshape(shape = var_2808, x = linear_19_cast_fp16)[name = string("p_3_cast_fp16")];
            bool input_81_transpose_x_0 = const()[name = string("input_81_transpose_x_0"), val = bool(false)];
            bool input_81_transpose_y_0 = const()[name = string("input_81_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_196_perm_0 = const()[name = string("transpose_196_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_197_perm_0 = const()[name = string("transpose_197_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_197 = transpose(perm = transpose_197_perm_0, x = p_3_cast_fp16)[name = string("transpose_471")];
            tensor<fp16, [1, 8, 188, 128]> transpose_196 = transpose(perm = transpose_196_perm_0, x = q_v_3_cast_fp16)[name = string("transpose_472")];
            tensor<fp16, [1, 8, 188, 375]> input_81_cast_fp16 = matmul(transpose_x = input_81_transpose_x_0, transpose_y = input_81_transpose_y_0, x = transpose_196, y = transpose_197)[name = string("input_81_cast_fp16")];
            tensor<int32, [8]> var_2814_pad_0 = const()[name = string("op_2814_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_2814_mode_0 = const()[name = string("op_2814_mode_0"), val = string("constant")];
            fp16 const_15_to_fp16 = const()[name = string("const_15_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_2814_cast_fp16 = pad(constant_val = const_15_to_fp16, mode = var_2814_mode_0, pad = var_2814_pad_0, x = input_81_cast_fp16)[name = string("op_2814_cast_fp16")];
            tensor<int32, [4]> var_2815 = const()[name = string("op_2815"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_5_cast_fp16 = reshape(shape = var_2815, x = var_2814_cast_fp16)[name = string("scores_5_cast_fp16")];
            tensor<int32, [4]> var_2819_begin_0 = const()[name = string("op_2819_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2819_end_0 = const()[name = string("op_2819_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_2819_end_mask_0 = const()[name = string("op_2819_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_2819_cast_fp16 = slice_by_index(begin = var_2819_begin_0, end = var_2819_end_0, end_mask = var_2819_end_mask_0, x = scores_5_cast_fp16)[name = string("op_2819_cast_fp16")];
            tensor<int32, [4]> var_2820 = const()[name = string("op_2820"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_2821_cast_fp16 = reshape(shape = var_2820, x = var_2819_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<int32, [4]> bd_3_begin_0 = const()[name = string("bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_3_end_0 = const()[name = string("bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_3_end_mask_0 = const()[name = string("bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_3_cast_fp16 = slice_by_index(begin = bd_3_begin_0, end = bd_3_end_0, end_mask = bd_3_end_mask_0, x = var_2821_cast_fp16)[name = string("bd_3_cast_fp16")];
            bool var_2824_transpose_x_0 = const()[name = string("op_2824_transpose_x_0"), val = bool(false)];
            bool var_2824_transpose_y_0 = const()[name = string("op_2824_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_198_perm_0 = const()[name = string("transpose_198_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_199_perm_0 = const()[name = string("transpose_199_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_199 = transpose(perm = transpose_199_perm_0, x = var_2793_cast_fp16)[name = string("transpose_469")];
            tensor<fp16, [1, 8, 188, 128]> transpose_198 = transpose(perm = transpose_198_perm_0, x = q_u_3_cast_fp16)[name = string("transpose_470")];
            tensor<fp16, [1, 8, 188, 188]> var_2824_cast_fp16 = matmul(transpose_x = var_2824_transpose_x_0, transpose_y = var_2824_transpose_y_0, x = transpose_198, y = transpose_199)[name = string("op_2824_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_2825_cast_fp16 = add(x = var_2824_cast_fp16, y = bd_3_cast_fp16)[name = string("op_2825_cast_fp16")];
            fp16 var_2826_to_fp16 = const()[name = string("op_2826_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_2827_cast_fp16 = mul(x = var_2825_cast_fp16, y = var_2826_to_fp16)[name = string("op_2827_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_7_cast_fp16 = add(x = var_2827_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_7_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_2829_cast_fp16 = softmax(axis = var_2752, x = scores_7_cast_fp16)[name = string("op_2829_cast_fp16")];
            bool out_3_transpose_x_0 = const()[name = string("out_3_transpose_x_0"), val = bool(false)];
            bool out_3_transpose_y_0 = const()[name = string("out_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_2798_cast_fp16)[name = string("transpose_468")];
            tensor<fp16, [1, 8, 188, 128]> out_3_cast_fp16 = matmul(transpose_x = out_3_transpose_x_0, transpose_y = out_3_transpose_y_0, x = var_2829_cast_fp16, y = v_3_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [4]> var_2831_perm_0 = const()[name = string("op_2831_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2832 = const()[name = string("op_2832"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_2831_cast_fp16 = transpose(perm = var_2831_perm_0, x = out_3_cast_fp16)[name = string("transpose_467")];
            tensor<fp16, [1, 188, 1024]> input_83_cast_fp16 = reshape(shape = var_2832, x = var_2831_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_1_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45115520))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46164160))))[name = string("blocks_1_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_20_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_1_o_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_85_cast_fp16 = add(x = input_77_cast_fp16, y = linear_20_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<int32, [1]> input_87_axes_0 = const()[name = string("input_87_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_1_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46166272)))];
            tensor<fp16, [1024]> enc_layers_1_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46168384)))];
            tensor<fp16, [1, 188, 1024]> input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = enc_layers_1_norm_conv_bias_to_fp16, epsilon = var_2760_to_fp16, gamma = enc_layers_1_norm_conv_weight_to_fp16, x = input_85_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_1_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46170496))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48267712))))[name = string("blocks_1_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_21_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_1_pw1_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_21_cast_fp16")];
            int32 var_2841_split_num_splits_0 = const()[name = string("op_2841_split_num_splits_0"), val = int32(2)];
            int32 var_2841_split_axis_0 = const()[name = string("op_2841_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_2841_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_2841_split_cast_fp16_1 = split(axis = var_2841_split_axis_0, num_splits = var_2841_split_num_splits_0, x = linear_21_cast_fp16)[name = string("op_2841_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_2841_split_1_sigmoid_cast_fp16 = sigmoid(x = var_2841_split_cast_fp16_1)[name = string("op_2841_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_2841_cast_fp16 = mul(x = var_2841_split_cast_fp16_0, y = var_2841_split_1_sigmoid_cast_fp16)[name = string("op_2841_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_5_cast_fp16 = mul(x = var_2841_cast_fp16, y = valid_cast_fp16)[name = string("h_5_cast_fp16")];
            tensor<int32, [3]> input_91_perm_0 = const()[name = string("input_91_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_93_pad_type_0 = const()[name = string("input_93_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_93_pad_0 = const()[name = string("input_93_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_93_groups_0 = const()[name = string("input_93_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_93_strides_0 = const()[name = string("input_93_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_93_dilations_0 = const()[name = string("input_93_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_40_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48271872))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48281152))))[name = string("const_40_to_fp16_quantized")];
            tensor<fp16, [1024]> const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48283264)))];
            tensor<fp16, [1, 1024, 188]> input_91_cast_fp16 = transpose(perm = input_91_perm_0, x = h_5_cast_fp16)[name = string("transpose_466")];
            tensor<fp16, [1, 1024, 188]> input_95_cast_fp16 = conv(bias = const_41_to_fp16, 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 = const_40_to_fp16_quantized, x = input_91_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_7_cast_fp16 = silu(x = input_95_cast_fp16)[name = string("h_7_cast_fp16")];
            tensor<int32, [3]> input_97_perm_0 = const()[name = string("input_97_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_1_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48285376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49334016))))[name = string("blocks_1_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_97_cast_fp16 = transpose(perm = input_97_perm_0, x = h_7_cast_fp16)[name = string("transpose_465")];
            tensor<fp16, [1, 188, 1024]> linear_22_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_1_pw2_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_99_cast_fp16 = add(x = input_85_cast_fp16, y = linear_22_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49336128)))];
            tensor<fp16, [1024]> enc_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49338240)))];
            tensor<fp16, [1, 188, 1024]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = enc_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_2760_to_fp16, gamma = enc_layers_1_norm_feed_forward2_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_1_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49340352))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53534720))))[name = string("blocks_1_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_1_ff2_1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_105_cast_fp16 = silu(x = linear_23_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_1_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53542976))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57737344))))[name = string("blocks_1_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_24_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_1_ff2_2_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_24_cast_fp16")];
            fp16 var_2862_to_fp16 = const()[name = string("op_2862_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_2863_cast_fp16 = mul(x = linear_24_cast_fp16, y = var_2862_to_fp16)[name = string("op_2863_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_107_cast_fp16 = add(x = input_99_cast_fp16, y = var_2863_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<int32, [1]> input_109_axes_0 = const()[name = string("input_109_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_1_norm_out_weight_to_fp16 = const()[name = string("enc_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57739456)))];
            tensor<fp16, [1024]> enc_layers_1_norm_out_bias_to_fp16 = const()[name = string("enc_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57741568)))];
            tensor<fp16, [1, 188, 1024]> input_109_cast_fp16 = layer_norm(axes = input_109_axes_0, beta = enc_layers_1_norm_out_bias_to_fp16, epsilon = var_2760_to_fp16, gamma = enc_layers_1_norm_out_weight_to_fp16, x = input_107_cast_fp16)[name = string("input_109_cast_fp16")];
            int32 var_2876 = const()[name = string("op_2876"), val = int32(-1)];
            tensor<int32, [1]> input_111_axes_0 = const()[name = string("input_111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57743680)))];
            tensor<fp16, [1024]> enc_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57745792)))];
            fp16 var_2884_to_fp16 = const()[name = string("op_2884_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = enc_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_2884_to_fp16, gamma = enc_layers_2_norm_feed_forward1_weight_to_fp16, x = input_109_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_2_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57747904))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61942272))))[name = string("blocks_2_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_25_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_2_ff1_1_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_115_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_2_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61950528))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66144896))))[name = string("blocks_2_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_2_ff1_2_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_26_cast_fp16")];
            fp16 var_2904_to_fp16 = const()[name = string("op_2904_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_2905_cast_fp16 = mul(x = linear_26_cast_fp16, y = var_2904_to_fp16)[name = string("op_2905_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_117_cast_fp16 = add(x = input_109_cast_fp16, y = var_2905_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<int32, [1]> input_119_axes_0 = const()[name = string("input_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66147008)))];
            tensor<fp16, [1024]> enc_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66149120)))];
            tensor<fp16, [1, 188, 1024]> input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = enc_layers_2_norm_self_att_bias_to_fp16, epsilon = var_2884_to_fp16, gamma = enc_layers_2_norm_self_att_weight_to_fp16, x = input_117_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_2_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66151232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67199872))))[name = string("blocks_2_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_27_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_2_q_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<int32, [4]> var_2911 = const()[name = string("op_2911"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_2912_cast_fp16 = reshape(shape = var_2911, x = linear_27_cast_fp16)[name = string("op_2912_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_2_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67201984))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68250624))))[name = string("blocks_2_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_28_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_2_k_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_2916 = const()[name = string("op_2916"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_2917_cast_fp16 = reshape(shape = var_2916, x = linear_28_cast_fp16)[name = string("op_2917_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_2_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68252736))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69301376))))[name = string("blocks_2_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_29_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_2_v_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_2921 = const()[name = string("op_2921"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_2922_cast_fp16 = reshape(shape = var_2921, x = linear_29_cast_fp16)[name = string("op_2922_cast_fp16")];
            tensor<int32, [4]> v_5_perm_0 = const()[name = string("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_90_to_fp16 = const()[name = string("const_90_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69303488)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_5_cast_fp16 = add(x = var_2912_cast_fp16, y = const_90_to_fp16)[name = string("q_u_5_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_91_to_fp16 = const()[name = string("const_91_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69305600)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_5_cast_fp16 = add(x = var_2912_cast_fp16, y = const_91_to_fp16)[name = string("q_v_5_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_2_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69307712))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70356352))))[name = string("blocks_2_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_30_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_2_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_2932 = const()[name = string("op_2932"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_5_cast_fp16 = reshape(shape = var_2932, x = linear_30_cast_fp16)[name = string("p_5_cast_fp16")];
            bool input_121_transpose_x_0 = const()[name = string("input_121_transpose_x_0"), val = bool(false)];
            bool input_121_transpose_y_0 = const()[name = string("input_121_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_200_perm_0 = const()[name = string("transpose_200_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_201_perm_0 = const()[name = string("transpose_201_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_201 = transpose(perm = transpose_201_perm_0, x = p_5_cast_fp16)[name = string("transpose_463")];
            tensor<fp16, [1, 8, 188, 128]> transpose_200 = transpose(perm = transpose_200_perm_0, x = q_v_5_cast_fp16)[name = string("transpose_464")];
            tensor<fp16, [1, 8, 188, 375]> input_121_cast_fp16 = matmul(transpose_x = input_121_transpose_x_0, transpose_y = input_121_transpose_y_0, x = transpose_200, y = transpose_201)[name = string("input_121_cast_fp16")];
            tensor<int32, [8]> var_2938_pad_0 = const()[name = string("op_2938_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_2938_mode_0 = const()[name = string("op_2938_mode_0"), val = string("constant")];
            fp16 const_16_to_fp16 = const()[name = string("const_16_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_2938_cast_fp16 = pad(constant_val = const_16_to_fp16, mode = var_2938_mode_0, pad = var_2938_pad_0, x = input_121_cast_fp16)[name = string("op_2938_cast_fp16")];
            tensor<int32, [4]> var_2939 = const()[name = string("op_2939"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_9_cast_fp16 = reshape(shape = var_2939, x = var_2938_cast_fp16)[name = string("scores_9_cast_fp16")];
            tensor<int32, [4]> var_2943_begin_0 = const()[name = string("op_2943_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_2943_end_0 = const()[name = string("op_2943_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_2943_end_mask_0 = const()[name = string("op_2943_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_2943_cast_fp16 = slice_by_index(begin = var_2943_begin_0, end = var_2943_end_0, end_mask = var_2943_end_mask_0, x = scores_9_cast_fp16)[name = string("op_2943_cast_fp16")];
            tensor<int32, [4]> var_2944 = const()[name = string("op_2944"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_2945_cast_fp16 = reshape(shape = var_2944, x = var_2943_cast_fp16)[name = string("op_2945_cast_fp16")];
            tensor<int32, [4]> bd_5_begin_0 = const()[name = string("bd_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_5_end_0 = const()[name = string("bd_5_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_5_end_mask_0 = const()[name = string("bd_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_5_cast_fp16 = slice_by_index(begin = bd_5_begin_0, end = bd_5_end_0, end_mask = bd_5_end_mask_0, x = var_2945_cast_fp16)[name = string("bd_5_cast_fp16")];
            bool var_2948_transpose_x_0 = const()[name = string("op_2948_transpose_x_0"), val = bool(false)];
            bool var_2948_transpose_y_0 = const()[name = string("op_2948_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_202_perm_0 = const()[name = string("transpose_202_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_203_perm_0 = const()[name = string("transpose_203_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_203 = transpose(perm = transpose_203_perm_0, x = var_2917_cast_fp16)[name = string("transpose_461")];
            tensor<fp16, [1, 8, 188, 128]> transpose_202 = transpose(perm = transpose_202_perm_0, x = q_u_5_cast_fp16)[name = string("transpose_462")];
            tensor<fp16, [1, 8, 188, 188]> var_2948_cast_fp16 = matmul(transpose_x = var_2948_transpose_x_0, transpose_y = var_2948_transpose_y_0, x = transpose_202, y = transpose_203)[name = string("op_2948_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_2949_cast_fp16 = add(x = var_2948_cast_fp16, y = bd_5_cast_fp16)[name = string("op_2949_cast_fp16")];
            fp16 var_2950_to_fp16 = const()[name = string("op_2950_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_2951_cast_fp16 = mul(x = var_2949_cast_fp16, y = var_2950_to_fp16)[name = string("op_2951_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_11_cast_fp16 = add(x = var_2951_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_11_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_2953_cast_fp16 = softmax(axis = var_2876, x = scores_11_cast_fp16)[name = string("op_2953_cast_fp16")];
            bool out_5_transpose_x_0 = const()[name = string("out_5_transpose_x_0"), val = bool(false)];
            bool out_5_transpose_y_0 = const()[name = string("out_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_2922_cast_fp16)[name = string("transpose_460")];
            tensor<fp16, [1, 8, 188, 128]> out_5_cast_fp16 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = var_2953_cast_fp16, y = v_5_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [4]> var_2955_perm_0 = const()[name = string("op_2955_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2956 = const()[name = string("op_2956"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_2955_cast_fp16 = transpose(perm = var_2955_perm_0, x = out_5_cast_fp16)[name = string("transpose_459")];
            tensor<fp16, [1, 188, 1024]> input_123_cast_fp16 = reshape(shape = var_2956, x = var_2955_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_2_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70358464))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71407104))))[name = string("blocks_2_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_31_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_2_o_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_125_cast_fp16 = add(x = input_117_cast_fp16, y = linear_31_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<int32, [1]> input_127_axes_0 = const()[name = string("input_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_2_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71409216)))];
            tensor<fp16, [1024]> enc_layers_2_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71411328)))];
            tensor<fp16, [1, 188, 1024]> input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = enc_layers_2_norm_conv_bias_to_fp16, epsilon = var_2884_to_fp16, gamma = enc_layers_2_norm_conv_weight_to_fp16, x = input_125_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_2_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71413440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73510656))))[name = string("blocks_2_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_32_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_2_pw1_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_32_cast_fp16")];
            int32 var_2965_split_num_splits_0 = const()[name = string("op_2965_split_num_splits_0"), val = int32(2)];
            int32 var_2965_split_axis_0 = const()[name = string("op_2965_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_2965_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_2965_split_cast_fp16_1 = split(axis = var_2965_split_axis_0, num_splits = var_2965_split_num_splits_0, x = linear_32_cast_fp16)[name = string("op_2965_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_2965_split_1_sigmoid_cast_fp16 = sigmoid(x = var_2965_split_cast_fp16_1)[name = string("op_2965_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_2965_cast_fp16 = mul(x = var_2965_split_cast_fp16_0, y = var_2965_split_1_sigmoid_cast_fp16)[name = string("op_2965_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_9_cast_fp16 = mul(x = var_2965_cast_fp16, y = valid_cast_fp16)[name = string("h_9_cast_fp16")];
            tensor<int32, [3]> input_131_perm_0 = const()[name = string("input_131_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_133_pad_type_0 = const()[name = string("input_133_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_133_pad_0 = const()[name = string("input_133_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_133_groups_0 = const()[name = string("input_133_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_133_strides_0 = const()[name = string("input_133_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_133_dilations_0 = const()[name = string("input_133_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_42_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73514816))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73524096))))[name = string("const_42_to_fp16_quantized")];
            tensor<fp16, [1024]> const_43_to_fp16 = const()[name = string("const_43_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73526208)))];
            tensor<fp16, [1, 1024, 188]> input_131_cast_fp16 = transpose(perm = input_131_perm_0, x = h_9_cast_fp16)[name = string("transpose_458")];
            tensor<fp16, [1, 1024, 188]> input_135_cast_fp16 = conv(bias = const_43_to_fp16, 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 = const_42_to_fp16_quantized, x = input_131_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_11_cast_fp16 = silu(x = input_135_cast_fp16)[name = string("h_11_cast_fp16")];
            tensor<int32, [3]> input_137_perm_0 = const()[name = string("input_137_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_2_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73528320))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74576960))))[name = string("blocks_2_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_137_cast_fp16 = transpose(perm = input_137_perm_0, x = h_11_cast_fp16)[name = string("transpose_457")];
            tensor<fp16, [1, 188, 1024]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_2_pw2_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_139_cast_fp16 = add(x = input_125_cast_fp16, y = linear_33_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<int32, [1]> input_141_axes_0 = const()[name = string("input_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74579072)))];
            tensor<fp16, [1024]> enc_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74581184)))];
            tensor<fp16, [1, 188, 1024]> input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = enc_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_2884_to_fp16, gamma = enc_layers_2_norm_feed_forward2_weight_to_fp16, x = input_139_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_2_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74583296))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78777664))))[name = string("blocks_2_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_2_ff2_1_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_145_cast_fp16 = silu(x = linear_34_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_2_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78785920))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82980288))))[name = string("blocks_2_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_35_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_2_ff2_2_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_35_cast_fp16")];
            fp16 var_2986_to_fp16 = const()[name = string("op_2986_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_2987_cast_fp16 = mul(x = linear_35_cast_fp16, y = var_2986_to_fp16)[name = string("op_2987_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_147_cast_fp16 = add(x = input_139_cast_fp16, y = var_2987_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<int32, [1]> input_149_axes_0 = const()[name = string("input_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_2_norm_out_weight_to_fp16 = const()[name = string("enc_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82982400)))];
            tensor<fp16, [1024]> enc_layers_2_norm_out_bias_to_fp16 = const()[name = string("enc_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82984512)))];
            tensor<fp16, [1, 188, 1024]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = enc_layers_2_norm_out_bias_to_fp16, epsilon = var_2884_to_fp16, gamma = enc_layers_2_norm_out_weight_to_fp16, x = input_147_cast_fp16)[name = string("input_149_cast_fp16")];
            int32 var_3000 = const()[name = string("op_3000"), val = int32(-1)];
            tensor<int32, [1]> input_151_axes_0 = const()[name = string("input_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82986624)))];
            tensor<fp16, [1024]> enc_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82988736)))];
            fp16 var_3008_to_fp16 = const()[name = string("op_3008_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_151_cast_fp16 = layer_norm(axes = input_151_axes_0, beta = enc_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_3008_to_fp16, gamma = enc_layers_3_norm_feed_forward1_weight_to_fp16, x = input_149_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_3_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82990848))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87185216))))[name = string("blocks_3_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_36_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_3_ff1_1_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_155_cast_fp16 = silu(x = linear_36_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_3_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87193472))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91387840))))[name = string("blocks_3_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_37_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_3_ff1_2_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_37_cast_fp16")];
            fp16 var_3028_to_fp16 = const()[name = string("op_3028_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3029_cast_fp16 = mul(x = linear_37_cast_fp16, y = var_3028_to_fp16)[name = string("op_3029_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_157_cast_fp16 = add(x = input_149_cast_fp16, y = var_3029_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<int32, [1]> input_159_axes_0 = const()[name = string("input_159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91389952)))];
            tensor<fp16, [1024]> enc_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91392064)))];
            tensor<fp16, [1, 188, 1024]> input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = enc_layers_3_norm_self_att_bias_to_fp16, epsilon = var_3008_to_fp16, gamma = enc_layers_3_norm_self_att_weight_to_fp16, x = input_157_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_3_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91394176))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92442816))))[name = string("blocks_3_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_38_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_3_q_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3035 = const()[name = string("op_3035"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3036_cast_fp16 = reshape(shape = var_3035, x = linear_38_cast_fp16)[name = string("op_3036_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_3_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92444928))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93493568))))[name = string("blocks_3_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_3_k_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<int32, [4]> var_3040 = const()[name = string("op_3040"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3041_cast_fp16 = reshape(shape = var_3040, x = linear_39_cast_fp16)[name = string("op_3041_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_3_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93495680))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94544320))))[name = string("blocks_3_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_3_v_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<int32, [4]> var_3045 = const()[name = string("op_3045"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3046_cast_fp16 = reshape(shape = var_3045, x = linear_40_cast_fp16)[name = string("op_3046_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94546432)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_7_cast_fp16 = add(x = var_3036_cast_fp16, y = const_92_to_fp16)[name = string("q_u_7_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_93_to_fp16 = const()[name = string("const_93_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94548544)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_7_cast_fp16 = add(x = var_3036_cast_fp16, y = const_93_to_fp16)[name = string("q_v_7_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_3_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94550656))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95599296))))[name = string("blocks_3_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_41_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_3_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<int32, [4]> var_3056 = const()[name = string("op_3056"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_7_cast_fp16 = reshape(shape = var_3056, x = linear_41_cast_fp16)[name = string("p_7_cast_fp16")];
            bool input_161_transpose_x_0 = const()[name = string("input_161_transpose_x_0"), val = bool(false)];
            bool input_161_transpose_y_0 = const()[name = string("input_161_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_204_perm_0 = const()[name = string("transpose_204_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_205_perm_0 = const()[name = string("transpose_205_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_205 = transpose(perm = transpose_205_perm_0, x = p_7_cast_fp16)[name = string("transpose_455")];
            tensor<fp16, [1, 8, 188, 128]> transpose_204 = transpose(perm = transpose_204_perm_0, x = q_v_7_cast_fp16)[name = string("transpose_456")];
            tensor<fp16, [1, 8, 188, 375]> input_161_cast_fp16 = matmul(transpose_x = input_161_transpose_x_0, transpose_y = input_161_transpose_y_0, x = transpose_204, y = transpose_205)[name = string("input_161_cast_fp16")];
            tensor<int32, [8]> var_3062_pad_0 = const()[name = string("op_3062_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_3062_mode_0 = const()[name = string("op_3062_mode_0"), val = string("constant")];
            fp16 const_17_to_fp16 = const()[name = string("const_17_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_3062_cast_fp16 = pad(constant_val = const_17_to_fp16, mode = var_3062_mode_0, pad = var_3062_pad_0, x = input_161_cast_fp16)[name = string("op_3062_cast_fp16")];
            tensor<int32, [4]> var_3063 = const()[name = string("op_3063"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_13_cast_fp16 = reshape(shape = var_3063, x = var_3062_cast_fp16)[name = string("scores_13_cast_fp16")];
            tensor<int32, [4]> var_3067_begin_0 = const()[name = string("op_3067_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3067_end_0 = const()[name = string("op_3067_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_3067_end_mask_0 = const()[name = string("op_3067_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_3067_cast_fp16 = slice_by_index(begin = var_3067_begin_0, end = var_3067_end_0, end_mask = var_3067_end_mask_0, x = scores_13_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<int32, [4]> var_3068 = const()[name = string("op_3068"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_3069_cast_fp16 = reshape(shape = var_3068, x = var_3067_cast_fp16)[name = string("op_3069_cast_fp16")];
            tensor<int32, [4]> bd_7_begin_0 = const()[name = string("bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_7_end_0 = const()[name = string("bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_7_end_mask_0 = const()[name = string("bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_7_cast_fp16 = slice_by_index(begin = bd_7_begin_0, end = bd_7_end_0, end_mask = bd_7_end_mask_0, x = var_3069_cast_fp16)[name = string("bd_7_cast_fp16")];
            bool var_3072_transpose_x_0 = const()[name = string("op_3072_transpose_x_0"), val = bool(false)];
            bool var_3072_transpose_y_0 = const()[name = string("op_3072_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_206_perm_0 = const()[name = string("transpose_206_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_207_perm_0 = const()[name = string("transpose_207_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_207 = transpose(perm = transpose_207_perm_0, x = var_3041_cast_fp16)[name = string("transpose_453")];
            tensor<fp16, [1, 8, 188, 128]> transpose_206 = transpose(perm = transpose_206_perm_0, x = q_u_7_cast_fp16)[name = string("transpose_454")];
            tensor<fp16, [1, 8, 188, 188]> var_3072_cast_fp16 = matmul(transpose_x = var_3072_transpose_x_0, transpose_y = var_3072_transpose_y_0, x = transpose_206, y = transpose_207)[name = string("op_3072_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3073_cast_fp16 = add(x = var_3072_cast_fp16, y = bd_7_cast_fp16)[name = string("op_3073_cast_fp16")];
            fp16 var_3074_to_fp16 = const()[name = string("op_3074_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_3075_cast_fp16 = mul(x = var_3073_cast_fp16, y = var_3074_to_fp16)[name = string("op_3075_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_15_cast_fp16 = add(x = var_3075_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_15_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3077_cast_fp16 = softmax(axis = var_3000, x = scores_15_cast_fp16)[name = string("op_3077_cast_fp16")];
            bool out_7_transpose_x_0 = const()[name = string("out_7_transpose_x_0"), val = bool(false)];
            bool out_7_transpose_y_0 = const()[name = string("out_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_3046_cast_fp16)[name = string("transpose_452")];
            tensor<fp16, [1, 8, 188, 128]> out_7_cast_fp16 = matmul(transpose_x = out_7_transpose_x_0, transpose_y = out_7_transpose_y_0, x = var_3077_cast_fp16, y = v_7_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [4]> var_3079_perm_0 = const()[name = string("op_3079_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3080 = const()[name = string("op_3080"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_3079_cast_fp16 = transpose(perm = var_3079_perm_0, x = out_7_cast_fp16)[name = string("transpose_451")];
            tensor<fp16, [1, 188, 1024]> input_163_cast_fp16 = reshape(shape = var_3080, x = var_3079_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_3_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95601408))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96650048))))[name = string("blocks_3_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_42_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_3_o_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_165_cast_fp16 = add(x = input_157_cast_fp16, y = linear_42_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<int32, [1]> input_167_axes_0 = const()[name = string("input_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_3_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96652160)))];
            tensor<fp16, [1024]> enc_layers_3_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96654272)))];
            tensor<fp16, [1, 188, 1024]> input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = enc_layers_3_norm_conv_bias_to_fp16, epsilon = var_3008_to_fp16, gamma = enc_layers_3_norm_conv_weight_to_fp16, x = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_3_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96656384))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98753600))))[name = string("blocks_3_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_43_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_3_pw1_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_43_cast_fp16")];
            int32 var_3089_split_num_splits_0 = const()[name = string("op_3089_split_num_splits_0"), val = int32(2)];
            int32 var_3089_split_axis_0 = const()[name = string("op_3089_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_3089_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_3089_split_cast_fp16_1 = split(axis = var_3089_split_axis_0, num_splits = var_3089_split_num_splits_0, x = linear_43_cast_fp16)[name = string("op_3089_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3089_split_1_sigmoid_cast_fp16 = sigmoid(x = var_3089_split_cast_fp16_1)[name = string("op_3089_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3089_cast_fp16 = mul(x = var_3089_split_cast_fp16_0, y = var_3089_split_1_sigmoid_cast_fp16)[name = string("op_3089_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_13_cast_fp16 = mul(x = var_3089_cast_fp16, y = valid_cast_fp16)[name = string("h_13_cast_fp16")];
            tensor<int32, [3]> input_171_perm_0 = const()[name = string("input_171_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_173_pad_type_0 = const()[name = string("input_173_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_173_pad_0 = const()[name = string("input_173_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_173_groups_0 = const()[name = string("input_173_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_173_strides_0 = const()[name = string("input_173_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_173_dilations_0 = const()[name = string("input_173_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_44_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98757760))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98767040))))[name = string("const_44_to_fp16_quantized")];
            tensor<fp16, [1024]> const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98769152)))];
            tensor<fp16, [1, 1024, 188]> input_171_cast_fp16 = transpose(perm = input_171_perm_0, x = h_13_cast_fp16)[name = string("transpose_450")];
            tensor<fp16, [1, 1024, 188]> input_175_cast_fp16 = conv(bias = const_45_to_fp16, dilations = input_173_dilations_0, groups = input_173_groups_0, pad = input_173_pad_0, pad_type = input_173_pad_type_0, strides = input_173_strides_0, weight = const_44_to_fp16_quantized, x = input_171_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_15_cast_fp16 = silu(x = input_175_cast_fp16)[name = string("h_15_cast_fp16")];
            tensor<int32, [3]> input_177_perm_0 = const()[name = string("input_177_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_3_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98771264))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99819904))))[name = string("blocks_3_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_177_cast_fp16 = transpose(perm = input_177_perm_0, x = h_15_cast_fp16)[name = string("transpose_449")];
            tensor<fp16, [1, 188, 1024]> linear_44_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_3_pw2_weight_to_fp16_quantized, x = input_177_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_179_cast_fp16 = add(x = input_165_cast_fp16, y = linear_44_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<int32, [1]> input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99822016)))];
            tensor<fp16, [1024]> enc_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99824128)))];
            tensor<fp16, [1, 188, 1024]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = enc_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_3008_to_fp16, gamma = enc_layers_3_norm_feed_forward2_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_3_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99826240))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104020608))))[name = string("blocks_3_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_3_ff2_1_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_185_cast_fp16 = silu(x = linear_45_cast_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_3_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104028864))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108223232))))[name = string("blocks_3_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_3_ff2_2_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_46_cast_fp16")];
            fp16 var_3110_to_fp16 = const()[name = string("op_3110_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3111_cast_fp16 = mul(x = linear_46_cast_fp16, y = var_3110_to_fp16)[name = string("op_3111_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_187_cast_fp16 = add(x = input_179_cast_fp16, y = var_3111_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<int32, [1]> input_189_axes_0 = const()[name = string("input_189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_3_norm_out_weight_to_fp16 = const()[name = string("enc_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108225344)))];
            tensor<fp16, [1024]> enc_layers_3_norm_out_bias_to_fp16 = const()[name = string("enc_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108227456)))];
            tensor<fp16, [1, 188, 1024]> input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = enc_layers_3_norm_out_bias_to_fp16, epsilon = var_3008_to_fp16, gamma = enc_layers_3_norm_out_weight_to_fp16, x = input_187_cast_fp16)[name = string("input_189_cast_fp16")];
            int32 var_3124 = const()[name = string("op_3124"), val = int32(-1)];
            tensor<int32, [1]> input_191_axes_0 = const()[name = string("input_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108229568)))];
            tensor<fp16, [1024]> enc_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108231680)))];
            fp16 var_3132_to_fp16 = const()[name = string("op_3132_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = enc_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_3132_to_fp16, gamma = enc_layers_4_norm_feed_forward1_weight_to_fp16, x = input_189_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_4_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108233792))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112428160))))[name = string("blocks_4_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_4_ff1_1_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_195_cast_fp16 = silu(x = linear_47_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_4_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112436416))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116630784))))[name = string("blocks_4_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_48_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_4_ff1_2_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_48_cast_fp16")];
            fp16 var_3152_to_fp16 = const()[name = string("op_3152_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3153_cast_fp16 = mul(x = linear_48_cast_fp16, y = var_3152_to_fp16)[name = string("op_3153_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_197_cast_fp16 = add(x = input_189_cast_fp16, y = var_3153_cast_fp16)[name = string("input_197_cast_fp16")];
            tensor<int32, [1]> input_199_axes_0 = const()[name = string("input_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116632896)))];
            tensor<fp16, [1024]> enc_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116635008)))];
            tensor<fp16, [1, 188, 1024]> input_199_cast_fp16 = layer_norm(axes = input_199_axes_0, beta = enc_layers_4_norm_self_att_bias_to_fp16, epsilon = var_3132_to_fp16, gamma = enc_layers_4_norm_self_att_weight_to_fp16, x = input_197_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_4_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116637120))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117685760))))[name = string("blocks_4_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_49_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_4_q_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_3159 = const()[name = string("op_3159"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3160_cast_fp16 = reshape(shape = var_3159, x = linear_49_cast_fp16)[name = string("op_3160_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_4_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117687872))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118736512))))[name = string("blocks_4_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_50_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_4_k_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_3164 = const()[name = string("op_3164"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3165_cast_fp16 = reshape(shape = var_3164, x = linear_50_cast_fp16)[name = string("op_3165_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_4_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118738624))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119787264))))[name = string("blocks_4_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_51_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_4_v_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_3169 = const()[name = string("op_3169"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3170_cast_fp16 = reshape(shape = var_3169, x = linear_51_cast_fp16)[name = string("op_3170_cast_fp16")];
            tensor<int32, [4]> v_9_perm_0 = const()[name = string("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_94_to_fp16 = const()[name = string("const_94_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119789376)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_9_cast_fp16 = add(x = var_3160_cast_fp16, y = const_94_to_fp16)[name = string("q_u_9_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_95_to_fp16 = const()[name = string("const_95_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119791488)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_9_cast_fp16 = add(x = var_3160_cast_fp16, y = const_95_to_fp16)[name = string("q_v_9_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_4_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119793600))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120842240))))[name = string("blocks_4_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_52_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_4_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<int32, [4]> var_3180 = const()[name = string("op_3180"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_9_cast_fp16 = reshape(shape = var_3180, x = linear_52_cast_fp16)[name = string("p_9_cast_fp16")];
            bool input_201_transpose_x_0 = const()[name = string("input_201_transpose_x_0"), val = bool(false)];
            bool input_201_transpose_y_0 = const()[name = string("input_201_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_208_perm_0 = const()[name = string("transpose_208_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_209_perm_0 = const()[name = string("transpose_209_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_209 = transpose(perm = transpose_209_perm_0, x = p_9_cast_fp16)[name = string("transpose_447")];
            tensor<fp16, [1, 8, 188, 128]> transpose_208 = transpose(perm = transpose_208_perm_0, x = q_v_9_cast_fp16)[name = string("transpose_448")];
            tensor<fp16, [1, 8, 188, 375]> input_201_cast_fp16 = matmul(transpose_x = input_201_transpose_x_0, transpose_y = input_201_transpose_y_0, x = transpose_208, y = transpose_209)[name = string("input_201_cast_fp16")];
            tensor<int32, [8]> var_3186_pad_0 = const()[name = string("op_3186_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_3186_mode_0 = const()[name = string("op_3186_mode_0"), val = string("constant")];
            fp16 const_18_to_fp16 = const()[name = string("const_18_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_3186_cast_fp16 = pad(constant_val = const_18_to_fp16, mode = var_3186_mode_0, pad = var_3186_pad_0, x = input_201_cast_fp16)[name = string("op_3186_cast_fp16")];
            tensor<int32, [4]> var_3187 = const()[name = string("op_3187"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_17_cast_fp16 = reshape(shape = var_3187, x = var_3186_cast_fp16)[name = string("scores_17_cast_fp16")];
            tensor<int32, [4]> var_3191_begin_0 = const()[name = string("op_3191_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3191_end_0 = const()[name = string("op_3191_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_3191_end_mask_0 = const()[name = string("op_3191_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_3191_cast_fp16 = slice_by_index(begin = var_3191_begin_0, end = var_3191_end_0, end_mask = var_3191_end_mask_0, x = scores_17_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> var_3192 = const()[name = string("op_3192"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_3193_cast_fp16 = reshape(shape = var_3192, x = var_3191_cast_fp16)[name = string("op_3193_cast_fp16")];
            tensor<int32, [4]> bd_9_begin_0 = const()[name = string("bd_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_9_end_0 = const()[name = string("bd_9_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_9_end_mask_0 = const()[name = string("bd_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_9_cast_fp16 = slice_by_index(begin = bd_9_begin_0, end = bd_9_end_0, end_mask = bd_9_end_mask_0, x = var_3193_cast_fp16)[name = string("bd_9_cast_fp16")];
            bool var_3196_transpose_x_0 = const()[name = string("op_3196_transpose_x_0"), val = bool(false)];
            bool var_3196_transpose_y_0 = const()[name = string("op_3196_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_210_perm_0 = const()[name = string("transpose_210_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_211_perm_0 = const()[name = string("transpose_211_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_211 = transpose(perm = transpose_211_perm_0, x = var_3165_cast_fp16)[name = string("transpose_445")];
            tensor<fp16, [1, 8, 188, 128]> transpose_210 = transpose(perm = transpose_210_perm_0, x = q_u_9_cast_fp16)[name = string("transpose_446")];
            tensor<fp16, [1, 8, 188, 188]> var_3196_cast_fp16 = matmul(transpose_x = var_3196_transpose_x_0, transpose_y = var_3196_transpose_y_0, x = transpose_210, y = transpose_211)[name = string("op_3196_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3197_cast_fp16 = add(x = var_3196_cast_fp16, y = bd_9_cast_fp16)[name = string("op_3197_cast_fp16")];
            fp16 var_3198_to_fp16 = const()[name = string("op_3198_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_3199_cast_fp16 = mul(x = var_3197_cast_fp16, y = var_3198_to_fp16)[name = string("op_3199_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_19_cast_fp16 = add(x = var_3199_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_19_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3201_cast_fp16 = softmax(axis = var_3124, x = scores_19_cast_fp16)[name = string("op_3201_cast_fp16")];
            bool out_9_transpose_x_0 = const()[name = string("out_9_transpose_x_0"), val = bool(false)];
            bool out_9_transpose_y_0 = const()[name = string("out_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_3170_cast_fp16)[name = string("transpose_444")];
            tensor<fp16, [1, 8, 188, 128]> out_9_cast_fp16 = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = var_3201_cast_fp16, y = v_9_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [4]> var_3203_perm_0 = const()[name = string("op_3203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3204 = const()[name = string("op_3204"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_3203_cast_fp16 = transpose(perm = var_3203_perm_0, x = out_9_cast_fp16)[name = string("transpose_443")];
            tensor<fp16, [1, 188, 1024]> input_203_cast_fp16 = reshape(shape = var_3204, x = var_3203_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_4_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120844352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121892992))))[name = string("blocks_4_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_4_o_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_205_cast_fp16 = add(x = input_197_cast_fp16, y = linear_53_cast_fp16)[name = string("input_205_cast_fp16")];
            tensor<int32, [1]> input_207_axes_0 = const()[name = string("input_207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_4_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121895104)))];
            tensor<fp16, [1024]> enc_layers_4_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121897216)))];
            tensor<fp16, [1, 188, 1024]> input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = enc_layers_4_norm_conv_bias_to_fp16, epsilon = var_3132_to_fp16, gamma = enc_layers_4_norm_conv_weight_to_fp16, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_4_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121899328))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123996544))))[name = string("blocks_4_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_54_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_4_pw1_weight_to_fp16_quantized, x = input_207_cast_fp16)[name = string("linear_54_cast_fp16")];
            int32 var_3213_split_num_splits_0 = const()[name = string("op_3213_split_num_splits_0"), val = int32(2)];
            int32 var_3213_split_axis_0 = const()[name = string("op_3213_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_3213_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_3213_split_cast_fp16_1 = split(axis = var_3213_split_axis_0, num_splits = var_3213_split_num_splits_0, x = linear_54_cast_fp16)[name = string("op_3213_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3213_split_1_sigmoid_cast_fp16 = sigmoid(x = var_3213_split_cast_fp16_1)[name = string("op_3213_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3213_cast_fp16 = mul(x = var_3213_split_cast_fp16_0, y = var_3213_split_1_sigmoid_cast_fp16)[name = string("op_3213_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_17_cast_fp16 = mul(x = var_3213_cast_fp16, y = valid_cast_fp16)[name = string("h_17_cast_fp16")];
            tensor<int32, [3]> input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_213_pad_type_0 = const()[name = string("input_213_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_213_pad_0 = const()[name = string("input_213_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_213_groups_0 = const()[name = string("input_213_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_213_strides_0 = const()[name = string("input_213_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_213_dilations_0 = const()[name = string("input_213_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_46_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124000704))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124009984))))[name = string("const_46_to_fp16_quantized")];
            tensor<fp16, [1024]> const_47_to_fp16 = const()[name = string("const_47_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124012096)))];
            tensor<fp16, [1, 1024, 188]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = h_17_cast_fp16)[name = string("transpose_442")];
            tensor<fp16, [1, 1024, 188]> input_215_cast_fp16 = conv(bias = const_47_to_fp16, dilations = input_213_dilations_0, groups = input_213_groups_0, pad = input_213_pad_0, pad_type = input_213_pad_type_0, strides = input_213_strides_0, weight = const_46_to_fp16_quantized, x = input_211_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_19_cast_fp16 = silu(x = input_215_cast_fp16)[name = string("h_19_cast_fp16")];
            tensor<int32, [3]> input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_4_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124014208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125062848))))[name = string("blocks_4_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = h_19_cast_fp16)[name = string("transpose_441")];
            tensor<fp16, [1, 188, 1024]> linear_55_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_4_pw2_weight_to_fp16_quantized, x = input_217_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_219_cast_fp16 = add(x = input_205_cast_fp16, y = linear_55_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125064960)))];
            tensor<fp16, [1024]> enc_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125067072)))];
            tensor<fp16, [1, 188, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = enc_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_3132_to_fp16, gamma = enc_layers_4_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_4_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125069184))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129263552))))[name = string("blocks_4_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_56_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_4_ff2_1_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_225_cast_fp16 = silu(x = linear_56_cast_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_4_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129271808))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133466176))))[name = string("blocks_4_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_57_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_4_ff2_2_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_57_cast_fp16")];
            fp16 var_3234_to_fp16 = const()[name = string("op_3234_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3235_cast_fp16 = mul(x = linear_57_cast_fp16, y = var_3234_to_fp16)[name = string("op_3235_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_227_cast_fp16 = add(x = input_219_cast_fp16, y = var_3235_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<int32, [1]> input_229_axes_0 = const()[name = string("input_229_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_4_norm_out_weight_to_fp16 = const()[name = string("enc_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133468288)))];
            tensor<fp16, [1024]> enc_layers_4_norm_out_bias_to_fp16 = const()[name = string("enc_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133470400)))];
            tensor<fp16, [1, 188, 1024]> input_229_cast_fp16 = layer_norm(axes = input_229_axes_0, beta = enc_layers_4_norm_out_bias_to_fp16, epsilon = var_3132_to_fp16, gamma = enc_layers_4_norm_out_weight_to_fp16, x = input_227_cast_fp16)[name = string("input_229_cast_fp16")];
            int32 var_3248 = const()[name = string("op_3248"), val = int32(-1)];
            tensor<int32, [1]> input_231_axes_0 = const()[name = string("input_231_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133472512)))];
            tensor<fp16, [1024]> enc_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133474624)))];
            fp16 var_3256_to_fp16 = const()[name = string("op_3256_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = enc_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_3256_to_fp16, gamma = enc_layers_5_norm_feed_forward1_weight_to_fp16, x = input_229_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_5_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133476736))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137671104))))[name = string("blocks_5_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_58_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_5_ff1_1_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_235_cast_fp16 = silu(x = linear_58_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_5_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137679360))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141873728))))[name = string("blocks_5_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_59_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_5_ff1_2_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_59_cast_fp16")];
            fp16 var_3276_to_fp16 = const()[name = string("op_3276_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3277_cast_fp16 = mul(x = linear_59_cast_fp16, y = var_3276_to_fp16)[name = string("op_3277_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_237_cast_fp16 = add(x = input_229_cast_fp16, y = var_3277_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<int32, [1]> input_239_axes_0 = const()[name = string("input_239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141875840)))];
            tensor<fp16, [1024]> enc_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141877952)))];
            tensor<fp16, [1, 188, 1024]> input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = enc_layers_5_norm_self_att_bias_to_fp16, epsilon = var_3256_to_fp16, gamma = enc_layers_5_norm_self_att_weight_to_fp16, x = input_237_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_5_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141880064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142928704))))[name = string("blocks_5_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_5_q_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<int32, [4]> var_3283 = const()[name = string("op_3283"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3284_cast_fp16 = reshape(shape = var_3283, x = linear_60_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_5_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142930816))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143979456))))[name = string("blocks_5_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_5_k_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<int32, [4]> var_3288 = const()[name = string("op_3288"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3289_cast_fp16 = reshape(shape = var_3288, x = linear_61_cast_fp16)[name = string("op_3289_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_5_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143981568))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145030208))))[name = string("blocks_5_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_62_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_5_v_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<int32, [4]> var_3293 = const()[name = string("op_3293"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3294_cast_fp16 = reshape(shape = var_3293, x = linear_62_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_96_to_fp16 = const()[name = string("const_96_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145032320)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_11_cast_fp16 = add(x = var_3284_cast_fp16, y = const_96_to_fp16)[name = string("q_u_11_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_97_to_fp16 = const()[name = string("const_97_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145034432)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_11_cast_fp16 = add(x = var_3284_cast_fp16, y = const_97_to_fp16)[name = string("q_v_11_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_5_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145036544))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146085184))))[name = string("blocks_5_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_63_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_5_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_3304 = const()[name = string("op_3304"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_11_cast_fp16 = reshape(shape = var_3304, x = linear_63_cast_fp16)[name = string("p_11_cast_fp16")];
            bool input_241_transpose_x_0 = const()[name = string("input_241_transpose_x_0"), val = bool(false)];
            bool input_241_transpose_y_0 = const()[name = string("input_241_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_212_perm_0 = const()[name = string("transpose_212_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_213_perm_0 = const()[name = string("transpose_213_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_213 = transpose(perm = transpose_213_perm_0, x = p_11_cast_fp16)[name = string("transpose_439")];
            tensor<fp16, [1, 8, 188, 128]> transpose_212 = transpose(perm = transpose_212_perm_0, x = q_v_11_cast_fp16)[name = string("transpose_440")];
            tensor<fp16, [1, 8, 188, 375]> input_241_cast_fp16 = matmul(transpose_x = input_241_transpose_x_0, transpose_y = input_241_transpose_y_0, x = transpose_212, y = transpose_213)[name = string("input_241_cast_fp16")];
            tensor<int32, [8]> var_3310_pad_0 = const()[name = string("op_3310_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_3310_mode_0 = const()[name = string("op_3310_mode_0"), val = string("constant")];
            fp16 const_19_to_fp16 = const()[name = string("const_19_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_3310_cast_fp16 = pad(constant_val = const_19_to_fp16, mode = var_3310_mode_0, pad = var_3310_pad_0, x = input_241_cast_fp16)[name = string("op_3310_cast_fp16")];
            tensor<int32, [4]> var_3311 = const()[name = string("op_3311"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_21_cast_fp16 = reshape(shape = var_3311, x = var_3310_cast_fp16)[name = string("scores_21_cast_fp16")];
            tensor<int32, [4]> var_3315_begin_0 = const()[name = string("op_3315_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3315_end_0 = const()[name = string("op_3315_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_3315_end_mask_0 = const()[name = string("op_3315_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_3315_cast_fp16 = slice_by_index(begin = var_3315_begin_0, end = var_3315_end_0, end_mask = var_3315_end_mask_0, x = scores_21_cast_fp16)[name = string("op_3315_cast_fp16")];
            tensor<int32, [4]> var_3316 = const()[name = string("op_3316"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_3317_cast_fp16 = reshape(shape = var_3316, x = var_3315_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<int32, [4]> bd_11_begin_0 = const()[name = string("bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_11_end_0 = const()[name = string("bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_11_end_mask_0 = const()[name = string("bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_11_cast_fp16 = slice_by_index(begin = bd_11_begin_0, end = bd_11_end_0, end_mask = bd_11_end_mask_0, x = var_3317_cast_fp16)[name = string("bd_11_cast_fp16")];
            bool var_3320_transpose_x_0 = const()[name = string("op_3320_transpose_x_0"), val = bool(false)];
            bool var_3320_transpose_y_0 = const()[name = string("op_3320_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_214_perm_0 = const()[name = string("transpose_214_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_215_perm_0 = const()[name = string("transpose_215_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_215 = transpose(perm = transpose_215_perm_0, x = var_3289_cast_fp16)[name = string("transpose_437")];
            tensor<fp16, [1, 8, 188, 128]> transpose_214 = transpose(perm = transpose_214_perm_0, x = q_u_11_cast_fp16)[name = string("transpose_438")];
            tensor<fp16, [1, 8, 188, 188]> var_3320_cast_fp16 = matmul(transpose_x = var_3320_transpose_x_0, transpose_y = var_3320_transpose_y_0, x = transpose_214, y = transpose_215)[name = string("op_3320_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3321_cast_fp16 = add(x = var_3320_cast_fp16, y = bd_11_cast_fp16)[name = string("op_3321_cast_fp16")];
            fp16 var_3322_to_fp16 = const()[name = string("op_3322_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_3323_cast_fp16 = mul(x = var_3321_cast_fp16, y = var_3322_to_fp16)[name = string("op_3323_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_23_cast_fp16 = add(x = var_3323_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_23_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3325_cast_fp16 = softmax(axis = var_3248, x = scores_23_cast_fp16)[name = string("op_3325_cast_fp16")];
            bool out_11_transpose_x_0 = const()[name = string("out_11_transpose_x_0"), val = bool(false)];
            bool out_11_transpose_y_0 = const()[name = string("out_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_3294_cast_fp16)[name = string("transpose_436")];
            tensor<fp16, [1, 8, 188, 128]> out_11_cast_fp16 = matmul(transpose_x = out_11_transpose_x_0, transpose_y = out_11_transpose_y_0, x = var_3325_cast_fp16, y = v_11_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [4]> var_3327_perm_0 = const()[name = string("op_3327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3328 = const()[name = string("op_3328"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_3327_cast_fp16 = transpose(perm = var_3327_perm_0, x = out_11_cast_fp16)[name = string("transpose_435")];
            tensor<fp16, [1, 188, 1024]> input_243_cast_fp16 = reshape(shape = var_3328, x = var_3327_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_5_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146087296))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147135936))))[name = string("blocks_5_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_64_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_5_o_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_245_cast_fp16 = add(x = input_237_cast_fp16, y = linear_64_cast_fp16)[name = string("input_245_cast_fp16")];
            tensor<int32, [1]> input_247_axes_0 = const()[name = string("input_247_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_5_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147138048)))];
            tensor<fp16, [1024]> enc_layers_5_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147140160)))];
            tensor<fp16, [1, 188, 1024]> input_247_cast_fp16 = layer_norm(axes = input_247_axes_0, beta = enc_layers_5_norm_conv_bias_to_fp16, epsilon = var_3256_to_fp16, gamma = enc_layers_5_norm_conv_weight_to_fp16, x = input_245_cast_fp16)[name = string("input_247_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_5_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147142272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149239488))))[name = string("blocks_5_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_65_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_5_pw1_weight_to_fp16_quantized, x = input_247_cast_fp16)[name = string("linear_65_cast_fp16")];
            int32 var_3337_split_num_splits_0 = const()[name = string("op_3337_split_num_splits_0"), val = int32(2)];
            int32 var_3337_split_axis_0 = const()[name = string("op_3337_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_3337_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_3337_split_cast_fp16_1 = split(axis = var_3337_split_axis_0, num_splits = var_3337_split_num_splits_0, x = linear_65_cast_fp16)[name = string("op_3337_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3337_split_1_sigmoid_cast_fp16 = sigmoid(x = var_3337_split_cast_fp16_1)[name = string("op_3337_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3337_cast_fp16 = mul(x = var_3337_split_cast_fp16_0, y = var_3337_split_1_sigmoid_cast_fp16)[name = string("op_3337_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_21_cast_fp16 = mul(x = var_3337_cast_fp16, y = valid_cast_fp16)[name = string("h_21_cast_fp16")];
            tensor<int32, [3]> input_251_perm_0 = const()[name = string("input_251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_253_pad_type_0 = const()[name = string("input_253_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_253_pad_0 = const()[name = string("input_253_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_253_groups_0 = const()[name = string("input_253_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_253_strides_0 = const()[name = string("input_253_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_253_dilations_0 = const()[name = string("input_253_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_48_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149243648))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149252928))))[name = string("const_48_to_fp16_quantized")];
            tensor<fp16, [1024]> const_49_to_fp16 = const()[name = string("const_49_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149255040)))];
            tensor<fp16, [1, 1024, 188]> input_251_cast_fp16 = transpose(perm = input_251_perm_0, x = h_21_cast_fp16)[name = string("transpose_434")];
            tensor<fp16, [1, 1024, 188]> input_255_cast_fp16 = conv(bias = const_49_to_fp16, dilations = input_253_dilations_0, groups = input_253_groups_0, pad = input_253_pad_0, pad_type = input_253_pad_type_0, strides = input_253_strides_0, weight = const_48_to_fp16_quantized, x = input_251_cast_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_23_cast_fp16 = silu(x = input_255_cast_fp16)[name = string("h_23_cast_fp16")];
            tensor<int32, [3]> input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_5_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149257152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150305792))))[name = string("blocks_5_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = h_23_cast_fp16)[name = string("transpose_433")];
            tensor<fp16, [1, 188, 1024]> linear_66_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_5_pw2_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_259_cast_fp16 = add(x = input_245_cast_fp16, y = linear_66_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<int32, [1]> input_261_axes_0 = const()[name = string("input_261_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150307904)))];
            tensor<fp16, [1024]> enc_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150310016)))];
            tensor<fp16, [1, 188, 1024]> input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = enc_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_3256_to_fp16, gamma = enc_layers_5_norm_feed_forward2_weight_to_fp16, x = input_259_cast_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_5_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150312128))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154506496))))[name = string("blocks_5_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_5_ff2_1_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_265_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_5_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154514752))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158709120))))[name = string("blocks_5_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_5_ff2_2_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_68_cast_fp16")];
            fp16 var_3358_to_fp16 = const()[name = string("op_3358_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3359_cast_fp16 = mul(x = linear_68_cast_fp16, y = var_3358_to_fp16)[name = string("op_3359_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_267_cast_fp16 = add(x = input_259_cast_fp16, y = var_3359_cast_fp16)[name = string("input_267_cast_fp16")];
            tensor<int32, [1]> input_269_axes_0 = const()[name = string("input_269_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_5_norm_out_weight_to_fp16 = const()[name = string("enc_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158711232)))];
            tensor<fp16, [1024]> enc_layers_5_norm_out_bias_to_fp16 = const()[name = string("enc_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158713344)))];
            tensor<fp16, [1, 188, 1024]> input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = enc_layers_5_norm_out_bias_to_fp16, epsilon = var_3256_to_fp16, gamma = enc_layers_5_norm_out_weight_to_fp16, x = input_267_cast_fp16)[name = string("input_269_cast_fp16")];
            int32 var_3372 = const()[name = string("op_3372"), val = int32(-1)];
            tensor<int32, [1]> input_271_axes_0 = const()[name = string("input_271_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158715456)))];
            tensor<fp16, [1024]> enc_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158717568)))];
            fp16 var_3380_to_fp16 = const()[name = string("op_3380_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_271_cast_fp16 = layer_norm(axes = input_271_axes_0, beta = enc_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_3380_to_fp16, gamma = enc_layers_6_norm_feed_forward1_weight_to_fp16, x = input_269_cast_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_6_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158719680))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162914048))))[name = string("blocks_6_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_6_ff1_1_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_275_cast_fp16 = silu(x = linear_69_cast_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_6_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(162922304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167116672))))[name = string("blocks_6_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_70_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_6_ff1_2_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_70_cast_fp16")];
            fp16 var_3400_to_fp16 = const()[name = string("op_3400_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3401_cast_fp16 = mul(x = linear_70_cast_fp16, y = var_3400_to_fp16)[name = string("op_3401_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_277_cast_fp16 = add(x = input_269_cast_fp16, y = var_3401_cast_fp16)[name = string("input_277_cast_fp16")];
            tensor<int32, [1]> input_279_axes_0 = const()[name = string("input_279_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167118784)))];
            tensor<fp16, [1024]> enc_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167120896)))];
            tensor<fp16, [1, 188, 1024]> input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = enc_layers_6_norm_self_att_bias_to_fp16, epsilon = var_3380_to_fp16, gamma = enc_layers_6_norm_self_att_weight_to_fp16, x = input_277_cast_fp16)[name = string("input_279_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_6_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167123008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168171648))))[name = string("blocks_6_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_71_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_6_q_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_3407 = const()[name = string("op_3407"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3408_cast_fp16 = reshape(shape = var_3407, x = linear_71_cast_fp16)[name = string("op_3408_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_6_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168173760))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169222400))))[name = string("blocks_6_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_72_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_6_k_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_3412 = const()[name = string("op_3412"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3413_cast_fp16 = reshape(shape = var_3412, x = linear_72_cast_fp16)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_6_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169224512))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170273152))))[name = string("blocks_6_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_73_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_6_v_weight_to_fp16_quantized, x = input_279_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_3417 = const()[name = string("op_3417"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3418_cast_fp16 = reshape(shape = var_3417, x = linear_73_cast_fp16)[name = string("op_3418_cast_fp16")];
            tensor<int32, [4]> v_13_perm_0 = const()[name = string("v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_98_to_fp16 = const()[name = string("const_98_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170275264)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_13_cast_fp16 = add(x = var_3408_cast_fp16, y = const_98_to_fp16)[name = string("q_u_13_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170277376)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_13_cast_fp16 = add(x = var_3408_cast_fp16, y = const_99_to_fp16)[name = string("q_v_13_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_6_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170279488))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171328128))))[name = string("blocks_6_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_6_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<int32, [4]> var_3428 = const()[name = string("op_3428"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_13_cast_fp16 = reshape(shape = var_3428, x = linear_74_cast_fp16)[name = string("p_13_cast_fp16")];
            bool input_281_transpose_x_0 = const()[name = string("input_281_transpose_x_0"), val = bool(false)];
            bool input_281_transpose_y_0 = const()[name = string("input_281_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_216_perm_0 = const()[name = string("transpose_216_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_217_perm_0 = const()[name = string("transpose_217_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_217 = transpose(perm = transpose_217_perm_0, x = p_13_cast_fp16)[name = string("transpose_431")];
            tensor<fp16, [1, 8, 188, 128]> transpose_216 = transpose(perm = transpose_216_perm_0, x = q_v_13_cast_fp16)[name = string("transpose_432")];
            tensor<fp16, [1, 8, 188, 375]> input_281_cast_fp16 = matmul(transpose_x = input_281_transpose_x_0, transpose_y = input_281_transpose_y_0, x = transpose_216, y = transpose_217)[name = string("input_281_cast_fp16")];
            tensor<int32, [8]> var_3434_pad_0 = const()[name = string("op_3434_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_3434_mode_0 = const()[name = string("op_3434_mode_0"), val = string("constant")];
            fp16 const_20_to_fp16 = const()[name = string("const_20_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_3434_cast_fp16 = pad(constant_val = const_20_to_fp16, mode = var_3434_mode_0, pad = var_3434_pad_0, x = input_281_cast_fp16)[name = string("op_3434_cast_fp16")];
            tensor<int32, [4]> var_3435 = const()[name = string("op_3435"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_25_cast_fp16 = reshape(shape = var_3435, x = var_3434_cast_fp16)[name = string("scores_25_cast_fp16")];
            tensor<int32, [4]> var_3439_begin_0 = const()[name = string("op_3439_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3439_end_0 = const()[name = string("op_3439_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_3439_end_mask_0 = const()[name = string("op_3439_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_3439_cast_fp16 = slice_by_index(begin = var_3439_begin_0, end = var_3439_end_0, end_mask = var_3439_end_mask_0, x = scores_25_cast_fp16)[name = string("op_3439_cast_fp16")];
            tensor<int32, [4]> var_3440 = const()[name = string("op_3440"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_3441_cast_fp16 = reshape(shape = var_3440, x = var_3439_cast_fp16)[name = string("op_3441_cast_fp16")];
            tensor<int32, [4]> bd_13_begin_0 = const()[name = string("bd_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_13_end_0 = const()[name = string("bd_13_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_13_end_mask_0 = const()[name = string("bd_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_13_cast_fp16 = slice_by_index(begin = bd_13_begin_0, end = bd_13_end_0, end_mask = bd_13_end_mask_0, x = var_3441_cast_fp16)[name = string("bd_13_cast_fp16")];
            bool var_3444_transpose_x_0 = const()[name = string("op_3444_transpose_x_0"), val = bool(false)];
            bool var_3444_transpose_y_0 = const()[name = string("op_3444_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_218_perm_0 = const()[name = string("transpose_218_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_219_perm_0 = const()[name = string("transpose_219_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_219 = transpose(perm = transpose_219_perm_0, x = var_3413_cast_fp16)[name = string("transpose_429")];
            tensor<fp16, [1, 8, 188, 128]> transpose_218 = transpose(perm = transpose_218_perm_0, x = q_u_13_cast_fp16)[name = string("transpose_430")];
            tensor<fp16, [1, 8, 188, 188]> var_3444_cast_fp16 = matmul(transpose_x = var_3444_transpose_x_0, transpose_y = var_3444_transpose_y_0, x = transpose_218, y = transpose_219)[name = string("op_3444_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3445_cast_fp16 = add(x = var_3444_cast_fp16, y = bd_13_cast_fp16)[name = string("op_3445_cast_fp16")];
            fp16 var_3446_to_fp16 = const()[name = string("op_3446_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_3447_cast_fp16 = mul(x = var_3445_cast_fp16, y = var_3446_to_fp16)[name = string("op_3447_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_27_cast_fp16 = add(x = var_3447_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_27_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3449_cast_fp16 = softmax(axis = var_3372, x = scores_27_cast_fp16)[name = string("op_3449_cast_fp16")];
            bool out_13_transpose_x_0 = const()[name = string("out_13_transpose_x_0"), val = bool(false)];
            bool out_13_transpose_y_0 = const()[name = string("out_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_3418_cast_fp16)[name = string("transpose_428")];
            tensor<fp16, [1, 8, 188, 128]> out_13_cast_fp16 = matmul(transpose_x = out_13_transpose_x_0, transpose_y = out_13_transpose_y_0, x = var_3449_cast_fp16, y = v_13_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [4]> var_3451_perm_0 = const()[name = string("op_3451_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3452 = const()[name = string("op_3452"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_3451_cast_fp16 = transpose(perm = var_3451_perm_0, x = out_13_cast_fp16)[name = string("transpose_427")];
            tensor<fp16, [1, 188, 1024]> input_283_cast_fp16 = reshape(shape = var_3452, x = var_3451_cast_fp16)[name = string("input_283_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_6_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171330240))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172378880))))[name = string("blocks_6_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_6_o_weight_to_fp16_quantized, x = input_283_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_285_cast_fp16 = add(x = input_277_cast_fp16, y = linear_75_cast_fp16)[name = string("input_285_cast_fp16")];
            tensor<int32, [1]> input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_6_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172380992)))];
            tensor<fp16, [1024]> enc_layers_6_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172383104)))];
            tensor<fp16, [1, 188, 1024]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = enc_layers_6_norm_conv_bias_to_fp16, epsilon = var_3380_to_fp16, gamma = enc_layers_6_norm_conv_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_6_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172385216))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174482432))))[name = string("blocks_6_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_76_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_6_pw1_weight_to_fp16_quantized, x = input_287_cast_fp16)[name = string("linear_76_cast_fp16")];
            int32 var_3461_split_num_splits_0 = const()[name = string("op_3461_split_num_splits_0"), val = int32(2)];
            int32 var_3461_split_axis_0 = const()[name = string("op_3461_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_3461_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_3461_split_cast_fp16_1 = split(axis = var_3461_split_axis_0, num_splits = var_3461_split_num_splits_0, x = linear_76_cast_fp16)[name = string("op_3461_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3461_split_1_sigmoid_cast_fp16 = sigmoid(x = var_3461_split_cast_fp16_1)[name = string("op_3461_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3461_cast_fp16 = mul(x = var_3461_split_cast_fp16_0, y = var_3461_split_1_sigmoid_cast_fp16)[name = string("op_3461_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_25_cast_fp16 = mul(x = var_3461_cast_fp16, y = valid_cast_fp16)[name = string("h_25_cast_fp16")];
            tensor<int32, [3]> input_291_perm_0 = const()[name = string("input_291_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_293_pad_type_0 = const()[name = string("input_293_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_293_pad_0 = const()[name = string("input_293_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_293_groups_0 = const()[name = string("input_293_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_293_strides_0 = const()[name = string("input_293_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_293_dilations_0 = const()[name = string("input_293_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_50_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174486592))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174495872))))[name = string("const_50_to_fp16_quantized")];
            tensor<fp16, [1024]> const_51_to_fp16 = const()[name = string("const_51_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174497984)))];
            tensor<fp16, [1, 1024, 188]> input_291_cast_fp16 = transpose(perm = input_291_perm_0, x = h_25_cast_fp16)[name = string("transpose_426")];
            tensor<fp16, [1, 1024, 188]> input_295_cast_fp16 = conv(bias = const_51_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_50_to_fp16_quantized, x = input_291_cast_fp16)[name = string("input_295_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_27_cast_fp16 = silu(x = input_295_cast_fp16)[name = string("h_27_cast_fp16")];
            tensor<int32, [3]> input_297_perm_0 = const()[name = string("input_297_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_6_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174500096))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175548736))))[name = string("blocks_6_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_297_cast_fp16 = transpose(perm = input_297_perm_0, x = h_27_cast_fp16)[name = string("transpose_425")];
            tensor<fp16, [1, 188, 1024]> linear_77_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_6_pw2_weight_to_fp16_quantized, x = input_297_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_299_cast_fp16 = add(x = input_285_cast_fp16, y = linear_77_cast_fp16)[name = string("input_299_cast_fp16")];
            tensor<int32, [1]> input_301_axes_0 = const()[name = string("input_301_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175550848)))];
            tensor<fp16, [1024]> enc_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175552960)))];
            tensor<fp16, [1, 188, 1024]> input_301_cast_fp16 = layer_norm(axes = input_301_axes_0, beta = enc_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_3380_to_fp16, gamma = enc_layers_6_norm_feed_forward2_weight_to_fp16, x = input_299_cast_fp16)[name = string("input_301_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_6_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175555072))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179749440))))[name = string("blocks_6_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_78_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_6_ff2_1_weight_to_fp16_quantized, x = input_301_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_305_cast_fp16 = silu(x = linear_78_cast_fp16)[name = string("input_305_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_6_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179757696))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183952064))))[name = string("blocks_6_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_79_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_6_ff2_2_weight_to_fp16_quantized, x = input_305_cast_fp16)[name = string("linear_79_cast_fp16")];
            fp16 var_3482_to_fp16 = const()[name = string("op_3482_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3483_cast_fp16 = mul(x = linear_79_cast_fp16, y = var_3482_to_fp16)[name = string("op_3483_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_307_cast_fp16 = add(x = input_299_cast_fp16, y = var_3483_cast_fp16)[name = string("input_307_cast_fp16")];
            tensor<int32, [1]> input_309_axes_0 = const()[name = string("input_309_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_6_norm_out_weight_to_fp16 = const()[name = string("enc_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183954176)))];
            tensor<fp16, [1024]> enc_layers_6_norm_out_bias_to_fp16 = const()[name = string("enc_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183956288)))];
            tensor<fp16, [1, 188, 1024]> input_309_cast_fp16 = layer_norm(axes = input_309_axes_0, beta = enc_layers_6_norm_out_bias_to_fp16, epsilon = var_3380_to_fp16, gamma = enc_layers_6_norm_out_weight_to_fp16, x = input_307_cast_fp16)[name = string("input_309_cast_fp16")];
            int32 var_3496 = const()[name = string("op_3496"), val = int32(-1)];
            tensor<int32, [1]> input_311_axes_0 = const()[name = string("input_311_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183958400)))];
            tensor<fp16, [1024]> enc_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183960512)))];
            fp16 var_3504_to_fp16 = const()[name = string("op_3504_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_311_cast_fp16 = layer_norm(axes = input_311_axes_0, beta = enc_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_3504_to_fp16, gamma = enc_layers_7_norm_feed_forward1_weight_to_fp16, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_7_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183962624))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188156992))))[name = string("blocks_7_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_7_ff1_1_weight_to_fp16_quantized, x = input_311_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_315_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_315_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_7_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188165248))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192359616))))[name = string("blocks_7_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_7_ff1_2_weight_to_fp16_quantized, x = input_315_cast_fp16)[name = string("linear_81_cast_fp16")];
            fp16 var_3524_to_fp16 = const()[name = string("op_3524_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3525_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_3524_to_fp16)[name = string("op_3525_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_317_cast_fp16 = add(x = input_309_cast_fp16, y = var_3525_cast_fp16)[name = string("input_317_cast_fp16")];
            tensor<int32, [1]> input_319_axes_0 = const()[name = string("input_319_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192361728)))];
            tensor<fp16, [1024]> enc_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192363840)))];
            tensor<fp16, [1, 188, 1024]> input_319_cast_fp16 = layer_norm(axes = input_319_axes_0, beta = enc_layers_7_norm_self_att_bias_to_fp16, epsilon = var_3504_to_fp16, gamma = enc_layers_7_norm_self_att_weight_to_fp16, x = input_317_cast_fp16)[name = string("input_319_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_7_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192365952))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193414592))))[name = string("blocks_7_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_7_q_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<int32, [4]> var_3531 = const()[name = string("op_3531"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3532_cast_fp16 = reshape(shape = var_3531, x = linear_82_cast_fp16)[name = string("op_3532_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_7_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193416704))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194465344))))[name = string("blocks_7_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_83_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_7_k_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<int32, [4]> var_3536 = const()[name = string("op_3536"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3537_cast_fp16 = reshape(shape = var_3536, x = linear_83_cast_fp16)[name = string("op_3537_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_7_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194467456))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195516096))))[name = string("blocks_7_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_84_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_7_v_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_3541 = const()[name = string("op_3541"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3542_cast_fp16 = reshape(shape = var_3541, x = linear_84_cast_fp16)[name = string("op_3542_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195518208)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_15_cast_fp16 = add(x = var_3532_cast_fp16, y = const_100_to_fp16)[name = string("q_u_15_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_101_to_fp16 = const()[name = string("const_101_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195520320)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_15_cast_fp16 = add(x = var_3532_cast_fp16, y = const_101_to_fp16)[name = string("q_v_15_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_7_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195522432))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196571072))))[name = string("blocks_7_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_85_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_7_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_3552 = const()[name = string("op_3552"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_15_cast_fp16 = reshape(shape = var_3552, x = linear_85_cast_fp16)[name = string("p_15_cast_fp16")];
            bool input_321_transpose_x_0 = const()[name = string("input_321_transpose_x_0"), val = bool(false)];
            bool input_321_transpose_y_0 = const()[name = string("input_321_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_220_perm_0 = const()[name = string("transpose_220_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_221_perm_0 = const()[name = string("transpose_221_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_221 = transpose(perm = transpose_221_perm_0, x = p_15_cast_fp16)[name = string("transpose_423")];
            tensor<fp16, [1, 8, 188, 128]> transpose_220 = transpose(perm = transpose_220_perm_0, x = q_v_15_cast_fp16)[name = string("transpose_424")];
            tensor<fp16, [1, 8, 188, 375]> input_321_cast_fp16 = matmul(transpose_x = input_321_transpose_x_0, transpose_y = input_321_transpose_y_0, x = transpose_220, y = transpose_221)[name = string("input_321_cast_fp16")];
            tensor<int32, [8]> var_3558_pad_0 = const()[name = string("op_3558_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_3558_mode_0 = const()[name = string("op_3558_mode_0"), val = string("constant")];
            fp16 const_21_to_fp16 = const()[name = string("const_21_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_3558_cast_fp16 = pad(constant_val = const_21_to_fp16, mode = var_3558_mode_0, pad = var_3558_pad_0, x = input_321_cast_fp16)[name = string("op_3558_cast_fp16")];
            tensor<int32, [4]> var_3559 = const()[name = string("op_3559"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_29_cast_fp16 = reshape(shape = var_3559, x = var_3558_cast_fp16)[name = string("scores_29_cast_fp16")];
            tensor<int32, [4]> var_3563_begin_0 = const()[name = string("op_3563_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3563_end_0 = const()[name = string("op_3563_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_3563_end_mask_0 = const()[name = string("op_3563_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_3563_cast_fp16 = slice_by_index(begin = var_3563_begin_0, end = var_3563_end_0, end_mask = var_3563_end_mask_0, x = scores_29_cast_fp16)[name = string("op_3563_cast_fp16")];
            tensor<int32, [4]> var_3564 = const()[name = string("op_3564"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_3565_cast_fp16 = reshape(shape = var_3564, x = var_3563_cast_fp16)[name = string("op_3565_cast_fp16")];
            tensor<int32, [4]> bd_15_begin_0 = const()[name = string("bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_15_end_0 = const()[name = string("bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_15_end_mask_0 = const()[name = string("bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_15_cast_fp16 = slice_by_index(begin = bd_15_begin_0, end = bd_15_end_0, end_mask = bd_15_end_mask_0, x = var_3565_cast_fp16)[name = string("bd_15_cast_fp16")];
            bool var_3568_transpose_x_0 = const()[name = string("op_3568_transpose_x_0"), val = bool(false)];
            bool var_3568_transpose_y_0 = const()[name = string("op_3568_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_222_perm_0 = const()[name = string("transpose_222_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_223_perm_0 = const()[name = string("transpose_223_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_223 = transpose(perm = transpose_223_perm_0, x = var_3537_cast_fp16)[name = string("transpose_421")];
            tensor<fp16, [1, 8, 188, 128]> transpose_222 = transpose(perm = transpose_222_perm_0, x = q_u_15_cast_fp16)[name = string("transpose_422")];
            tensor<fp16, [1, 8, 188, 188]> var_3568_cast_fp16 = matmul(transpose_x = var_3568_transpose_x_0, transpose_y = var_3568_transpose_y_0, x = transpose_222, y = transpose_223)[name = string("op_3568_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3569_cast_fp16 = add(x = var_3568_cast_fp16, y = bd_15_cast_fp16)[name = string("op_3569_cast_fp16")];
            fp16 var_3570_to_fp16 = const()[name = string("op_3570_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_3571_cast_fp16 = mul(x = var_3569_cast_fp16, y = var_3570_to_fp16)[name = string("op_3571_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_31_cast_fp16 = add(x = var_3571_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_31_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3573_cast_fp16 = softmax(axis = var_3496, x = scores_31_cast_fp16)[name = string("op_3573_cast_fp16")];
            bool out_15_transpose_x_0 = const()[name = string("out_15_transpose_x_0"), val = bool(false)];
            bool out_15_transpose_y_0 = const()[name = string("out_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_3542_cast_fp16)[name = string("transpose_420")];
            tensor<fp16, [1, 8, 188, 128]> out_15_cast_fp16 = matmul(transpose_x = out_15_transpose_x_0, transpose_y = out_15_transpose_y_0, x = var_3573_cast_fp16, y = v_15_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [4]> var_3575_perm_0 = const()[name = string("op_3575_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3576 = const()[name = string("op_3576"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_3575_cast_fp16 = transpose(perm = var_3575_perm_0, x = out_15_cast_fp16)[name = string("transpose_419")];
            tensor<fp16, [1, 188, 1024]> input_323_cast_fp16 = reshape(shape = var_3576, x = var_3575_cast_fp16)[name = string("input_323_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_7_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196573184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197621824))))[name = string("blocks_7_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_86_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_7_o_weight_to_fp16_quantized, x = input_323_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_325_cast_fp16 = add(x = input_317_cast_fp16, y = linear_86_cast_fp16)[name = string("input_325_cast_fp16")];
            tensor<int32, [1]> input_327_axes_0 = const()[name = string("input_327_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_7_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197623936)))];
            tensor<fp16, [1024]> enc_layers_7_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197626048)))];
            tensor<fp16, [1, 188, 1024]> input_327_cast_fp16 = layer_norm(axes = input_327_axes_0, beta = enc_layers_7_norm_conv_bias_to_fp16, epsilon = var_3504_to_fp16, gamma = enc_layers_7_norm_conv_weight_to_fp16, x = input_325_cast_fp16)[name = string("input_327_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_7_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197628160))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199725376))))[name = string("blocks_7_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_87_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_7_pw1_weight_to_fp16_quantized, x = input_327_cast_fp16)[name = string("linear_87_cast_fp16")];
            int32 var_3585_split_num_splits_0 = const()[name = string("op_3585_split_num_splits_0"), val = int32(2)];
            int32 var_3585_split_axis_0 = const()[name = string("op_3585_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_3585_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_3585_split_cast_fp16_1 = split(axis = var_3585_split_axis_0, num_splits = var_3585_split_num_splits_0, x = linear_87_cast_fp16)[name = string("op_3585_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3585_split_1_sigmoid_cast_fp16 = sigmoid(x = var_3585_split_cast_fp16_1)[name = string("op_3585_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3585_cast_fp16 = mul(x = var_3585_split_cast_fp16_0, y = var_3585_split_1_sigmoid_cast_fp16)[name = string("op_3585_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_29_cast_fp16 = mul(x = var_3585_cast_fp16, y = valid_cast_fp16)[name = string("h_29_cast_fp16")];
            tensor<int32, [3]> input_331_perm_0 = const()[name = string("input_331_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_333_pad_type_0 = const()[name = string("input_333_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_333_pad_0 = const()[name = string("input_333_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_333_groups_0 = const()[name = string("input_333_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_333_strides_0 = const()[name = string("input_333_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_333_dilations_0 = const()[name = string("input_333_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_52_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199729536))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199738816))))[name = string("const_52_to_fp16_quantized")];
            tensor<fp16, [1024]> const_53_to_fp16 = const()[name = string("const_53_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199740928)))];
            tensor<fp16, [1, 1024, 188]> input_331_cast_fp16 = transpose(perm = input_331_perm_0, x = h_29_cast_fp16)[name = string("transpose_418")];
            tensor<fp16, [1, 1024, 188]> input_335_cast_fp16 = conv(bias = const_53_to_fp16, 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 = const_52_to_fp16_quantized, x = input_331_cast_fp16)[name = string("input_335_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_31_cast_fp16 = silu(x = input_335_cast_fp16)[name = string("h_31_cast_fp16")];
            tensor<int32, [3]> input_337_perm_0 = const()[name = string("input_337_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_7_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199743040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200791680))))[name = string("blocks_7_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_337_cast_fp16 = transpose(perm = input_337_perm_0, x = h_31_cast_fp16)[name = string("transpose_417")];
            tensor<fp16, [1, 188, 1024]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_7_pw2_weight_to_fp16_quantized, x = input_337_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_339_cast_fp16 = add(x = input_325_cast_fp16, y = linear_88_cast_fp16)[name = string("input_339_cast_fp16")];
            tensor<int32, [1]> input_341_axes_0 = const()[name = string("input_341_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200793792)))];
            tensor<fp16, [1024]> enc_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200795904)))];
            tensor<fp16, [1, 188, 1024]> input_341_cast_fp16 = layer_norm(axes = input_341_axes_0, beta = enc_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_3504_to_fp16, gamma = enc_layers_7_norm_feed_forward2_weight_to_fp16, x = input_339_cast_fp16)[name = string("input_341_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_7_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200798016))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204992384))))[name = string("blocks_7_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_89_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_7_ff2_1_weight_to_fp16_quantized, x = input_341_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_345_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_345_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_7_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205000640))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209195008))))[name = string("blocks_7_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_90_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_7_ff2_2_weight_to_fp16_quantized, x = input_345_cast_fp16)[name = string("linear_90_cast_fp16")];
            fp16 var_3606_to_fp16 = const()[name = string("op_3606_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3607_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_3606_to_fp16)[name = string("op_3607_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_347_cast_fp16 = add(x = input_339_cast_fp16, y = var_3607_cast_fp16)[name = string("input_347_cast_fp16")];
            tensor<int32, [1]> input_349_axes_0 = const()[name = string("input_349_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_7_norm_out_weight_to_fp16 = const()[name = string("enc_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209197120)))];
            tensor<fp16, [1024]> enc_layers_7_norm_out_bias_to_fp16 = const()[name = string("enc_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209199232)))];
            tensor<fp16, [1, 188, 1024]> input_349_cast_fp16 = layer_norm(axes = input_349_axes_0, beta = enc_layers_7_norm_out_bias_to_fp16, epsilon = var_3504_to_fp16, gamma = enc_layers_7_norm_out_weight_to_fp16, x = input_347_cast_fp16)[name = string("input_349_cast_fp16")];
            int32 var_3620 = const()[name = string("op_3620"), val = int32(-1)];
            tensor<int32, [1]> input_351_axes_0 = const()[name = string("input_351_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209201344)))];
            tensor<fp16, [1024]> enc_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209203456)))];
            fp16 var_3628_to_fp16 = const()[name = string("op_3628_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_351_cast_fp16 = layer_norm(axes = input_351_axes_0, beta = enc_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_3628_to_fp16, gamma = enc_layers_8_norm_feed_forward1_weight_to_fp16, x = input_349_cast_fp16)[name = string("input_351_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_8_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209205568))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213399936))))[name = string("blocks_8_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_8_ff1_1_weight_to_fp16_quantized, x = input_351_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_355_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_355_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_8_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213408192))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217602560))))[name = string("blocks_8_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_92_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_8_ff1_2_weight_to_fp16_quantized, x = input_355_cast_fp16)[name = string("linear_92_cast_fp16")];
            fp16 var_3648_to_fp16 = const()[name = string("op_3648_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3649_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_3648_to_fp16)[name = string("op_3649_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_357_cast_fp16 = add(x = input_349_cast_fp16, y = var_3649_cast_fp16)[name = string("input_357_cast_fp16")];
            tensor<int32, [1]> input_359_axes_0 = const()[name = string("input_359_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217604672)))];
            tensor<fp16, [1024]> enc_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217606784)))];
            tensor<fp16, [1, 188, 1024]> input_359_cast_fp16 = layer_norm(axes = input_359_axes_0, beta = enc_layers_8_norm_self_att_bias_to_fp16, epsilon = var_3628_to_fp16, gamma = enc_layers_8_norm_self_att_weight_to_fp16, x = input_357_cast_fp16)[name = string("input_359_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_8_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217608896))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218657536))))[name = string("blocks_8_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_93_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_8_q_weight_to_fp16_quantized, x = input_359_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_3655 = const()[name = string("op_3655"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3656_cast_fp16 = reshape(shape = var_3655, x = linear_93_cast_fp16)[name = string("op_3656_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_8_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218659648))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219708288))))[name = string("blocks_8_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_94_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_8_k_weight_to_fp16_quantized, x = input_359_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<int32, [4]> var_3660 = const()[name = string("op_3660"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3661_cast_fp16 = reshape(shape = var_3660, x = linear_94_cast_fp16)[name = string("op_3661_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_8_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219710400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220759040))))[name = string("blocks_8_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_8_v_weight_to_fp16_quantized, x = input_359_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<int32, [4]> var_3665 = const()[name = string("op_3665"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3666_cast_fp16 = reshape(shape = var_3665, x = linear_95_cast_fp16)[name = string("op_3666_cast_fp16")];
            tensor<int32, [4]> v_17_perm_0 = const()[name = string("v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_102_to_fp16 = const()[name = string("const_102_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220761152)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_17_cast_fp16 = add(x = var_3656_cast_fp16, y = const_102_to_fp16)[name = string("q_u_17_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_103_to_fp16 = const()[name = string("const_103_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220763264)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_17_cast_fp16 = add(x = var_3656_cast_fp16, y = const_103_to_fp16)[name = string("q_v_17_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_8_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220765376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221814016))))[name = string("blocks_8_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_8_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<int32, [4]> var_3676 = const()[name = string("op_3676"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_17_cast_fp16 = reshape(shape = var_3676, x = linear_96_cast_fp16)[name = string("p_17_cast_fp16")];
            bool input_361_transpose_x_0 = const()[name = string("input_361_transpose_x_0"), val = bool(false)];
            bool input_361_transpose_y_0 = const()[name = string("input_361_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_224_perm_0 = const()[name = string("transpose_224_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_225_perm_0 = const()[name = string("transpose_225_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_225 = transpose(perm = transpose_225_perm_0, x = p_17_cast_fp16)[name = string("transpose_415")];
            tensor<fp16, [1, 8, 188, 128]> transpose_224 = transpose(perm = transpose_224_perm_0, x = q_v_17_cast_fp16)[name = string("transpose_416")];
            tensor<fp16, [1, 8, 188, 375]> input_361_cast_fp16 = matmul(transpose_x = input_361_transpose_x_0, transpose_y = input_361_transpose_y_0, x = transpose_224, y = transpose_225)[name = string("input_361_cast_fp16")];
            tensor<int32, [8]> var_3682_pad_0 = const()[name = string("op_3682_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_3682_mode_0 = const()[name = string("op_3682_mode_0"), val = string("constant")];
            fp16 const_22_to_fp16 = const()[name = string("const_22_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_3682_cast_fp16 = pad(constant_val = const_22_to_fp16, mode = var_3682_mode_0, pad = var_3682_pad_0, x = input_361_cast_fp16)[name = string("op_3682_cast_fp16")];
            tensor<int32, [4]> var_3683 = const()[name = string("op_3683"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_33_cast_fp16 = reshape(shape = var_3683, x = var_3682_cast_fp16)[name = string("scores_33_cast_fp16")];
            tensor<int32, [4]> var_3687_begin_0 = const()[name = string("op_3687_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3687_end_0 = const()[name = string("op_3687_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_3687_end_mask_0 = const()[name = string("op_3687_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_3687_cast_fp16 = slice_by_index(begin = var_3687_begin_0, end = var_3687_end_0, end_mask = var_3687_end_mask_0, x = scores_33_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [4]> var_3688 = const()[name = string("op_3688"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_3689_cast_fp16 = reshape(shape = var_3688, x = var_3687_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<int32, [4]> bd_17_begin_0 = const()[name = string("bd_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_17_end_0 = const()[name = string("bd_17_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_17_end_mask_0 = const()[name = string("bd_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_17_cast_fp16 = slice_by_index(begin = bd_17_begin_0, end = bd_17_end_0, end_mask = bd_17_end_mask_0, x = var_3689_cast_fp16)[name = string("bd_17_cast_fp16")];
            bool var_3692_transpose_x_0 = const()[name = string("op_3692_transpose_x_0"), val = bool(false)];
            bool var_3692_transpose_y_0 = const()[name = string("op_3692_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_226_perm_0 = const()[name = string("transpose_226_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_227_perm_0 = const()[name = string("transpose_227_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_227 = transpose(perm = transpose_227_perm_0, x = var_3661_cast_fp16)[name = string("transpose_413")];
            tensor<fp16, [1, 8, 188, 128]> transpose_226 = transpose(perm = transpose_226_perm_0, x = q_u_17_cast_fp16)[name = string("transpose_414")];
            tensor<fp16, [1, 8, 188, 188]> var_3692_cast_fp16 = matmul(transpose_x = var_3692_transpose_x_0, transpose_y = var_3692_transpose_y_0, x = transpose_226, y = transpose_227)[name = string("op_3692_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3693_cast_fp16 = add(x = var_3692_cast_fp16, y = bd_17_cast_fp16)[name = string("op_3693_cast_fp16")];
            fp16 var_3694_to_fp16 = const()[name = string("op_3694_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_3695_cast_fp16 = mul(x = var_3693_cast_fp16, y = var_3694_to_fp16)[name = string("op_3695_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_35_cast_fp16 = add(x = var_3695_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_35_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3697_cast_fp16 = softmax(axis = var_3620, x = scores_35_cast_fp16)[name = string("op_3697_cast_fp16")];
            bool out_17_transpose_x_0 = const()[name = string("out_17_transpose_x_0"), val = bool(false)];
            bool out_17_transpose_y_0 = const()[name = string("out_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = var_3666_cast_fp16)[name = string("transpose_412")];
            tensor<fp16, [1, 8, 188, 128]> out_17_cast_fp16 = matmul(transpose_x = out_17_transpose_x_0, transpose_y = out_17_transpose_y_0, x = var_3697_cast_fp16, y = v_17_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [4]> var_3699_perm_0 = const()[name = string("op_3699_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3700 = const()[name = string("op_3700"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_3699_cast_fp16 = transpose(perm = var_3699_perm_0, x = out_17_cast_fp16)[name = string("transpose_411")];
            tensor<fp16, [1, 188, 1024]> input_363_cast_fp16 = reshape(shape = var_3700, x = var_3699_cast_fp16)[name = string("input_363_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_8_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221816128))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222864768))))[name = string("blocks_8_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_97_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_8_o_weight_to_fp16_quantized, x = input_363_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_365_cast_fp16 = add(x = input_357_cast_fp16, y = linear_97_cast_fp16)[name = string("input_365_cast_fp16")];
            tensor<int32, [1]> input_367_axes_0 = const()[name = string("input_367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_8_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222866880)))];
            tensor<fp16, [1024]> enc_layers_8_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222868992)))];
            tensor<fp16, [1, 188, 1024]> input_367_cast_fp16 = layer_norm(axes = input_367_axes_0, beta = enc_layers_8_norm_conv_bias_to_fp16, epsilon = var_3628_to_fp16, gamma = enc_layers_8_norm_conv_weight_to_fp16, x = input_365_cast_fp16)[name = string("input_367_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_8_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222871104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224968320))))[name = string("blocks_8_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_98_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_8_pw1_weight_to_fp16_quantized, x = input_367_cast_fp16)[name = string("linear_98_cast_fp16")];
            int32 var_3709_split_num_splits_0 = const()[name = string("op_3709_split_num_splits_0"), val = int32(2)];
            int32 var_3709_split_axis_0 = const()[name = string("op_3709_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_3709_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_3709_split_cast_fp16_1 = split(axis = var_3709_split_axis_0, num_splits = var_3709_split_num_splits_0, x = linear_98_cast_fp16)[name = string("op_3709_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3709_split_1_sigmoid_cast_fp16 = sigmoid(x = var_3709_split_cast_fp16_1)[name = string("op_3709_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3709_cast_fp16 = mul(x = var_3709_split_cast_fp16_0, y = var_3709_split_1_sigmoid_cast_fp16)[name = string("op_3709_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_33_cast_fp16 = mul(x = var_3709_cast_fp16, y = valid_cast_fp16)[name = string("h_33_cast_fp16")];
            tensor<int32, [3]> input_371_perm_0 = const()[name = string("input_371_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_373_pad_type_0 = const()[name = string("input_373_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_373_pad_0 = const()[name = string("input_373_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_373_groups_0 = const()[name = string("input_373_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_373_strides_0 = const()[name = string("input_373_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_373_dilations_0 = const()[name = string("input_373_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_54_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224972480))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224981760))))[name = string("const_54_to_fp16_quantized")];
            tensor<fp16, [1024]> const_55_to_fp16 = const()[name = string("const_55_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224983872)))];
            tensor<fp16, [1, 1024, 188]> input_371_cast_fp16 = transpose(perm = input_371_perm_0, x = h_33_cast_fp16)[name = string("transpose_410")];
            tensor<fp16, [1, 1024, 188]> input_375_cast_fp16 = conv(bias = const_55_to_fp16, dilations = input_373_dilations_0, groups = input_373_groups_0, pad = input_373_pad_0, pad_type = input_373_pad_type_0, strides = input_373_strides_0, weight = const_54_to_fp16_quantized, x = input_371_cast_fp16)[name = string("input_375_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_35_cast_fp16 = silu(x = input_375_cast_fp16)[name = string("h_35_cast_fp16")];
            tensor<int32, [3]> input_377_perm_0 = const()[name = string("input_377_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_8_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224985984))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226034624))))[name = string("blocks_8_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_377_cast_fp16 = transpose(perm = input_377_perm_0, x = h_35_cast_fp16)[name = string("transpose_409")];
            tensor<fp16, [1, 188, 1024]> linear_99_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_8_pw2_weight_to_fp16_quantized, x = input_377_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_379_cast_fp16 = add(x = input_365_cast_fp16, y = linear_99_cast_fp16)[name = string("input_379_cast_fp16")];
            tensor<int32, [1]> input_381_axes_0 = const()[name = string("input_381_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226036736)))];
            tensor<fp16, [1024]> enc_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226038848)))];
            tensor<fp16, [1, 188, 1024]> input_381_cast_fp16 = layer_norm(axes = input_381_axes_0, beta = enc_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_3628_to_fp16, gamma = enc_layers_8_norm_feed_forward2_weight_to_fp16, x = input_379_cast_fp16)[name = string("input_381_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_8_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226040960))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230235328))))[name = string("blocks_8_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_100_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_8_ff2_1_weight_to_fp16_quantized, x = input_381_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_385_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_385_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_8_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230243584))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234437952))))[name = string("blocks_8_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_101_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_8_ff2_2_weight_to_fp16_quantized, x = input_385_cast_fp16)[name = string("linear_101_cast_fp16")];
            fp16 var_3730_to_fp16 = const()[name = string("op_3730_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3731_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_3730_to_fp16)[name = string("op_3731_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_387_cast_fp16 = add(x = input_379_cast_fp16, y = var_3731_cast_fp16)[name = string("input_387_cast_fp16")];
            tensor<int32, [1]> input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_8_norm_out_weight_to_fp16 = const()[name = string("enc_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234440064)))];
            tensor<fp16, [1024]> enc_layers_8_norm_out_bias_to_fp16 = const()[name = string("enc_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234442176)))];
            tensor<fp16, [1, 188, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = enc_layers_8_norm_out_bias_to_fp16, epsilon = var_3628_to_fp16, gamma = enc_layers_8_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")];
            int32 var_3744 = const()[name = string("op_3744"), val = int32(-1)];
            tensor<int32, [1]> input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234444288)))];
            tensor<fp16, [1024]> enc_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234446400)))];
            fp16 var_3752_to_fp16 = const()[name = string("op_3752_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = enc_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_3752_to_fp16, gamma = enc_layers_9_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_9_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234448512))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238642880))))[name = string("blocks_9_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_102_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_9_ff1_1_weight_to_fp16_quantized, x = input_391_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_395_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("input_395_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_9_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238651136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242845504))))[name = string("blocks_9_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_9_ff1_2_weight_to_fp16_quantized, x = input_395_cast_fp16)[name = string("linear_103_cast_fp16")];
            fp16 var_3772_to_fp16 = const()[name = string("op_3772_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3773_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_3772_to_fp16)[name = string("op_3773_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_397_cast_fp16 = add(x = input_389_cast_fp16, y = var_3773_cast_fp16)[name = string("input_397_cast_fp16")];
            tensor<int32, [1]> input_399_axes_0 = const()[name = string("input_399_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242847616)))];
            tensor<fp16, [1024]> enc_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242849728)))];
            tensor<fp16, [1, 188, 1024]> input_399_cast_fp16 = layer_norm(axes = input_399_axes_0, beta = enc_layers_9_norm_self_att_bias_to_fp16, epsilon = var_3752_to_fp16, gamma = enc_layers_9_norm_self_att_weight_to_fp16, x = input_397_cast_fp16)[name = string("input_399_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_9_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242851840))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243900480))))[name = string("blocks_9_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_104_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_9_q_weight_to_fp16_quantized, x = input_399_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<int32, [4]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3780_cast_fp16 = reshape(shape = var_3779, x = linear_104_cast_fp16)[name = string("op_3780_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_9_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243902592))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244951232))))[name = string("blocks_9_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_105_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_9_k_weight_to_fp16_quantized, x = input_399_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_3784 = const()[name = string("op_3784"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3785_cast_fp16 = reshape(shape = var_3784, x = linear_105_cast_fp16)[name = string("op_3785_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_9_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244953344))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246001984))))[name = string("blocks_9_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_106_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_9_v_weight_to_fp16_quantized, x = input_399_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_3789 = const()[name = string("op_3789"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3790_cast_fp16 = reshape(shape = var_3789, x = linear_106_cast_fp16)[name = string("op_3790_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_104_to_fp16 = const()[name = string("const_104_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246004096)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_19_cast_fp16 = add(x = var_3780_cast_fp16, y = const_104_to_fp16)[name = string("q_u_19_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246006208)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_19_cast_fp16 = add(x = var_3780_cast_fp16, y = const_105_to_fp16)[name = string("q_v_19_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_9_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246008320))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247056960))))[name = string("blocks_9_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_107_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_9_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_3800 = const()[name = string("op_3800"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_19_cast_fp16 = reshape(shape = var_3800, x = linear_107_cast_fp16)[name = string("p_19_cast_fp16")];
            bool input_401_transpose_x_0 = const()[name = string("input_401_transpose_x_0"), val = bool(false)];
            bool input_401_transpose_y_0 = const()[name = string("input_401_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_228_perm_0 = const()[name = string("transpose_228_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_229_perm_0 = const()[name = string("transpose_229_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_229 = transpose(perm = transpose_229_perm_0, x = p_19_cast_fp16)[name = string("transpose_407")];
            tensor<fp16, [1, 8, 188, 128]> transpose_228 = transpose(perm = transpose_228_perm_0, x = q_v_19_cast_fp16)[name = string("transpose_408")];
            tensor<fp16, [1, 8, 188, 375]> input_401_cast_fp16 = matmul(transpose_x = input_401_transpose_x_0, transpose_y = input_401_transpose_y_0, x = transpose_228, y = transpose_229)[name = string("input_401_cast_fp16")];
            tensor<int32, [8]> var_3806_pad_0 = const()[name = string("op_3806_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_3806_mode_0 = const()[name = string("op_3806_mode_0"), val = string("constant")];
            fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_3806_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = var_3806_mode_0, pad = var_3806_pad_0, x = input_401_cast_fp16)[name = string("op_3806_cast_fp16")];
            tensor<int32, [4]> var_3807 = const()[name = string("op_3807"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_37_cast_fp16 = reshape(shape = var_3807, x = var_3806_cast_fp16)[name = string("scores_37_cast_fp16")];
            tensor<int32, [4]> var_3811_begin_0 = const()[name = string("op_3811_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3811_end_0 = const()[name = string("op_3811_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_3811_end_mask_0 = const()[name = string("op_3811_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_3811_cast_fp16 = slice_by_index(begin = var_3811_begin_0, end = var_3811_end_0, end_mask = var_3811_end_mask_0, x = scores_37_cast_fp16)[name = string("op_3811_cast_fp16")];
            tensor<int32, [4]> var_3812 = const()[name = string("op_3812"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_3813_cast_fp16 = reshape(shape = var_3812, x = var_3811_cast_fp16)[name = string("op_3813_cast_fp16")];
            tensor<int32, [4]> bd_19_begin_0 = const()[name = string("bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_19_end_0 = const()[name = string("bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_19_end_mask_0 = const()[name = string("bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_19_cast_fp16 = slice_by_index(begin = bd_19_begin_0, end = bd_19_end_0, end_mask = bd_19_end_mask_0, x = var_3813_cast_fp16)[name = string("bd_19_cast_fp16")];
            bool var_3816_transpose_x_0 = const()[name = string("op_3816_transpose_x_0"), val = bool(false)];
            bool var_3816_transpose_y_0 = const()[name = string("op_3816_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_230_perm_0 = const()[name = string("transpose_230_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_231_perm_0 = const()[name = string("transpose_231_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_231 = transpose(perm = transpose_231_perm_0, x = var_3785_cast_fp16)[name = string("transpose_405")];
            tensor<fp16, [1, 8, 188, 128]> transpose_230 = transpose(perm = transpose_230_perm_0, x = q_u_19_cast_fp16)[name = string("transpose_406")];
            tensor<fp16, [1, 8, 188, 188]> var_3816_cast_fp16 = matmul(transpose_x = var_3816_transpose_x_0, transpose_y = var_3816_transpose_y_0, x = transpose_230, y = transpose_231)[name = string("op_3816_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3817_cast_fp16 = add(x = var_3816_cast_fp16, y = bd_19_cast_fp16)[name = string("op_3817_cast_fp16")];
            fp16 var_3818_to_fp16 = const()[name = string("op_3818_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_3819_cast_fp16 = mul(x = var_3817_cast_fp16, y = var_3818_to_fp16)[name = string("op_3819_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_39_cast_fp16 = add(x = var_3819_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_39_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3821_cast_fp16 = softmax(axis = var_3744, x = scores_39_cast_fp16)[name = string("op_3821_cast_fp16")];
            bool out_19_transpose_x_0 = const()[name = string("out_19_transpose_x_0"), val = bool(false)];
            bool out_19_transpose_y_0 = const()[name = string("out_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_3790_cast_fp16)[name = string("transpose_404")];
            tensor<fp16, [1, 8, 188, 128]> out_19_cast_fp16 = matmul(transpose_x = out_19_transpose_x_0, transpose_y = out_19_transpose_y_0, x = var_3821_cast_fp16, y = v_19_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [4]> var_3823_perm_0 = const()[name = string("op_3823_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3824 = const()[name = string("op_3824"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_3823_cast_fp16 = transpose(perm = var_3823_perm_0, x = out_19_cast_fp16)[name = string("transpose_403")];
            tensor<fp16, [1, 188, 1024]> input_403_cast_fp16 = reshape(shape = var_3824, x = var_3823_cast_fp16)[name = string("input_403_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_9_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247059072))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248107712))))[name = string("blocks_9_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_108_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_9_o_weight_to_fp16_quantized, x = input_403_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_405_cast_fp16 = add(x = input_397_cast_fp16, y = linear_108_cast_fp16)[name = string("input_405_cast_fp16")];
            tensor<int32, [1]> input_407_axes_0 = const()[name = string("input_407_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_9_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248109824)))];
            tensor<fp16, [1024]> enc_layers_9_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248111936)))];
            tensor<fp16, [1, 188, 1024]> input_407_cast_fp16 = layer_norm(axes = input_407_axes_0, beta = enc_layers_9_norm_conv_bias_to_fp16, epsilon = var_3752_to_fp16, gamma = enc_layers_9_norm_conv_weight_to_fp16, x = input_405_cast_fp16)[name = string("input_407_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_9_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248114048))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250211264))))[name = string("blocks_9_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_109_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_9_pw1_weight_to_fp16_quantized, x = input_407_cast_fp16)[name = string("linear_109_cast_fp16")];
            int32 var_3833_split_num_splits_0 = const()[name = string("op_3833_split_num_splits_0"), val = int32(2)];
            int32 var_3833_split_axis_0 = const()[name = string("op_3833_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_3833_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_3833_split_cast_fp16_1 = split(axis = var_3833_split_axis_0, num_splits = var_3833_split_num_splits_0, x = linear_109_cast_fp16)[name = string("op_3833_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3833_split_1_sigmoid_cast_fp16 = sigmoid(x = var_3833_split_cast_fp16_1)[name = string("op_3833_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3833_cast_fp16 = mul(x = var_3833_split_cast_fp16_0, y = var_3833_split_1_sigmoid_cast_fp16)[name = string("op_3833_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_37_cast_fp16 = mul(x = var_3833_cast_fp16, y = valid_cast_fp16)[name = string("h_37_cast_fp16")];
            tensor<int32, [3]> input_411_perm_0 = const()[name = string("input_411_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_413_pad_type_0 = const()[name = string("input_413_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_413_pad_0 = const()[name = string("input_413_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_413_groups_0 = const()[name = string("input_413_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_413_strides_0 = const()[name = string("input_413_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_413_dilations_0 = const()[name = string("input_413_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_56_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250215424))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250224704))))[name = string("const_56_to_fp16_quantized")];
            tensor<fp16, [1024]> const_57_to_fp16 = const()[name = string("const_57_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250226816)))];
            tensor<fp16, [1, 1024, 188]> input_411_cast_fp16 = transpose(perm = input_411_perm_0, x = h_37_cast_fp16)[name = string("transpose_402")];
            tensor<fp16, [1, 1024, 188]> input_415_cast_fp16 = conv(bias = const_57_to_fp16, dilations = input_413_dilations_0, groups = input_413_groups_0, pad = input_413_pad_0, pad_type = input_413_pad_type_0, strides = input_413_strides_0, weight = const_56_to_fp16_quantized, x = input_411_cast_fp16)[name = string("input_415_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_39_cast_fp16 = silu(x = input_415_cast_fp16)[name = string("h_39_cast_fp16")];
            tensor<int32, [3]> input_417_perm_0 = const()[name = string("input_417_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_9_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250228928))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251277568))))[name = string("blocks_9_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_417_cast_fp16 = transpose(perm = input_417_perm_0, x = h_39_cast_fp16)[name = string("transpose_401")];
            tensor<fp16, [1, 188, 1024]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_9_pw2_weight_to_fp16_quantized, x = input_417_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_419_cast_fp16 = add(x = input_405_cast_fp16, y = linear_110_cast_fp16)[name = string("input_419_cast_fp16")];
            tensor<int32, [1]> input_421_axes_0 = const()[name = string("input_421_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251279680)))];
            tensor<fp16, [1024]> enc_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251281792)))];
            tensor<fp16, [1, 188, 1024]> input_421_cast_fp16 = layer_norm(axes = input_421_axes_0, beta = enc_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_3752_to_fp16, gamma = enc_layers_9_norm_feed_forward2_weight_to_fp16, x = input_419_cast_fp16)[name = string("input_421_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_9_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251283904))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255478272))))[name = string("blocks_9_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_9_ff2_1_weight_to_fp16_quantized, x = input_421_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_425_cast_fp16 = silu(x = linear_111_cast_fp16)[name = string("input_425_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_9_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255486528))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259680896))))[name = string("blocks_9_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_112_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_9_ff2_2_weight_to_fp16_quantized, x = input_425_cast_fp16)[name = string("linear_112_cast_fp16")];
            fp16 var_3854_to_fp16 = const()[name = string("op_3854_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3855_cast_fp16 = mul(x = linear_112_cast_fp16, y = var_3854_to_fp16)[name = string("op_3855_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_427_cast_fp16 = add(x = input_419_cast_fp16, y = var_3855_cast_fp16)[name = string("input_427_cast_fp16")];
            tensor<int32, [1]> input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_9_norm_out_weight_to_fp16 = const()[name = string("enc_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259683008)))];
            tensor<fp16, [1024]> enc_layers_9_norm_out_bias_to_fp16 = const()[name = string("enc_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259685120)))];
            tensor<fp16, [1, 188, 1024]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = enc_layers_9_norm_out_bias_to_fp16, epsilon = var_3752_to_fp16, gamma = enc_layers_9_norm_out_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")];
            int32 var_3868 = const()[name = string("op_3868"), val = int32(-1)];
            tensor<int32, [1]> input_431_axes_0 = const()[name = string("input_431_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259687232)))];
            tensor<fp16, [1024]> enc_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259689344)))];
            fp16 var_3876_to_fp16 = const()[name = string("op_3876_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_431_cast_fp16 = layer_norm(axes = input_431_axes_0, beta = enc_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_3876_to_fp16, gamma = enc_layers_10_norm_feed_forward1_weight_to_fp16, x = input_429_cast_fp16)[name = string("input_431_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_10_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259691456))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263885824))))[name = string("blocks_10_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_113_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_10_ff1_1_weight_to_fp16_quantized, x = input_431_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_435_cast_fp16 = silu(x = linear_113_cast_fp16)[name = string("input_435_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_10_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263894080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268088448))))[name = string("blocks_10_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_114_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_10_ff1_2_weight_to_fp16_quantized, x = input_435_cast_fp16)[name = string("linear_114_cast_fp16")];
            fp16 var_3896_to_fp16 = const()[name = string("op_3896_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3897_cast_fp16 = mul(x = linear_114_cast_fp16, y = var_3896_to_fp16)[name = string("op_3897_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_437_cast_fp16 = add(x = input_429_cast_fp16, y = var_3897_cast_fp16)[name = string("input_437_cast_fp16")];
            tensor<int32, [1]> input_439_axes_0 = const()[name = string("input_439_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268090560)))];
            tensor<fp16, [1024]> enc_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268092672)))];
            tensor<fp16, [1, 188, 1024]> input_439_cast_fp16 = layer_norm(axes = input_439_axes_0, beta = enc_layers_10_norm_self_att_bias_to_fp16, epsilon = var_3876_to_fp16, gamma = enc_layers_10_norm_self_att_weight_to_fp16, x = input_437_cast_fp16)[name = string("input_439_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_10_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268094784))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269143424))))[name = string("blocks_10_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_115_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_10_q_weight_to_fp16_quantized, x = input_439_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<int32, [4]> var_3903 = const()[name = string("op_3903"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3904_cast_fp16 = reshape(shape = var_3903, x = linear_115_cast_fp16)[name = string("op_3904_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_10_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269145536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270194176))))[name = string("blocks_10_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_10_k_weight_to_fp16_quantized, x = input_439_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<int32, [4]> var_3908 = const()[name = string("op_3908"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3909_cast_fp16 = reshape(shape = var_3908, x = linear_116_cast_fp16)[name = string("op_3909_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_10_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270196288))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271244928))))[name = string("blocks_10_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_10_v_weight_to_fp16_quantized, x = input_439_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<int32, [4]> var_3913 = const()[name = string("op_3913"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_3914_cast_fp16 = reshape(shape = var_3913, x = linear_117_cast_fp16)[name = string("op_3914_cast_fp16")];
            tensor<int32, [4]> v_21_perm_0 = const()[name = string("v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_106_to_fp16 = const()[name = string("const_106_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271247040)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_21_cast_fp16 = add(x = var_3904_cast_fp16, y = const_106_to_fp16)[name = string("q_u_21_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_107_to_fp16 = const()[name = string("const_107_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271249152)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_21_cast_fp16 = add(x = var_3904_cast_fp16, y = const_107_to_fp16)[name = string("q_v_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_10_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271251264))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272299904))))[name = string("blocks_10_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_118_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_10_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<int32, [4]> var_3924 = const()[name = string("op_3924"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_21_cast_fp16 = reshape(shape = var_3924, x = linear_118_cast_fp16)[name = string("p_21_cast_fp16")];
            bool input_441_transpose_x_0 = const()[name = string("input_441_transpose_x_0"), val = bool(false)];
            bool input_441_transpose_y_0 = const()[name = string("input_441_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_232_perm_0 = const()[name = string("transpose_232_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_233_perm_0 = const()[name = string("transpose_233_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_233 = transpose(perm = transpose_233_perm_0, x = p_21_cast_fp16)[name = string("transpose_399")];
            tensor<fp16, [1, 8, 188, 128]> transpose_232 = transpose(perm = transpose_232_perm_0, x = q_v_21_cast_fp16)[name = string("transpose_400")];
            tensor<fp16, [1, 8, 188, 375]> input_441_cast_fp16 = matmul(transpose_x = input_441_transpose_x_0, transpose_y = input_441_transpose_y_0, x = transpose_232, y = transpose_233)[name = string("input_441_cast_fp16")];
            tensor<int32, [8]> var_3930_pad_0 = const()[name = string("op_3930_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_3930_mode_0 = const()[name = string("op_3930_mode_0"), val = string("constant")];
            fp16 const_24_to_fp16 = const()[name = string("const_24_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_3930_cast_fp16 = pad(constant_val = const_24_to_fp16, mode = var_3930_mode_0, pad = var_3930_pad_0, x = input_441_cast_fp16)[name = string("op_3930_cast_fp16")];
            tensor<int32, [4]> var_3931 = const()[name = string("op_3931"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_41_cast_fp16 = reshape(shape = var_3931, x = var_3930_cast_fp16)[name = string("scores_41_cast_fp16")];
            tensor<int32, [4]> var_3935_begin_0 = const()[name = string("op_3935_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_3935_end_0 = const()[name = string("op_3935_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_3935_end_mask_0 = const()[name = string("op_3935_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_3935_cast_fp16 = slice_by_index(begin = var_3935_begin_0, end = var_3935_end_0, end_mask = var_3935_end_mask_0, x = scores_41_cast_fp16)[name = string("op_3935_cast_fp16")];
            tensor<int32, [4]> var_3936 = const()[name = string("op_3936"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_3937_cast_fp16 = reshape(shape = var_3936, x = var_3935_cast_fp16)[name = string("op_3937_cast_fp16")];
            tensor<int32, [4]> bd_21_begin_0 = const()[name = string("bd_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_21_end_0 = const()[name = string("bd_21_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_21_end_mask_0 = const()[name = string("bd_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_21_cast_fp16 = slice_by_index(begin = bd_21_begin_0, end = bd_21_end_0, end_mask = bd_21_end_mask_0, x = var_3937_cast_fp16)[name = string("bd_21_cast_fp16")];
            bool var_3940_transpose_x_0 = const()[name = string("op_3940_transpose_x_0"), val = bool(false)];
            bool var_3940_transpose_y_0 = const()[name = string("op_3940_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_234_perm_0 = const()[name = string("transpose_234_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_235_perm_0 = const()[name = string("transpose_235_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_235 = transpose(perm = transpose_235_perm_0, x = var_3909_cast_fp16)[name = string("transpose_397")];
            tensor<fp16, [1, 8, 188, 128]> transpose_234 = transpose(perm = transpose_234_perm_0, x = q_u_21_cast_fp16)[name = string("transpose_398")];
            tensor<fp16, [1, 8, 188, 188]> var_3940_cast_fp16 = matmul(transpose_x = var_3940_transpose_x_0, transpose_y = var_3940_transpose_y_0, x = transpose_234, y = transpose_235)[name = string("op_3940_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3941_cast_fp16 = add(x = var_3940_cast_fp16, y = bd_21_cast_fp16)[name = string("op_3941_cast_fp16")];
            fp16 var_3942_to_fp16 = const()[name = string("op_3942_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_3943_cast_fp16 = mul(x = var_3941_cast_fp16, y = var_3942_to_fp16)[name = string("op_3943_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_43_cast_fp16 = add(x = var_3943_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_43_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_3945_cast_fp16 = softmax(axis = var_3868, x = scores_43_cast_fp16)[name = string("op_3945_cast_fp16")];
            bool out_21_transpose_x_0 = const()[name = string("out_21_transpose_x_0"), val = bool(false)];
            bool out_21_transpose_y_0 = const()[name = string("out_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_3914_cast_fp16)[name = string("transpose_396")];
            tensor<fp16, [1, 8, 188, 128]> out_21_cast_fp16 = matmul(transpose_x = out_21_transpose_x_0, transpose_y = out_21_transpose_y_0, x = var_3945_cast_fp16, y = v_21_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [4]> var_3947_perm_0 = const()[name = string("op_3947_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3948 = const()[name = string("op_3948"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_3947_cast_fp16 = transpose(perm = var_3947_perm_0, x = out_21_cast_fp16)[name = string("transpose_395")];
            tensor<fp16, [1, 188, 1024]> input_443_cast_fp16 = reshape(shape = var_3948, x = var_3947_cast_fp16)[name = string("input_443_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_10_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272302016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273350656))))[name = string("blocks_10_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_119_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_10_o_weight_to_fp16_quantized, x = input_443_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_445_cast_fp16 = add(x = input_437_cast_fp16, y = linear_119_cast_fp16)[name = string("input_445_cast_fp16")];
            tensor<int32, [1]> input_447_axes_0 = const()[name = string("input_447_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_10_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273352768)))];
            tensor<fp16, [1024]> enc_layers_10_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273354880)))];
            tensor<fp16, [1, 188, 1024]> input_447_cast_fp16 = layer_norm(axes = input_447_axes_0, beta = enc_layers_10_norm_conv_bias_to_fp16, epsilon = var_3876_to_fp16, gamma = enc_layers_10_norm_conv_weight_to_fp16, x = input_445_cast_fp16)[name = string("input_447_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_10_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273356992))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275454208))))[name = string("blocks_10_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_120_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_10_pw1_weight_to_fp16_quantized, x = input_447_cast_fp16)[name = string("linear_120_cast_fp16")];
            int32 var_3957_split_num_splits_0 = const()[name = string("op_3957_split_num_splits_0"), val = int32(2)];
            int32 var_3957_split_axis_0 = const()[name = string("op_3957_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_3957_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_3957_split_cast_fp16_1 = split(axis = var_3957_split_axis_0, num_splits = var_3957_split_num_splits_0, x = linear_120_cast_fp16)[name = string("op_3957_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3957_split_1_sigmoid_cast_fp16 = sigmoid(x = var_3957_split_cast_fp16_1)[name = string("op_3957_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_3957_cast_fp16 = mul(x = var_3957_split_cast_fp16_0, y = var_3957_split_1_sigmoid_cast_fp16)[name = string("op_3957_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_41_cast_fp16 = mul(x = var_3957_cast_fp16, y = valid_cast_fp16)[name = string("h_41_cast_fp16")];
            tensor<int32, [3]> input_451_perm_0 = const()[name = string("input_451_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_453_pad_type_0 = const()[name = string("input_453_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_453_pad_0 = const()[name = string("input_453_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_453_groups_0 = const()[name = string("input_453_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_453_strides_0 = const()[name = string("input_453_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_453_dilations_0 = const()[name = string("input_453_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_58_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275458368))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275467648))))[name = string("const_58_to_fp16_quantized")];
            tensor<fp16, [1024]> const_59_to_fp16 = const()[name = string("const_59_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275469760)))];
            tensor<fp16, [1, 1024, 188]> input_451_cast_fp16 = transpose(perm = input_451_perm_0, x = h_41_cast_fp16)[name = string("transpose_394")];
            tensor<fp16, [1, 1024, 188]> input_455_cast_fp16 = conv(bias = const_59_to_fp16, dilations = input_453_dilations_0, groups = input_453_groups_0, pad = input_453_pad_0, pad_type = input_453_pad_type_0, strides = input_453_strides_0, weight = const_58_to_fp16_quantized, x = input_451_cast_fp16)[name = string("input_455_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_43_cast_fp16 = silu(x = input_455_cast_fp16)[name = string("h_43_cast_fp16")];
            tensor<int32, [3]> input_457_perm_0 = const()[name = string("input_457_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_10_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275471872))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276520512))))[name = string("blocks_10_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_457_cast_fp16 = transpose(perm = input_457_perm_0, x = h_43_cast_fp16)[name = string("transpose_393")];
            tensor<fp16, [1, 188, 1024]> linear_121_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_10_pw2_weight_to_fp16_quantized, x = input_457_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_459_cast_fp16 = add(x = input_445_cast_fp16, y = linear_121_cast_fp16)[name = string("input_459_cast_fp16")];
            tensor<int32, [1]> input_461_axes_0 = const()[name = string("input_461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276522624)))];
            tensor<fp16, [1024]> enc_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276524736)))];
            tensor<fp16, [1, 188, 1024]> input_461_cast_fp16 = layer_norm(axes = input_461_axes_0, beta = enc_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_3876_to_fp16, gamma = enc_layers_10_norm_feed_forward2_weight_to_fp16, x = input_459_cast_fp16)[name = string("input_461_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_10_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276526848))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280721216))))[name = string("blocks_10_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_10_ff2_1_weight_to_fp16_quantized, x = input_461_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_465_cast_fp16 = silu(x = linear_122_cast_fp16)[name = string("input_465_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_10_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(280729472))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284923840))))[name = string("blocks_10_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_10_ff2_2_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("linear_123_cast_fp16")];
            fp16 var_3978_to_fp16 = const()[name = string("op_3978_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_3979_cast_fp16 = mul(x = linear_123_cast_fp16, y = var_3978_to_fp16)[name = string("op_3979_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_467_cast_fp16 = add(x = input_459_cast_fp16, y = var_3979_cast_fp16)[name = string("input_467_cast_fp16")];
            tensor<int32, [1]> input_469_axes_0 = const()[name = string("input_469_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_10_norm_out_weight_to_fp16 = const()[name = string("enc_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284925952)))];
            tensor<fp16, [1024]> enc_layers_10_norm_out_bias_to_fp16 = const()[name = string("enc_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284928064)))];
            tensor<fp16, [1, 188, 1024]> input_469_cast_fp16 = layer_norm(axes = input_469_axes_0, beta = enc_layers_10_norm_out_bias_to_fp16, epsilon = var_3876_to_fp16, gamma = enc_layers_10_norm_out_weight_to_fp16, x = input_467_cast_fp16)[name = string("input_469_cast_fp16")];
            int32 var_3992 = const()[name = string("op_3992"), val = int32(-1)];
            tensor<int32, [1]> input_471_axes_0 = const()[name = string("input_471_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284930176)))];
            tensor<fp16, [1024]> enc_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284932288)))];
            fp16 var_4000_to_fp16 = const()[name = string("op_4000_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_471_cast_fp16 = layer_norm(axes = input_471_axes_0, beta = enc_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_4000_to_fp16, gamma = enc_layers_11_norm_feed_forward1_weight_to_fp16, x = input_469_cast_fp16)[name = string("input_471_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_11_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284934400))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289128768))))[name = string("blocks_11_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_124_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_11_ff1_1_weight_to_fp16_quantized, x = input_471_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_475_cast_fp16 = silu(x = linear_124_cast_fp16)[name = string("input_475_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_11_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289137024))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293331392))))[name = string("blocks_11_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_125_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_11_ff1_2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("linear_125_cast_fp16")];
            fp16 var_4020_to_fp16 = const()[name = string("op_4020_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4021_cast_fp16 = mul(x = linear_125_cast_fp16, y = var_4020_to_fp16)[name = string("op_4021_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_477_cast_fp16 = add(x = input_469_cast_fp16, y = var_4021_cast_fp16)[name = string("input_477_cast_fp16")];
            tensor<int32, [1]> input_479_axes_0 = const()[name = string("input_479_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293333504)))];
            tensor<fp16, [1024]> enc_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293335616)))];
            tensor<fp16, [1, 188, 1024]> input_479_cast_fp16 = layer_norm(axes = input_479_axes_0, beta = enc_layers_11_norm_self_att_bias_to_fp16, epsilon = var_4000_to_fp16, gamma = enc_layers_11_norm_self_att_weight_to_fp16, x = input_477_cast_fp16)[name = string("input_479_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_11_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293337728))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294386368))))[name = string("blocks_11_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_126_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_11_q_weight_to_fp16_quantized, x = input_479_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_4027 = const()[name = string("op_4027"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4028_cast_fp16 = reshape(shape = var_4027, x = linear_126_cast_fp16)[name = string("op_4028_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_11_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294388480))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295437120))))[name = string("blocks_11_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_127_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_11_k_weight_to_fp16_quantized, x = input_479_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_4032 = const()[name = string("op_4032"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4033_cast_fp16 = reshape(shape = var_4032, x = linear_127_cast_fp16)[name = string("op_4033_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_11_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295439232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296487872))))[name = string("blocks_11_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_128_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_11_v_weight_to_fp16_quantized, x = input_479_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_4037 = const()[name = string("op_4037"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4038_cast_fp16 = reshape(shape = var_4037, x = linear_128_cast_fp16)[name = string("op_4038_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_108_to_fp16 = const()[name = string("const_108_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296489984)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_23_cast_fp16 = add(x = var_4028_cast_fp16, y = const_108_to_fp16)[name = string("q_u_23_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_109_to_fp16 = const()[name = string("const_109_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296492096)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_23_cast_fp16 = add(x = var_4028_cast_fp16, y = const_109_to_fp16)[name = string("q_v_23_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_11_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296494208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297542848))))[name = string("blocks_11_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_129_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_11_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<int32, [4]> var_4048 = const()[name = string("op_4048"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_23_cast_fp16 = reshape(shape = var_4048, x = linear_129_cast_fp16)[name = string("p_23_cast_fp16")];
            bool input_481_transpose_x_0 = const()[name = string("input_481_transpose_x_0"), val = bool(false)];
            bool input_481_transpose_y_0 = const()[name = string("input_481_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_236_perm_0 = const()[name = string("transpose_236_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_237_perm_0 = const()[name = string("transpose_237_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_237 = transpose(perm = transpose_237_perm_0, x = p_23_cast_fp16)[name = string("transpose_391")];
            tensor<fp16, [1, 8, 188, 128]> transpose_236 = transpose(perm = transpose_236_perm_0, x = q_v_23_cast_fp16)[name = string("transpose_392")];
            tensor<fp16, [1, 8, 188, 375]> input_481_cast_fp16 = matmul(transpose_x = input_481_transpose_x_0, transpose_y = input_481_transpose_y_0, x = transpose_236, y = transpose_237)[name = string("input_481_cast_fp16")];
            tensor<int32, [8]> var_4054_pad_0 = const()[name = string("op_4054_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_4054_mode_0 = const()[name = string("op_4054_mode_0"), val = string("constant")];
            fp16 const_25_to_fp16 = const()[name = string("const_25_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_4054_cast_fp16 = pad(constant_val = const_25_to_fp16, mode = var_4054_mode_0, pad = var_4054_pad_0, x = input_481_cast_fp16)[name = string("op_4054_cast_fp16")];
            tensor<int32, [4]> var_4055 = const()[name = string("op_4055"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_45_cast_fp16 = reshape(shape = var_4055, x = var_4054_cast_fp16)[name = string("scores_45_cast_fp16")];
            tensor<int32, [4]> var_4059_begin_0 = const()[name = string("op_4059_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4059_end_0 = const()[name = string("op_4059_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_4059_end_mask_0 = const()[name = string("op_4059_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_4059_cast_fp16 = slice_by_index(begin = var_4059_begin_0, end = var_4059_end_0, end_mask = var_4059_end_mask_0, x = scores_45_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<int32, [4]> var_4060 = const()[name = string("op_4060"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_4061_cast_fp16 = reshape(shape = var_4060, x = var_4059_cast_fp16)[name = string("op_4061_cast_fp16")];
            tensor<int32, [4]> bd_23_begin_0 = const()[name = string("bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_23_end_0 = const()[name = string("bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_23_end_mask_0 = const()[name = string("bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_23_cast_fp16 = slice_by_index(begin = bd_23_begin_0, end = bd_23_end_0, end_mask = bd_23_end_mask_0, x = var_4061_cast_fp16)[name = string("bd_23_cast_fp16")];
            bool var_4064_transpose_x_0 = const()[name = string("op_4064_transpose_x_0"), val = bool(false)];
            bool var_4064_transpose_y_0 = const()[name = string("op_4064_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_238_perm_0 = const()[name = string("transpose_238_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_239_perm_0 = const()[name = string("transpose_239_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_239 = transpose(perm = transpose_239_perm_0, x = var_4033_cast_fp16)[name = string("transpose_389")];
            tensor<fp16, [1, 8, 188, 128]> transpose_238 = transpose(perm = transpose_238_perm_0, x = q_u_23_cast_fp16)[name = string("transpose_390")];
            tensor<fp16, [1, 8, 188, 188]> var_4064_cast_fp16 = matmul(transpose_x = var_4064_transpose_x_0, transpose_y = var_4064_transpose_y_0, x = transpose_238, y = transpose_239)[name = string("op_4064_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4065_cast_fp16 = add(x = var_4064_cast_fp16, y = bd_23_cast_fp16)[name = string("op_4065_cast_fp16")];
            fp16 var_4066_to_fp16 = const()[name = string("op_4066_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_4067_cast_fp16 = mul(x = var_4065_cast_fp16, y = var_4066_to_fp16)[name = string("op_4067_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_47_cast_fp16 = add(x = var_4067_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_47_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4069_cast_fp16 = softmax(axis = var_3992, x = scores_47_cast_fp16)[name = string("op_4069_cast_fp16")];
            bool out_23_transpose_x_0 = const()[name = string("out_23_transpose_x_0"), val = bool(false)];
            bool out_23_transpose_y_0 = const()[name = string("out_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_4038_cast_fp16)[name = string("transpose_388")];
            tensor<fp16, [1, 8, 188, 128]> out_23_cast_fp16 = matmul(transpose_x = out_23_transpose_x_0, transpose_y = out_23_transpose_y_0, x = var_4069_cast_fp16, y = v_23_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [4]> var_4071_perm_0 = const()[name = string("op_4071_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4072 = const()[name = string("op_4072"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_4071_cast_fp16 = transpose(perm = var_4071_perm_0, x = out_23_cast_fp16)[name = string("transpose_387")];
            tensor<fp16, [1, 188, 1024]> input_483_cast_fp16 = reshape(shape = var_4072, x = var_4071_cast_fp16)[name = string("input_483_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_11_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297544960))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298593600))))[name = string("blocks_11_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_11_o_weight_to_fp16_quantized, x = input_483_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_485_cast_fp16 = add(x = input_477_cast_fp16, y = linear_130_cast_fp16)[name = string("input_485_cast_fp16")];
            tensor<int32, [1]> input_487_axes_0 = const()[name = string("input_487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_11_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298595712)))];
            tensor<fp16, [1024]> enc_layers_11_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298597824)))];
            tensor<fp16, [1, 188, 1024]> input_487_cast_fp16 = layer_norm(axes = input_487_axes_0, beta = enc_layers_11_norm_conv_bias_to_fp16, epsilon = var_4000_to_fp16, gamma = enc_layers_11_norm_conv_weight_to_fp16, x = input_485_cast_fp16)[name = string("input_487_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_11_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298599936))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300697152))))[name = string("blocks_11_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_131_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_11_pw1_weight_to_fp16_quantized, x = input_487_cast_fp16)[name = string("linear_131_cast_fp16")];
            int32 var_4081_split_num_splits_0 = const()[name = string("op_4081_split_num_splits_0"), val = int32(2)];
            int32 var_4081_split_axis_0 = const()[name = string("op_4081_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_4081_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_4081_split_cast_fp16_1 = split(axis = var_4081_split_axis_0, num_splits = var_4081_split_num_splits_0, x = linear_131_cast_fp16)[name = string("op_4081_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4081_split_1_sigmoid_cast_fp16 = sigmoid(x = var_4081_split_cast_fp16_1)[name = string("op_4081_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4081_cast_fp16 = mul(x = var_4081_split_cast_fp16_0, y = var_4081_split_1_sigmoid_cast_fp16)[name = string("op_4081_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_45_cast_fp16 = mul(x = var_4081_cast_fp16, y = valid_cast_fp16)[name = string("h_45_cast_fp16")];
            tensor<int32, [3]> input_491_perm_0 = const()[name = string("input_491_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_493_pad_type_0 = const()[name = string("input_493_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_493_pad_0 = const()[name = string("input_493_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_493_groups_0 = const()[name = string("input_493_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_493_strides_0 = const()[name = string("input_493_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_493_dilations_0 = const()[name = string("input_493_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_60_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300701312))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300710592))))[name = string("const_60_to_fp16_quantized")];
            tensor<fp16, [1024]> const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300712704)))];
            tensor<fp16, [1, 1024, 188]> input_491_cast_fp16 = transpose(perm = input_491_perm_0, x = h_45_cast_fp16)[name = string("transpose_386")];
            tensor<fp16, [1, 1024, 188]> input_495_cast_fp16 = conv(bias = const_61_to_fp16, dilations = input_493_dilations_0, groups = input_493_groups_0, pad = input_493_pad_0, pad_type = input_493_pad_type_0, strides = input_493_strides_0, weight = const_60_to_fp16_quantized, x = input_491_cast_fp16)[name = string("input_495_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_47_cast_fp16 = silu(x = input_495_cast_fp16)[name = string("h_47_cast_fp16")];
            tensor<int32, [3]> input_497_perm_0 = const()[name = string("input_497_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_11_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300714816))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301763456))))[name = string("blocks_11_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_497_cast_fp16 = transpose(perm = input_497_perm_0, x = h_47_cast_fp16)[name = string("transpose_385")];
            tensor<fp16, [1, 188, 1024]> linear_132_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_11_pw2_weight_to_fp16_quantized, x = input_497_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_499_cast_fp16 = add(x = input_485_cast_fp16, y = linear_132_cast_fp16)[name = string("input_499_cast_fp16")];
            tensor<int32, [1]> input_501_axes_0 = const()[name = string("input_501_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301765568)))];
            tensor<fp16, [1024]> enc_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301767680)))];
            tensor<fp16, [1, 188, 1024]> input_501_cast_fp16 = layer_norm(axes = input_501_axes_0, beta = enc_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_4000_to_fp16, gamma = enc_layers_11_norm_feed_forward2_weight_to_fp16, x = input_499_cast_fp16)[name = string("input_501_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_11_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301769792))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305964160))))[name = string("blocks_11_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_133_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_11_ff2_1_weight_to_fp16_quantized, x = input_501_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_505_cast_fp16 = silu(x = linear_133_cast_fp16)[name = string("input_505_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_11_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305972416))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310166784))))[name = string("blocks_11_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_134_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_11_ff2_2_weight_to_fp16_quantized, x = input_505_cast_fp16)[name = string("linear_134_cast_fp16")];
            fp16 var_4102_to_fp16 = const()[name = string("op_4102_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4103_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_4102_to_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_507_cast_fp16 = add(x = input_499_cast_fp16, y = var_4103_cast_fp16)[name = string("input_507_cast_fp16")];
            tensor<int32, [1]> input_509_axes_0 = const()[name = string("input_509_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_11_norm_out_weight_to_fp16 = const()[name = string("enc_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310168896)))];
            tensor<fp16, [1024]> enc_layers_11_norm_out_bias_to_fp16 = const()[name = string("enc_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310171008)))];
            tensor<fp16, [1, 188, 1024]> input_509_cast_fp16 = layer_norm(axes = input_509_axes_0, beta = enc_layers_11_norm_out_bias_to_fp16, epsilon = var_4000_to_fp16, gamma = enc_layers_11_norm_out_weight_to_fp16, x = input_507_cast_fp16)[name = string("input_509_cast_fp16")];
            int32 var_4116 = const()[name = string("op_4116"), val = int32(-1)];
            tensor<int32, [1]> input_511_axes_0 = const()[name = string("input_511_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310173120)))];
            tensor<fp16, [1024]> enc_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310175232)))];
            fp16 var_4124_to_fp16 = const()[name = string("op_4124_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_511_cast_fp16 = layer_norm(axes = input_511_axes_0, beta = enc_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_4124_to_fp16, gamma = enc_layers_12_norm_feed_forward1_weight_to_fp16, x = input_509_cast_fp16)[name = string("input_511_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_12_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310177344))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314371712))))[name = string("blocks_12_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_135_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_12_ff1_1_weight_to_fp16_quantized, x = input_511_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_515_cast_fp16 = silu(x = linear_135_cast_fp16)[name = string("input_515_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_12_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314379968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318574336))))[name = string("blocks_12_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_136_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_12_ff1_2_weight_to_fp16_quantized, x = input_515_cast_fp16)[name = string("linear_136_cast_fp16")];
            fp16 var_4144_to_fp16 = const()[name = string("op_4144_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4145_cast_fp16 = mul(x = linear_136_cast_fp16, y = var_4144_to_fp16)[name = string("op_4145_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_517_cast_fp16 = add(x = input_509_cast_fp16, y = var_4145_cast_fp16)[name = string("input_517_cast_fp16")];
            tensor<int32, [1]> input_519_axes_0 = const()[name = string("input_519_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318576448)))];
            tensor<fp16, [1024]> enc_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318578560)))];
            tensor<fp16, [1, 188, 1024]> input_519_cast_fp16 = layer_norm(axes = input_519_axes_0, beta = enc_layers_12_norm_self_att_bias_to_fp16, epsilon = var_4124_to_fp16, gamma = enc_layers_12_norm_self_att_weight_to_fp16, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_12_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318580672))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319629312))))[name = string("blocks_12_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_12_q_weight_to_fp16_quantized, x = input_519_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<int32, [4]> var_4151 = const()[name = string("op_4151"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4152_cast_fp16 = reshape(shape = var_4151, x = linear_137_cast_fp16)[name = string("op_4152_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_12_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319631424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320680064))))[name = string("blocks_12_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_12_k_weight_to_fp16_quantized, x = input_519_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<int32, [4]> var_4156 = const()[name = string("op_4156"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4157_cast_fp16 = reshape(shape = var_4156, x = linear_138_cast_fp16)[name = string("op_4157_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_12_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320682176))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321730816))))[name = string("blocks_12_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_139_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_12_v_weight_to_fp16_quantized, x = input_519_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<int32, [4]> var_4161 = const()[name = string("op_4161"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4162_cast_fp16 = reshape(shape = var_4161, x = linear_139_cast_fp16)[name = string("op_4162_cast_fp16")];
            tensor<int32, [4]> v_25_perm_0 = const()[name = string("v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_110_to_fp16 = const()[name = string("const_110_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321732928)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_25_cast_fp16 = add(x = var_4152_cast_fp16, y = const_110_to_fp16)[name = string("q_u_25_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_111_to_fp16 = const()[name = string("const_111_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321735040)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_25_cast_fp16 = add(x = var_4152_cast_fp16, y = const_111_to_fp16)[name = string("q_v_25_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_12_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321737152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322785792))))[name = string("blocks_12_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_140_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_12_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_4172 = const()[name = string("op_4172"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_25_cast_fp16 = reshape(shape = var_4172, x = linear_140_cast_fp16)[name = string("p_25_cast_fp16")];
            bool input_521_transpose_x_0 = const()[name = string("input_521_transpose_x_0"), val = bool(false)];
            bool input_521_transpose_y_0 = const()[name = string("input_521_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_240_perm_0 = const()[name = string("transpose_240_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_241_perm_0 = const()[name = string("transpose_241_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_241 = transpose(perm = transpose_241_perm_0, x = p_25_cast_fp16)[name = string("transpose_383")];
            tensor<fp16, [1, 8, 188, 128]> transpose_240 = transpose(perm = transpose_240_perm_0, x = q_v_25_cast_fp16)[name = string("transpose_384")];
            tensor<fp16, [1, 8, 188, 375]> input_521_cast_fp16 = matmul(transpose_x = input_521_transpose_x_0, transpose_y = input_521_transpose_y_0, x = transpose_240, y = transpose_241)[name = string("input_521_cast_fp16")];
            tensor<int32, [8]> var_4178_pad_0 = const()[name = string("op_4178_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_4178_mode_0 = const()[name = string("op_4178_mode_0"), val = string("constant")];
            fp16 const_26_to_fp16 = const()[name = string("const_26_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_4178_cast_fp16 = pad(constant_val = const_26_to_fp16, mode = var_4178_mode_0, pad = var_4178_pad_0, x = input_521_cast_fp16)[name = string("op_4178_cast_fp16")];
            tensor<int32, [4]> var_4179 = const()[name = string("op_4179"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_49_cast_fp16 = reshape(shape = var_4179, x = var_4178_cast_fp16)[name = string("scores_49_cast_fp16")];
            tensor<int32, [4]> var_4183_begin_0 = const()[name = string("op_4183_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4183_end_0 = const()[name = string("op_4183_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_4183_end_mask_0 = const()[name = string("op_4183_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_4183_cast_fp16 = slice_by_index(begin = var_4183_begin_0, end = var_4183_end_0, end_mask = var_4183_end_mask_0, x = scores_49_cast_fp16)[name = string("op_4183_cast_fp16")];
            tensor<int32, [4]> var_4184 = const()[name = string("op_4184"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_4185_cast_fp16 = reshape(shape = var_4184, x = var_4183_cast_fp16)[name = string("op_4185_cast_fp16")];
            tensor<int32, [4]> bd_25_begin_0 = const()[name = string("bd_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_25_end_0 = const()[name = string("bd_25_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_25_end_mask_0 = const()[name = string("bd_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_25_cast_fp16 = slice_by_index(begin = bd_25_begin_0, end = bd_25_end_0, end_mask = bd_25_end_mask_0, x = var_4185_cast_fp16)[name = string("bd_25_cast_fp16")];
            bool var_4188_transpose_x_0 = const()[name = string("op_4188_transpose_x_0"), val = bool(false)];
            bool var_4188_transpose_y_0 = const()[name = string("op_4188_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_242_perm_0 = const()[name = string("transpose_242_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_243_perm_0 = const()[name = string("transpose_243_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_243 = transpose(perm = transpose_243_perm_0, x = var_4157_cast_fp16)[name = string("transpose_381")];
            tensor<fp16, [1, 8, 188, 128]> transpose_242 = transpose(perm = transpose_242_perm_0, x = q_u_25_cast_fp16)[name = string("transpose_382")];
            tensor<fp16, [1, 8, 188, 188]> var_4188_cast_fp16 = matmul(transpose_x = var_4188_transpose_x_0, transpose_y = var_4188_transpose_y_0, x = transpose_242, y = transpose_243)[name = string("op_4188_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4189_cast_fp16 = add(x = var_4188_cast_fp16, y = bd_25_cast_fp16)[name = string("op_4189_cast_fp16")];
            fp16 var_4190_to_fp16 = const()[name = string("op_4190_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_4191_cast_fp16 = mul(x = var_4189_cast_fp16, y = var_4190_to_fp16)[name = string("op_4191_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_51_cast_fp16 = add(x = var_4191_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_51_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4193_cast_fp16 = softmax(axis = var_4116, x = scores_51_cast_fp16)[name = string("op_4193_cast_fp16")];
            bool out_25_transpose_x_0 = const()[name = string("out_25_transpose_x_0"), val = bool(false)];
            bool out_25_transpose_y_0 = const()[name = string("out_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_4162_cast_fp16)[name = string("transpose_380")];
            tensor<fp16, [1, 8, 188, 128]> out_25_cast_fp16 = matmul(transpose_x = out_25_transpose_x_0, transpose_y = out_25_transpose_y_0, x = var_4193_cast_fp16, y = v_25_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [4]> var_4195_perm_0 = const()[name = string("op_4195_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4196 = const()[name = string("op_4196"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_4195_cast_fp16 = transpose(perm = var_4195_perm_0, x = out_25_cast_fp16)[name = string("transpose_379")];
            tensor<fp16, [1, 188, 1024]> input_523_cast_fp16 = reshape(shape = var_4196, x = var_4195_cast_fp16)[name = string("input_523_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_12_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322787904))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323836544))))[name = string("blocks_12_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_141_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_12_o_weight_to_fp16_quantized, x = input_523_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_525_cast_fp16 = add(x = input_517_cast_fp16, y = linear_141_cast_fp16)[name = string("input_525_cast_fp16")];
            tensor<int32, [1]> input_527_axes_0 = const()[name = string("input_527_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_12_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323838656)))];
            tensor<fp16, [1024]> enc_layers_12_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323840768)))];
            tensor<fp16, [1, 188, 1024]> input_527_cast_fp16 = layer_norm(axes = input_527_axes_0, beta = enc_layers_12_norm_conv_bias_to_fp16, epsilon = var_4124_to_fp16, gamma = enc_layers_12_norm_conv_weight_to_fp16, x = input_525_cast_fp16)[name = string("input_527_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_12_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323842880))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325940096))))[name = string("blocks_12_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_142_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_12_pw1_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("linear_142_cast_fp16")];
            int32 var_4205_split_num_splits_0 = const()[name = string("op_4205_split_num_splits_0"), val = int32(2)];
            int32 var_4205_split_axis_0 = const()[name = string("op_4205_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_4205_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_4205_split_cast_fp16_1 = split(axis = var_4205_split_axis_0, num_splits = var_4205_split_num_splits_0, x = linear_142_cast_fp16)[name = string("op_4205_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4205_split_1_sigmoid_cast_fp16 = sigmoid(x = var_4205_split_cast_fp16_1)[name = string("op_4205_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4205_cast_fp16 = mul(x = var_4205_split_cast_fp16_0, y = var_4205_split_1_sigmoid_cast_fp16)[name = string("op_4205_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_49_cast_fp16 = mul(x = var_4205_cast_fp16, y = valid_cast_fp16)[name = string("h_49_cast_fp16")];
            tensor<int32, [3]> input_531_perm_0 = const()[name = string("input_531_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_533_pad_type_0 = const()[name = string("input_533_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_533_pad_0 = const()[name = string("input_533_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_533_groups_0 = const()[name = string("input_533_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_533_strides_0 = const()[name = string("input_533_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_533_dilations_0 = const()[name = string("input_533_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_62_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325944256))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325953536))))[name = string("const_62_to_fp16_quantized")];
            tensor<fp16, [1024]> const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325955648)))];
            tensor<fp16, [1, 1024, 188]> input_531_cast_fp16 = transpose(perm = input_531_perm_0, x = h_49_cast_fp16)[name = string("transpose_378")];
            tensor<fp16, [1, 1024, 188]> input_535_cast_fp16 = conv(bias = const_63_to_fp16, dilations = input_533_dilations_0, groups = input_533_groups_0, pad = input_533_pad_0, pad_type = input_533_pad_type_0, strides = input_533_strides_0, weight = const_62_to_fp16_quantized, x = input_531_cast_fp16)[name = string("input_535_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_51_cast_fp16 = silu(x = input_535_cast_fp16)[name = string("h_51_cast_fp16")];
            tensor<int32, [3]> input_537_perm_0 = const()[name = string("input_537_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_12_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325957760))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327006400))))[name = string("blocks_12_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_537_cast_fp16 = transpose(perm = input_537_perm_0, x = h_51_cast_fp16)[name = string("transpose_377")];
            tensor<fp16, [1, 188, 1024]> linear_143_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_12_pw2_weight_to_fp16_quantized, x = input_537_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_539_cast_fp16 = add(x = input_525_cast_fp16, y = linear_143_cast_fp16)[name = string("input_539_cast_fp16")];
            tensor<int32, [1]> input_541_axes_0 = const()[name = string("input_541_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327008512)))];
            tensor<fp16, [1024]> enc_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327010624)))];
            tensor<fp16, [1, 188, 1024]> input_541_cast_fp16 = layer_norm(axes = input_541_axes_0, beta = enc_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_4124_to_fp16, gamma = enc_layers_12_norm_feed_forward2_weight_to_fp16, x = input_539_cast_fp16)[name = string("input_541_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_12_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327012736))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331207104))))[name = string("blocks_12_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_144_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_12_ff2_1_weight_to_fp16_quantized, x = input_541_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_545_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("input_545_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_12_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331215360))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335409728))))[name = string("blocks_12_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_12_ff2_2_weight_to_fp16_quantized, x = input_545_cast_fp16)[name = string("linear_145_cast_fp16")];
            fp16 var_4226_to_fp16 = const()[name = string("op_4226_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4227_cast_fp16 = mul(x = linear_145_cast_fp16, y = var_4226_to_fp16)[name = string("op_4227_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_547_cast_fp16 = add(x = input_539_cast_fp16, y = var_4227_cast_fp16)[name = string("input_547_cast_fp16")];
            tensor<int32, [1]> input_549_axes_0 = const()[name = string("input_549_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_12_norm_out_weight_to_fp16 = const()[name = string("enc_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335411840)))];
            tensor<fp16, [1024]> enc_layers_12_norm_out_bias_to_fp16 = const()[name = string("enc_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335413952)))];
            tensor<fp16, [1, 188, 1024]> input_549_cast_fp16 = layer_norm(axes = input_549_axes_0, beta = enc_layers_12_norm_out_bias_to_fp16, epsilon = var_4124_to_fp16, gamma = enc_layers_12_norm_out_weight_to_fp16, x = input_547_cast_fp16)[name = string("input_549_cast_fp16")];
            int32 var_4240 = const()[name = string("op_4240"), val = int32(-1)];
            tensor<int32, [1]> input_551_axes_0 = const()[name = string("input_551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335416064)))];
            tensor<fp16, [1024]> enc_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335418176)))];
            fp16 var_4248_to_fp16 = const()[name = string("op_4248_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_551_cast_fp16 = layer_norm(axes = input_551_axes_0, beta = enc_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_4248_to_fp16, gamma = enc_layers_13_norm_feed_forward1_weight_to_fp16, x = input_549_cast_fp16)[name = string("input_551_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_13_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335420288))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339614656))))[name = string("blocks_13_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_13_ff1_1_weight_to_fp16_quantized, x = input_551_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_555_cast_fp16 = silu(x = linear_146_cast_fp16)[name = string("input_555_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_13_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339622912))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343817280))))[name = string("blocks_13_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_147_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_13_ff1_2_weight_to_fp16_quantized, x = input_555_cast_fp16)[name = string("linear_147_cast_fp16")];
            fp16 var_4268_to_fp16 = const()[name = string("op_4268_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4269_cast_fp16 = mul(x = linear_147_cast_fp16, y = var_4268_to_fp16)[name = string("op_4269_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_557_cast_fp16 = add(x = input_549_cast_fp16, y = var_4269_cast_fp16)[name = string("input_557_cast_fp16")];
            tensor<int32, [1]> input_559_axes_0 = const()[name = string("input_559_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343819392)))];
            tensor<fp16, [1024]> enc_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343821504)))];
            tensor<fp16, [1, 188, 1024]> input_559_cast_fp16 = layer_norm(axes = input_559_axes_0, beta = enc_layers_13_norm_self_att_bias_to_fp16, epsilon = var_4248_to_fp16, gamma = enc_layers_13_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("input_559_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_13_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343823616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344872256))))[name = string("blocks_13_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_148_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_13_q_weight_to_fp16_quantized, x = input_559_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_4275 = const()[name = string("op_4275"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4276_cast_fp16 = reshape(shape = var_4275, x = linear_148_cast_fp16)[name = string("op_4276_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_13_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344874368))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345923008))))[name = string("blocks_13_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_149_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_13_k_weight_to_fp16_quantized, x = input_559_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_4280 = const()[name = string("op_4280"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4281_cast_fp16 = reshape(shape = var_4280, x = linear_149_cast_fp16)[name = string("op_4281_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_13_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345925120))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346973760))))[name = string("blocks_13_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_150_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_13_v_weight_to_fp16_quantized, x = input_559_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<int32, [4]> var_4285 = const()[name = string("op_4285"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4286_cast_fp16 = reshape(shape = var_4285, x = linear_150_cast_fp16)[name = string("op_4286_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_112_to_fp16 = const()[name = string("const_112_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346975872)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_27_cast_fp16 = add(x = var_4276_cast_fp16, y = const_112_to_fp16)[name = string("q_u_27_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_113_to_fp16 = const()[name = string("const_113_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346977984)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_27_cast_fp16 = add(x = var_4276_cast_fp16, y = const_113_to_fp16)[name = string("q_v_27_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_13_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346980096))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348028736))))[name = string("blocks_13_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_13_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<int32, [4]> var_4296 = const()[name = string("op_4296"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_27_cast_fp16 = reshape(shape = var_4296, x = linear_151_cast_fp16)[name = string("p_27_cast_fp16")];
            bool input_561_transpose_x_0 = const()[name = string("input_561_transpose_x_0"), val = bool(false)];
            bool input_561_transpose_y_0 = const()[name = string("input_561_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_244_perm_0 = const()[name = string("transpose_244_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_245_perm_0 = const()[name = string("transpose_245_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_245 = transpose(perm = transpose_245_perm_0, x = p_27_cast_fp16)[name = string("transpose_375")];
            tensor<fp16, [1, 8, 188, 128]> transpose_244 = transpose(perm = transpose_244_perm_0, x = q_v_27_cast_fp16)[name = string("transpose_376")];
            tensor<fp16, [1, 8, 188, 375]> input_561_cast_fp16 = matmul(transpose_x = input_561_transpose_x_0, transpose_y = input_561_transpose_y_0, x = transpose_244, y = transpose_245)[name = string("input_561_cast_fp16")];
            tensor<int32, [8]> var_4302_pad_0 = const()[name = string("op_4302_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_4302_mode_0 = const()[name = string("op_4302_mode_0"), val = string("constant")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_4302_cast_fp16 = pad(constant_val = const_27_to_fp16, mode = var_4302_mode_0, pad = var_4302_pad_0, x = input_561_cast_fp16)[name = string("op_4302_cast_fp16")];
            tensor<int32, [4]> var_4303 = const()[name = string("op_4303"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_53_cast_fp16 = reshape(shape = var_4303, x = var_4302_cast_fp16)[name = string("scores_53_cast_fp16")];
            tensor<int32, [4]> var_4307_begin_0 = const()[name = string("op_4307_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4307_end_0 = const()[name = string("op_4307_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_4307_end_mask_0 = const()[name = string("op_4307_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_4307_cast_fp16 = slice_by_index(begin = var_4307_begin_0, end = var_4307_end_0, end_mask = var_4307_end_mask_0, x = scores_53_cast_fp16)[name = string("op_4307_cast_fp16")];
            tensor<int32, [4]> var_4308 = const()[name = string("op_4308"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_4309_cast_fp16 = reshape(shape = var_4308, x = var_4307_cast_fp16)[name = string("op_4309_cast_fp16")];
            tensor<int32, [4]> bd_27_begin_0 = const()[name = string("bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_27_end_0 = const()[name = string("bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_27_end_mask_0 = const()[name = string("bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_27_cast_fp16 = slice_by_index(begin = bd_27_begin_0, end = bd_27_end_0, end_mask = bd_27_end_mask_0, x = var_4309_cast_fp16)[name = string("bd_27_cast_fp16")];
            bool var_4312_transpose_x_0 = const()[name = string("op_4312_transpose_x_0"), val = bool(false)];
            bool var_4312_transpose_y_0 = const()[name = string("op_4312_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_246_perm_0 = const()[name = string("transpose_246_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_247_perm_0 = const()[name = string("transpose_247_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_247 = transpose(perm = transpose_247_perm_0, x = var_4281_cast_fp16)[name = string("transpose_373")];
            tensor<fp16, [1, 8, 188, 128]> transpose_246 = transpose(perm = transpose_246_perm_0, x = q_u_27_cast_fp16)[name = string("transpose_374")];
            tensor<fp16, [1, 8, 188, 188]> var_4312_cast_fp16 = matmul(transpose_x = var_4312_transpose_x_0, transpose_y = var_4312_transpose_y_0, x = transpose_246, y = transpose_247)[name = string("op_4312_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4313_cast_fp16 = add(x = var_4312_cast_fp16, y = bd_27_cast_fp16)[name = string("op_4313_cast_fp16")];
            fp16 var_4314_to_fp16 = const()[name = string("op_4314_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_4315_cast_fp16 = mul(x = var_4313_cast_fp16, y = var_4314_to_fp16)[name = string("op_4315_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_55_cast_fp16 = add(x = var_4315_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_55_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4317_cast_fp16 = softmax(axis = var_4240, x = scores_55_cast_fp16)[name = string("op_4317_cast_fp16")];
            bool out_27_transpose_x_0 = const()[name = string("out_27_transpose_x_0"), val = bool(false)];
            bool out_27_transpose_y_0 = const()[name = string("out_27_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_4286_cast_fp16)[name = string("transpose_372")];
            tensor<fp16, [1, 8, 188, 128]> out_27_cast_fp16 = matmul(transpose_x = out_27_transpose_x_0, transpose_y = out_27_transpose_y_0, x = var_4317_cast_fp16, y = v_27_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<int32, [4]> var_4319_perm_0 = const()[name = string("op_4319_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4320 = const()[name = string("op_4320"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_4319_cast_fp16 = transpose(perm = var_4319_perm_0, x = out_27_cast_fp16)[name = string("transpose_371")];
            tensor<fp16, [1, 188, 1024]> input_563_cast_fp16 = reshape(shape = var_4320, x = var_4319_cast_fp16)[name = string("input_563_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_13_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348030848))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349079488))))[name = string("blocks_13_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_13_o_weight_to_fp16_quantized, x = input_563_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_565_cast_fp16 = add(x = input_557_cast_fp16, y = linear_152_cast_fp16)[name = string("input_565_cast_fp16")];
            tensor<int32, [1]> input_567_axes_0 = const()[name = string("input_567_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_13_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349081600)))];
            tensor<fp16, [1024]> enc_layers_13_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349083712)))];
            tensor<fp16, [1, 188, 1024]> input_567_cast_fp16 = layer_norm(axes = input_567_axes_0, beta = enc_layers_13_norm_conv_bias_to_fp16, epsilon = var_4248_to_fp16, gamma = enc_layers_13_norm_conv_weight_to_fp16, x = input_565_cast_fp16)[name = string("input_567_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_13_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349085824))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351183040))))[name = string("blocks_13_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_153_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_13_pw1_weight_to_fp16_quantized, x = input_567_cast_fp16)[name = string("linear_153_cast_fp16")];
            int32 var_4329_split_num_splits_0 = const()[name = string("op_4329_split_num_splits_0"), val = int32(2)];
            int32 var_4329_split_axis_0 = const()[name = string("op_4329_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_4329_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_4329_split_cast_fp16_1 = split(axis = var_4329_split_axis_0, num_splits = var_4329_split_num_splits_0, x = linear_153_cast_fp16)[name = string("op_4329_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4329_split_1_sigmoid_cast_fp16 = sigmoid(x = var_4329_split_cast_fp16_1)[name = string("op_4329_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4329_cast_fp16 = mul(x = var_4329_split_cast_fp16_0, y = var_4329_split_1_sigmoid_cast_fp16)[name = string("op_4329_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_53_cast_fp16 = mul(x = var_4329_cast_fp16, y = valid_cast_fp16)[name = string("h_53_cast_fp16")];
            tensor<int32, [3]> input_571_perm_0 = const()[name = string("input_571_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_573_pad_type_0 = const()[name = string("input_573_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_573_pad_0 = const()[name = string("input_573_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_573_groups_0 = const()[name = string("input_573_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_573_strides_0 = const()[name = string("input_573_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_573_dilations_0 = const()[name = string("input_573_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_64_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351187200))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351196480))))[name = string("const_64_to_fp16_quantized")];
            tensor<fp16, [1024]> const_65_to_fp16 = const()[name = string("const_65_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351198592)))];
            tensor<fp16, [1, 1024, 188]> input_571_cast_fp16 = transpose(perm = input_571_perm_0, x = h_53_cast_fp16)[name = string("transpose_370")];
            tensor<fp16, [1, 1024, 188]> input_575_cast_fp16 = conv(bias = const_65_to_fp16, dilations = input_573_dilations_0, groups = input_573_groups_0, pad = input_573_pad_0, pad_type = input_573_pad_type_0, strides = input_573_strides_0, weight = const_64_to_fp16_quantized, x = input_571_cast_fp16)[name = string("input_575_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_55_cast_fp16 = silu(x = input_575_cast_fp16)[name = string("h_55_cast_fp16")];
            tensor<int32, [3]> input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_13_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351200704))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352249344))))[name = string("blocks_13_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = h_55_cast_fp16)[name = string("transpose_369")];
            tensor<fp16, [1, 188, 1024]> linear_154_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_13_pw2_weight_to_fp16_quantized, x = input_577_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_579_cast_fp16 = add(x = input_565_cast_fp16, y = linear_154_cast_fp16)[name = string("input_579_cast_fp16")];
            tensor<int32, [1]> input_581_axes_0 = const()[name = string("input_581_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352251456)))];
            tensor<fp16, [1024]> enc_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352253568)))];
            tensor<fp16, [1, 188, 1024]> input_581_cast_fp16 = layer_norm(axes = input_581_axes_0, beta = enc_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_4248_to_fp16, gamma = enc_layers_13_norm_feed_forward2_weight_to_fp16, x = input_579_cast_fp16)[name = string("input_581_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_13_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352255680))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356450048))))[name = string("blocks_13_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_155_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_13_ff2_1_weight_to_fp16_quantized, x = input_581_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_585_cast_fp16 = silu(x = linear_155_cast_fp16)[name = string("input_585_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_13_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356458304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360652672))))[name = string("blocks_13_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_156_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_13_ff2_2_weight_to_fp16_quantized, x = input_585_cast_fp16)[name = string("linear_156_cast_fp16")];
            fp16 var_4350_to_fp16 = const()[name = string("op_4350_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4351_cast_fp16 = mul(x = linear_156_cast_fp16, y = var_4350_to_fp16)[name = string("op_4351_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_587_cast_fp16 = add(x = input_579_cast_fp16, y = var_4351_cast_fp16)[name = string("input_587_cast_fp16")];
            tensor<int32, [1]> input_589_axes_0 = const()[name = string("input_589_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_13_norm_out_weight_to_fp16 = const()[name = string("enc_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360654784)))];
            tensor<fp16, [1024]> enc_layers_13_norm_out_bias_to_fp16 = const()[name = string("enc_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360656896)))];
            tensor<fp16, [1, 188, 1024]> input_589_cast_fp16 = layer_norm(axes = input_589_axes_0, beta = enc_layers_13_norm_out_bias_to_fp16, epsilon = var_4248_to_fp16, gamma = enc_layers_13_norm_out_weight_to_fp16, x = input_587_cast_fp16)[name = string("input_589_cast_fp16")];
            int32 var_4364 = const()[name = string("op_4364"), val = int32(-1)];
            tensor<int32, [1]> input_591_axes_0 = const()[name = string("input_591_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360659008)))];
            tensor<fp16, [1024]> enc_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360661120)))];
            fp16 var_4372_to_fp16 = const()[name = string("op_4372_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_591_cast_fp16 = layer_norm(axes = input_591_axes_0, beta = enc_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_4372_to_fp16, gamma = enc_layers_14_norm_feed_forward1_weight_to_fp16, x = input_589_cast_fp16)[name = string("input_591_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_14_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360663232))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364857600))))[name = string("blocks_14_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_14_ff1_1_weight_to_fp16_quantized, x = input_591_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_595_cast_fp16 = silu(x = linear_157_cast_fp16)[name = string("input_595_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_14_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364865856))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369060224))))[name = string("blocks_14_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_14_ff1_2_weight_to_fp16_quantized, x = input_595_cast_fp16)[name = string("linear_158_cast_fp16")];
            fp16 var_4392_to_fp16 = const()[name = string("op_4392_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4393_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_4392_to_fp16)[name = string("op_4393_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_597_cast_fp16 = add(x = input_589_cast_fp16, y = var_4393_cast_fp16)[name = string("input_597_cast_fp16")];
            tensor<int32, [1]> input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369062336)))];
            tensor<fp16, [1024]> enc_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369064448)))];
            tensor<fp16, [1, 188, 1024]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = enc_layers_14_norm_self_att_bias_to_fp16, epsilon = var_4372_to_fp16, gamma = enc_layers_14_norm_self_att_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_14_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369066560))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370115200))))[name = string("blocks_14_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_14_q_weight_to_fp16_quantized, x = input_599_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<int32, [4]> var_4399 = const()[name = string("op_4399"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4400_cast_fp16 = reshape(shape = var_4399, x = linear_159_cast_fp16)[name = string("op_4400_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_14_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370117312))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371165952))))[name = string("blocks_14_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_160_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_14_k_weight_to_fp16_quantized, x = input_599_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<int32, [4]> var_4404 = const()[name = string("op_4404"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4405_cast_fp16 = reshape(shape = var_4404, x = linear_160_cast_fp16)[name = string("op_4405_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_14_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371168064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372216704))))[name = string("blocks_14_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_161_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_14_v_weight_to_fp16_quantized, x = input_599_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_4409 = const()[name = string("op_4409"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4410_cast_fp16 = reshape(shape = var_4409, x = linear_161_cast_fp16)[name = string("op_4410_cast_fp16")];
            tensor<int32, [4]> v_29_perm_0 = const()[name = string("v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_114_to_fp16 = const()[name = string("const_114_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372218816)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_29_cast_fp16 = add(x = var_4400_cast_fp16, y = const_114_to_fp16)[name = string("q_u_29_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_115_to_fp16 = const()[name = string("const_115_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372220928)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_29_cast_fp16 = add(x = var_4400_cast_fp16, y = const_115_to_fp16)[name = string("q_v_29_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_14_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372223040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373271680))))[name = string("blocks_14_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_162_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_14_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_4420 = const()[name = string("op_4420"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_29_cast_fp16 = reshape(shape = var_4420, x = linear_162_cast_fp16)[name = string("p_29_cast_fp16")];
            bool input_601_transpose_x_0 = const()[name = string("input_601_transpose_x_0"), val = bool(false)];
            bool input_601_transpose_y_0 = const()[name = string("input_601_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_248_perm_0 = const()[name = string("transpose_248_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_249_perm_0 = const()[name = string("transpose_249_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_249 = transpose(perm = transpose_249_perm_0, x = p_29_cast_fp16)[name = string("transpose_367")];
            tensor<fp16, [1, 8, 188, 128]> transpose_248 = transpose(perm = transpose_248_perm_0, x = q_v_29_cast_fp16)[name = string("transpose_368")];
            tensor<fp16, [1, 8, 188, 375]> input_601_cast_fp16 = matmul(transpose_x = input_601_transpose_x_0, transpose_y = input_601_transpose_y_0, x = transpose_248, y = transpose_249)[name = string("input_601_cast_fp16")];
            tensor<int32, [8]> var_4426_pad_0 = const()[name = string("op_4426_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_4426_mode_0 = const()[name = string("op_4426_mode_0"), val = string("constant")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_4426_cast_fp16 = pad(constant_val = const_28_to_fp16, mode = var_4426_mode_0, pad = var_4426_pad_0, x = input_601_cast_fp16)[name = string("op_4426_cast_fp16")];
            tensor<int32, [4]> var_4427 = const()[name = string("op_4427"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_57_cast_fp16 = reshape(shape = var_4427, x = var_4426_cast_fp16)[name = string("scores_57_cast_fp16")];
            tensor<int32, [4]> var_4431_begin_0 = const()[name = string("op_4431_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4431_end_0 = const()[name = string("op_4431_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_4431_end_mask_0 = const()[name = string("op_4431_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_4431_cast_fp16 = slice_by_index(begin = var_4431_begin_0, end = var_4431_end_0, end_mask = var_4431_end_mask_0, x = scores_57_cast_fp16)[name = string("op_4431_cast_fp16")];
            tensor<int32, [4]> var_4432 = const()[name = string("op_4432"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_4433_cast_fp16 = reshape(shape = var_4432, x = var_4431_cast_fp16)[name = string("op_4433_cast_fp16")];
            tensor<int32, [4]> bd_29_begin_0 = const()[name = string("bd_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_29_end_0 = const()[name = string("bd_29_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_29_end_mask_0 = const()[name = string("bd_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_29_cast_fp16 = slice_by_index(begin = bd_29_begin_0, end = bd_29_end_0, end_mask = bd_29_end_mask_0, x = var_4433_cast_fp16)[name = string("bd_29_cast_fp16")];
            bool var_4436_transpose_x_0 = const()[name = string("op_4436_transpose_x_0"), val = bool(false)];
            bool var_4436_transpose_y_0 = const()[name = string("op_4436_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_250_perm_0 = const()[name = string("transpose_250_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_251_perm_0 = const()[name = string("transpose_251_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_251 = transpose(perm = transpose_251_perm_0, x = var_4405_cast_fp16)[name = string("transpose_365")];
            tensor<fp16, [1, 8, 188, 128]> transpose_250 = transpose(perm = transpose_250_perm_0, x = q_u_29_cast_fp16)[name = string("transpose_366")];
            tensor<fp16, [1, 8, 188, 188]> var_4436_cast_fp16 = matmul(transpose_x = var_4436_transpose_x_0, transpose_y = var_4436_transpose_y_0, x = transpose_250, y = transpose_251)[name = string("op_4436_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4437_cast_fp16 = add(x = var_4436_cast_fp16, y = bd_29_cast_fp16)[name = string("op_4437_cast_fp16")];
            fp16 var_4438_to_fp16 = const()[name = string("op_4438_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_4439_cast_fp16 = mul(x = var_4437_cast_fp16, y = var_4438_to_fp16)[name = string("op_4439_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_59_cast_fp16 = add(x = var_4439_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_59_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4441_cast_fp16 = softmax(axis = var_4364, x = scores_59_cast_fp16)[name = string("op_4441_cast_fp16")];
            bool out_29_transpose_x_0 = const()[name = string("out_29_transpose_x_0"), val = bool(false)];
            bool out_29_transpose_y_0 = const()[name = string("out_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_4410_cast_fp16)[name = string("transpose_364")];
            tensor<fp16, [1, 8, 188, 128]> out_29_cast_fp16 = matmul(transpose_x = out_29_transpose_x_0, transpose_y = out_29_transpose_y_0, x = var_4441_cast_fp16, y = v_29_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<int32, [4]> var_4443_perm_0 = const()[name = string("op_4443_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4444 = const()[name = string("op_4444"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_4443_cast_fp16 = transpose(perm = var_4443_perm_0, x = out_29_cast_fp16)[name = string("transpose_363")];
            tensor<fp16, [1, 188, 1024]> input_603_cast_fp16 = reshape(shape = var_4444, x = var_4443_cast_fp16)[name = string("input_603_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_14_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373273792))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374322432))))[name = string("blocks_14_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_163_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_14_o_weight_to_fp16_quantized, x = input_603_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_605_cast_fp16 = add(x = input_597_cast_fp16, y = linear_163_cast_fp16)[name = string("input_605_cast_fp16")];
            tensor<int32, [1]> input_607_axes_0 = const()[name = string("input_607_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_14_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374324544)))];
            tensor<fp16, [1024]> enc_layers_14_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374326656)))];
            tensor<fp16, [1, 188, 1024]> input_607_cast_fp16 = layer_norm(axes = input_607_axes_0, beta = enc_layers_14_norm_conv_bias_to_fp16, epsilon = var_4372_to_fp16, gamma = enc_layers_14_norm_conv_weight_to_fp16, x = input_605_cast_fp16)[name = string("input_607_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_14_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374328768))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376425984))))[name = string("blocks_14_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_164_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_14_pw1_weight_to_fp16_quantized, x = input_607_cast_fp16)[name = string("linear_164_cast_fp16")];
            int32 var_4453_split_num_splits_0 = const()[name = string("op_4453_split_num_splits_0"), val = int32(2)];
            int32 var_4453_split_axis_0 = const()[name = string("op_4453_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_4453_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_4453_split_cast_fp16_1 = split(axis = var_4453_split_axis_0, num_splits = var_4453_split_num_splits_0, x = linear_164_cast_fp16)[name = string("op_4453_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4453_split_1_sigmoid_cast_fp16 = sigmoid(x = var_4453_split_cast_fp16_1)[name = string("op_4453_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4453_cast_fp16 = mul(x = var_4453_split_cast_fp16_0, y = var_4453_split_1_sigmoid_cast_fp16)[name = string("op_4453_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_57_cast_fp16 = mul(x = var_4453_cast_fp16, y = valid_cast_fp16)[name = string("h_57_cast_fp16")];
            tensor<int32, [3]> input_611_perm_0 = const()[name = string("input_611_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_613_pad_type_0 = const()[name = string("input_613_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_613_pad_0 = const()[name = string("input_613_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_613_groups_0 = const()[name = string("input_613_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_613_strides_0 = const()[name = string("input_613_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_613_dilations_0 = const()[name = string("input_613_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_66_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376430144))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376439424))))[name = string("const_66_to_fp16_quantized")];
            tensor<fp16, [1024]> const_67_to_fp16 = const()[name = string("const_67_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376441536)))];
            tensor<fp16, [1, 1024, 188]> input_611_cast_fp16 = transpose(perm = input_611_perm_0, x = h_57_cast_fp16)[name = string("transpose_362")];
            tensor<fp16, [1, 1024, 188]> input_615_cast_fp16 = conv(bias = const_67_to_fp16, 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 = const_66_to_fp16_quantized, x = input_611_cast_fp16)[name = string("input_615_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_59_cast_fp16 = silu(x = input_615_cast_fp16)[name = string("h_59_cast_fp16")];
            tensor<int32, [3]> input_617_perm_0 = const()[name = string("input_617_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_14_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376443648))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377492288))))[name = string("blocks_14_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_617_cast_fp16 = transpose(perm = input_617_perm_0, x = h_59_cast_fp16)[name = string("transpose_361")];
            tensor<fp16, [1, 188, 1024]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_14_pw2_weight_to_fp16_quantized, x = input_617_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_619_cast_fp16 = add(x = input_605_cast_fp16, y = linear_165_cast_fp16)[name = string("input_619_cast_fp16")];
            tensor<int32, [1]> input_621_axes_0 = const()[name = string("input_621_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377494400)))];
            tensor<fp16, [1024]> enc_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377496512)))];
            tensor<fp16, [1, 188, 1024]> input_621_cast_fp16 = layer_norm(axes = input_621_axes_0, beta = enc_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_4372_to_fp16, gamma = enc_layers_14_norm_feed_forward2_weight_to_fp16, x = input_619_cast_fp16)[name = string("input_621_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_14_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377498624))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381692992))))[name = string("blocks_14_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_166_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_14_ff2_1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_625_cast_fp16 = silu(x = linear_166_cast_fp16)[name = string("input_625_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_14_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381701248))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385895616))))[name = string("blocks_14_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_167_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_14_ff2_2_weight_to_fp16_quantized, x = input_625_cast_fp16)[name = string("linear_167_cast_fp16")];
            fp16 var_4474_to_fp16 = const()[name = string("op_4474_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4475_cast_fp16 = mul(x = linear_167_cast_fp16, y = var_4474_to_fp16)[name = string("op_4475_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_627_cast_fp16 = add(x = input_619_cast_fp16, y = var_4475_cast_fp16)[name = string("input_627_cast_fp16")];
            tensor<int32, [1]> input_629_axes_0 = const()[name = string("input_629_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_14_norm_out_weight_to_fp16 = const()[name = string("enc_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385897728)))];
            tensor<fp16, [1024]> enc_layers_14_norm_out_bias_to_fp16 = const()[name = string("enc_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385899840)))];
            tensor<fp16, [1, 188, 1024]> input_629_cast_fp16 = layer_norm(axes = input_629_axes_0, beta = enc_layers_14_norm_out_bias_to_fp16, epsilon = var_4372_to_fp16, gamma = enc_layers_14_norm_out_weight_to_fp16, x = input_627_cast_fp16)[name = string("input_629_cast_fp16")];
            int32 var_4488 = const()[name = string("op_4488"), val = int32(-1)];
            tensor<int32, [1]> input_631_axes_0 = const()[name = string("input_631_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385901952)))];
            tensor<fp16, [1024]> enc_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385904064)))];
            fp16 var_4496_to_fp16 = const()[name = string("op_4496_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_631_cast_fp16 = layer_norm(axes = input_631_axes_0, beta = enc_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_4496_to_fp16, gamma = enc_layers_15_norm_feed_forward1_weight_to_fp16, x = input_629_cast_fp16)[name = string("input_631_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_15_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385906176))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390100544))))[name = string("blocks_15_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_168_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_15_ff1_1_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_635_cast_fp16 = silu(x = linear_168_cast_fp16)[name = string("input_635_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_15_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390108800))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394303168))))[name = string("blocks_15_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_169_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_15_ff1_2_weight_to_fp16_quantized, x = input_635_cast_fp16)[name = string("linear_169_cast_fp16")];
            fp16 var_4516_to_fp16 = const()[name = string("op_4516_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4517_cast_fp16 = mul(x = linear_169_cast_fp16, y = var_4516_to_fp16)[name = string("op_4517_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_637_cast_fp16 = add(x = input_629_cast_fp16, y = var_4517_cast_fp16)[name = string("input_637_cast_fp16")];
            tensor<int32, [1]> input_639_axes_0 = const()[name = string("input_639_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394305280)))];
            tensor<fp16, [1024]> enc_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394307392)))];
            tensor<fp16, [1, 188, 1024]> input_639_cast_fp16 = layer_norm(axes = input_639_axes_0, beta = enc_layers_15_norm_self_att_bias_to_fp16, epsilon = var_4496_to_fp16, gamma = enc_layers_15_norm_self_att_weight_to_fp16, x = input_637_cast_fp16)[name = string("input_639_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_15_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394309504))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395358144))))[name = string("blocks_15_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_170_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_15_q_weight_to_fp16_quantized, x = input_639_cast_fp16)[name = string("linear_170_cast_fp16")];
            tensor<int32, [4]> var_4523 = const()[name = string("op_4523"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4524_cast_fp16 = reshape(shape = var_4523, x = linear_170_cast_fp16)[name = string("op_4524_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_15_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395360256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396408896))))[name = string("blocks_15_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_171_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_15_k_weight_to_fp16_quantized, x = input_639_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<int32, [4]> var_4528 = const()[name = string("op_4528"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4529_cast_fp16 = reshape(shape = var_4528, x = linear_171_cast_fp16)[name = string("op_4529_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_15_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396411008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397459648))))[name = string("blocks_15_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_172_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_15_v_weight_to_fp16_quantized, x = input_639_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<int32, [4]> var_4533 = const()[name = string("op_4533"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4534_cast_fp16 = reshape(shape = var_4533, x = linear_172_cast_fp16)[name = string("op_4534_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_116_to_fp16 = const()[name = string("const_116_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397461760)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_31_cast_fp16 = add(x = var_4524_cast_fp16, y = const_116_to_fp16)[name = string("q_u_31_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397463872)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_31_cast_fp16 = add(x = var_4524_cast_fp16, y = const_117_to_fp16)[name = string("q_v_31_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_15_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397465984))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398514624))))[name = string("blocks_15_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_173_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_15_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<int32, [4]> var_4544 = const()[name = string("op_4544"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_31_cast_fp16 = reshape(shape = var_4544, x = linear_173_cast_fp16)[name = string("p_31_cast_fp16")];
            bool input_641_transpose_x_0 = const()[name = string("input_641_transpose_x_0"), val = bool(false)];
            bool input_641_transpose_y_0 = const()[name = string("input_641_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_252_perm_0 = const()[name = string("transpose_252_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_253_perm_0 = const()[name = string("transpose_253_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_253 = transpose(perm = transpose_253_perm_0, x = p_31_cast_fp16)[name = string("transpose_359")];
            tensor<fp16, [1, 8, 188, 128]> transpose_252 = transpose(perm = transpose_252_perm_0, x = q_v_31_cast_fp16)[name = string("transpose_360")];
            tensor<fp16, [1, 8, 188, 375]> input_641_cast_fp16 = matmul(transpose_x = input_641_transpose_x_0, transpose_y = input_641_transpose_y_0, x = transpose_252, y = transpose_253)[name = string("input_641_cast_fp16")];
            tensor<int32, [8]> var_4550_pad_0 = const()[name = string("op_4550_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_4550_mode_0 = const()[name = string("op_4550_mode_0"), val = string("constant")];
            fp16 const_29_to_fp16 = const()[name = string("const_29_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_4550_cast_fp16 = pad(constant_val = const_29_to_fp16, mode = var_4550_mode_0, pad = var_4550_pad_0, x = input_641_cast_fp16)[name = string("op_4550_cast_fp16")];
            tensor<int32, [4]> var_4551 = const()[name = string("op_4551"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_61_cast_fp16 = reshape(shape = var_4551, x = var_4550_cast_fp16)[name = string("scores_61_cast_fp16")];
            tensor<int32, [4]> var_4555_begin_0 = const()[name = string("op_4555_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4555_end_0 = const()[name = string("op_4555_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_4555_end_mask_0 = const()[name = string("op_4555_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_4555_cast_fp16 = slice_by_index(begin = var_4555_begin_0, end = var_4555_end_0, end_mask = var_4555_end_mask_0, x = scores_61_cast_fp16)[name = string("op_4555_cast_fp16")];
            tensor<int32, [4]> var_4556 = const()[name = string("op_4556"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_4557_cast_fp16 = reshape(shape = var_4556, x = var_4555_cast_fp16)[name = string("op_4557_cast_fp16")];
            tensor<int32, [4]> bd_31_begin_0 = const()[name = string("bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_31_end_0 = const()[name = string("bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_31_end_mask_0 = const()[name = string("bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_31_cast_fp16 = slice_by_index(begin = bd_31_begin_0, end = bd_31_end_0, end_mask = bd_31_end_mask_0, x = var_4557_cast_fp16)[name = string("bd_31_cast_fp16")];
            bool var_4560_transpose_x_0 = const()[name = string("op_4560_transpose_x_0"), val = bool(false)];
            bool var_4560_transpose_y_0 = const()[name = string("op_4560_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_254_perm_0 = const()[name = string("transpose_254_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_255_perm_0 = const()[name = string("transpose_255_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_255 = transpose(perm = transpose_255_perm_0, x = var_4529_cast_fp16)[name = string("transpose_357")];
            tensor<fp16, [1, 8, 188, 128]> transpose_254 = transpose(perm = transpose_254_perm_0, x = q_u_31_cast_fp16)[name = string("transpose_358")];
            tensor<fp16, [1, 8, 188, 188]> var_4560_cast_fp16 = matmul(transpose_x = var_4560_transpose_x_0, transpose_y = var_4560_transpose_y_0, x = transpose_254, y = transpose_255)[name = string("op_4560_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4561_cast_fp16 = add(x = var_4560_cast_fp16, y = bd_31_cast_fp16)[name = string("op_4561_cast_fp16")];
            fp16 var_4562_to_fp16 = const()[name = string("op_4562_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_4563_cast_fp16 = mul(x = var_4561_cast_fp16, y = var_4562_to_fp16)[name = string("op_4563_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_63_cast_fp16 = add(x = var_4563_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_63_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4565_cast_fp16 = softmax(axis = var_4488, x = scores_63_cast_fp16)[name = string("op_4565_cast_fp16")];
            bool out_31_transpose_x_0 = const()[name = string("out_31_transpose_x_0"), val = bool(false)];
            bool out_31_transpose_y_0 = const()[name = string("out_31_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_4534_cast_fp16)[name = string("transpose_356")];
            tensor<fp16, [1, 8, 188, 128]> out_31_cast_fp16 = matmul(transpose_x = out_31_transpose_x_0, transpose_y = out_31_transpose_y_0, x = var_4565_cast_fp16, y = v_31_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<int32, [4]> var_4567_perm_0 = const()[name = string("op_4567_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4568 = const()[name = string("op_4568"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_4567_cast_fp16 = transpose(perm = var_4567_perm_0, x = out_31_cast_fp16)[name = string("transpose_355")];
            tensor<fp16, [1, 188, 1024]> input_643_cast_fp16 = reshape(shape = var_4568, x = var_4567_cast_fp16)[name = string("input_643_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_15_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398516736))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399565376))))[name = string("blocks_15_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_174_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_15_o_weight_to_fp16_quantized, x = input_643_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_645_cast_fp16 = add(x = input_637_cast_fp16, y = linear_174_cast_fp16)[name = string("input_645_cast_fp16")];
            tensor<int32, [1]> input_647_axes_0 = const()[name = string("input_647_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_15_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399567488)))];
            tensor<fp16, [1024]> enc_layers_15_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399569600)))];
            tensor<fp16, [1, 188, 1024]> input_647_cast_fp16 = layer_norm(axes = input_647_axes_0, beta = enc_layers_15_norm_conv_bias_to_fp16, epsilon = var_4496_to_fp16, gamma = enc_layers_15_norm_conv_weight_to_fp16, x = input_645_cast_fp16)[name = string("input_647_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_15_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399571712))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401668928))))[name = string("blocks_15_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_175_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_15_pw1_weight_to_fp16_quantized, x = input_647_cast_fp16)[name = string("linear_175_cast_fp16")];
            int32 var_4577_split_num_splits_0 = const()[name = string("op_4577_split_num_splits_0"), val = int32(2)];
            int32 var_4577_split_axis_0 = const()[name = string("op_4577_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_4577_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_4577_split_cast_fp16_1 = split(axis = var_4577_split_axis_0, num_splits = var_4577_split_num_splits_0, x = linear_175_cast_fp16)[name = string("op_4577_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4577_split_1_sigmoid_cast_fp16 = sigmoid(x = var_4577_split_cast_fp16_1)[name = string("op_4577_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4577_cast_fp16 = mul(x = var_4577_split_cast_fp16_0, y = var_4577_split_1_sigmoid_cast_fp16)[name = string("op_4577_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_61_cast_fp16 = mul(x = var_4577_cast_fp16, y = valid_cast_fp16)[name = string("h_61_cast_fp16")];
            tensor<int32, [3]> input_651_perm_0 = const()[name = string("input_651_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_653_pad_type_0 = const()[name = string("input_653_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_653_pad_0 = const()[name = string("input_653_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_653_groups_0 = const()[name = string("input_653_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_653_strides_0 = const()[name = string("input_653_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_653_dilations_0 = const()[name = string("input_653_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_68_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401673088))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401682368))))[name = string("const_68_to_fp16_quantized")];
            tensor<fp16, [1024]> const_69_to_fp16 = const()[name = string("const_69_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401684480)))];
            tensor<fp16, [1, 1024, 188]> input_651_cast_fp16 = transpose(perm = input_651_perm_0, x = h_61_cast_fp16)[name = string("transpose_354")];
            tensor<fp16, [1, 1024, 188]> input_655_cast_fp16 = conv(bias = const_69_to_fp16, dilations = input_653_dilations_0, groups = input_653_groups_0, pad = input_653_pad_0, pad_type = input_653_pad_type_0, strides = input_653_strides_0, weight = const_68_to_fp16_quantized, x = input_651_cast_fp16)[name = string("input_655_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_63_cast_fp16 = silu(x = input_655_cast_fp16)[name = string("h_63_cast_fp16")];
            tensor<int32, [3]> input_657_perm_0 = const()[name = string("input_657_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_15_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401686592))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402735232))))[name = string("blocks_15_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_657_cast_fp16 = transpose(perm = input_657_perm_0, x = h_63_cast_fp16)[name = string("transpose_353")];
            tensor<fp16, [1, 188, 1024]> linear_176_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_15_pw2_weight_to_fp16_quantized, x = input_657_cast_fp16)[name = string("linear_176_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_659_cast_fp16 = add(x = input_645_cast_fp16, y = linear_176_cast_fp16)[name = string("input_659_cast_fp16")];
            tensor<int32, [1]> input_661_axes_0 = const()[name = string("input_661_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402737344)))];
            tensor<fp16, [1024]> enc_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402739456)))];
            tensor<fp16, [1, 188, 1024]> input_661_cast_fp16 = layer_norm(axes = input_661_axes_0, beta = enc_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_4496_to_fp16, gamma = enc_layers_15_norm_feed_forward2_weight_to_fp16, x = input_659_cast_fp16)[name = string("input_661_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_15_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402741568))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406935936))))[name = string("blocks_15_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_177_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_15_ff2_1_weight_to_fp16_quantized, x = input_661_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_665_cast_fp16 = silu(x = linear_177_cast_fp16)[name = string("input_665_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_15_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406944192))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411138560))))[name = string("blocks_15_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_178_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_15_ff2_2_weight_to_fp16_quantized, x = input_665_cast_fp16)[name = string("linear_178_cast_fp16")];
            fp16 var_4598_to_fp16 = const()[name = string("op_4598_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4599_cast_fp16 = mul(x = linear_178_cast_fp16, y = var_4598_to_fp16)[name = string("op_4599_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_667_cast_fp16 = add(x = input_659_cast_fp16, y = var_4599_cast_fp16)[name = string("input_667_cast_fp16")];
            tensor<int32, [1]> input_669_axes_0 = const()[name = string("input_669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_15_norm_out_weight_to_fp16 = const()[name = string("enc_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411140672)))];
            tensor<fp16, [1024]> enc_layers_15_norm_out_bias_to_fp16 = const()[name = string("enc_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411142784)))];
            tensor<fp16, [1, 188, 1024]> input_669_cast_fp16 = layer_norm(axes = input_669_axes_0, beta = enc_layers_15_norm_out_bias_to_fp16, epsilon = var_4496_to_fp16, gamma = enc_layers_15_norm_out_weight_to_fp16, x = input_667_cast_fp16)[name = string("input_669_cast_fp16")];
            int32 var_4612 = const()[name = string("op_4612"), val = int32(-1)];
            tensor<int32, [1]> input_671_axes_0 = const()[name = string("input_671_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411144896)))];
            tensor<fp16, [1024]> enc_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411147008)))];
            fp16 var_4620_to_fp16 = const()[name = string("op_4620_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_671_cast_fp16 = layer_norm(axes = input_671_axes_0, beta = enc_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_4620_to_fp16, gamma = enc_layers_16_norm_feed_forward1_weight_to_fp16, x = input_669_cast_fp16)[name = string("input_671_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_16_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411149120))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415343488))))[name = string("blocks_16_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_179_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_16_ff1_1_weight_to_fp16_quantized, x = input_671_cast_fp16)[name = string("linear_179_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_675_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_675_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_16_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415351744))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419546112))))[name = string("blocks_16_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_180_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_16_ff1_2_weight_to_fp16_quantized, x = input_675_cast_fp16)[name = string("linear_180_cast_fp16")];
            fp16 var_4640_to_fp16 = const()[name = string("op_4640_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4641_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4640_to_fp16)[name = string("op_4641_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_677_cast_fp16 = add(x = input_669_cast_fp16, y = var_4641_cast_fp16)[name = string("input_677_cast_fp16")];
            tensor<int32, [1]> input_679_axes_0 = const()[name = string("input_679_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419548224)))];
            tensor<fp16, [1024]> enc_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419550336)))];
            tensor<fp16, [1, 188, 1024]> input_679_cast_fp16 = layer_norm(axes = input_679_axes_0, beta = enc_layers_16_norm_self_att_bias_to_fp16, epsilon = var_4620_to_fp16, gamma = enc_layers_16_norm_self_att_weight_to_fp16, x = input_677_cast_fp16)[name = string("input_679_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_16_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419552448))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420601088))))[name = string("blocks_16_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_181_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_16_q_weight_to_fp16_quantized, x = input_679_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<int32, [4]> var_4647 = const()[name = string("op_4647"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4648_cast_fp16 = reshape(shape = var_4647, x = linear_181_cast_fp16)[name = string("op_4648_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_16_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420603200))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421651840))))[name = string("blocks_16_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_182_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_16_k_weight_to_fp16_quantized, x = input_679_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_4652 = const()[name = string("op_4652"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4653_cast_fp16 = reshape(shape = var_4652, x = linear_182_cast_fp16)[name = string("op_4653_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_16_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421653952))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422702592))))[name = string("blocks_16_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_183_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_16_v_weight_to_fp16_quantized, x = input_679_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_4657 = const()[name = string("op_4657"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4658_cast_fp16 = reshape(shape = var_4657, x = linear_183_cast_fp16)[name = string("op_4658_cast_fp16")];
            tensor<int32, [4]> v_33_perm_0 = const()[name = string("v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422704704)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_33_cast_fp16 = add(x = var_4648_cast_fp16, y = const_118_to_fp16)[name = string("q_u_33_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_119_to_fp16 = const()[name = string("const_119_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422706816)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_33_cast_fp16 = add(x = var_4648_cast_fp16, y = const_119_to_fp16)[name = string("q_v_33_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_16_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422708928))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423757568))))[name = string("blocks_16_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_184_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_16_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<int32, [4]> var_4668 = const()[name = string("op_4668"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_33_cast_fp16 = reshape(shape = var_4668, x = linear_184_cast_fp16)[name = string("p_33_cast_fp16")];
            bool input_681_transpose_x_0 = const()[name = string("input_681_transpose_x_0"), val = bool(false)];
            bool input_681_transpose_y_0 = const()[name = string("input_681_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_256_perm_0 = const()[name = string("transpose_256_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_257_perm_0 = const()[name = string("transpose_257_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_257 = transpose(perm = transpose_257_perm_0, x = p_33_cast_fp16)[name = string("transpose_351")];
            tensor<fp16, [1, 8, 188, 128]> transpose_256 = transpose(perm = transpose_256_perm_0, x = q_v_33_cast_fp16)[name = string("transpose_352")];
            tensor<fp16, [1, 8, 188, 375]> input_681_cast_fp16 = matmul(transpose_x = input_681_transpose_x_0, transpose_y = input_681_transpose_y_0, x = transpose_256, y = transpose_257)[name = string("input_681_cast_fp16")];
            tensor<int32, [8]> var_4674_pad_0 = const()[name = string("op_4674_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_4674_mode_0 = const()[name = string("op_4674_mode_0"), val = string("constant")];
            fp16 const_30_to_fp16 = const()[name = string("const_30_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_4674_cast_fp16 = pad(constant_val = const_30_to_fp16, mode = var_4674_mode_0, pad = var_4674_pad_0, x = input_681_cast_fp16)[name = string("op_4674_cast_fp16")];
            tensor<int32, [4]> var_4675 = const()[name = string("op_4675"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_65_cast_fp16 = reshape(shape = var_4675, x = var_4674_cast_fp16)[name = string("scores_65_cast_fp16")];
            tensor<int32, [4]> var_4679_begin_0 = const()[name = string("op_4679_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4679_end_0 = const()[name = string("op_4679_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_4679_end_mask_0 = const()[name = string("op_4679_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_4679_cast_fp16 = slice_by_index(begin = var_4679_begin_0, end = var_4679_end_0, end_mask = var_4679_end_mask_0, x = scores_65_cast_fp16)[name = string("op_4679_cast_fp16")];
            tensor<int32, [4]> var_4680 = const()[name = string("op_4680"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_4681_cast_fp16 = reshape(shape = var_4680, x = var_4679_cast_fp16)[name = string("op_4681_cast_fp16")];
            tensor<int32, [4]> bd_33_begin_0 = const()[name = string("bd_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_33_end_0 = const()[name = string("bd_33_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_33_end_mask_0 = const()[name = string("bd_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_33_cast_fp16 = slice_by_index(begin = bd_33_begin_0, end = bd_33_end_0, end_mask = bd_33_end_mask_0, x = var_4681_cast_fp16)[name = string("bd_33_cast_fp16")];
            bool var_4684_transpose_x_0 = const()[name = string("op_4684_transpose_x_0"), val = bool(false)];
            bool var_4684_transpose_y_0 = const()[name = string("op_4684_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_258_perm_0 = const()[name = string("transpose_258_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_259_perm_0 = const()[name = string("transpose_259_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_259 = transpose(perm = transpose_259_perm_0, x = var_4653_cast_fp16)[name = string("transpose_349")];
            tensor<fp16, [1, 8, 188, 128]> transpose_258 = transpose(perm = transpose_258_perm_0, x = q_u_33_cast_fp16)[name = string("transpose_350")];
            tensor<fp16, [1, 8, 188, 188]> var_4684_cast_fp16 = matmul(transpose_x = var_4684_transpose_x_0, transpose_y = var_4684_transpose_y_0, x = transpose_258, y = transpose_259)[name = string("op_4684_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4685_cast_fp16 = add(x = var_4684_cast_fp16, y = bd_33_cast_fp16)[name = string("op_4685_cast_fp16")];
            fp16 var_4686_to_fp16 = const()[name = string("op_4686_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_4687_cast_fp16 = mul(x = var_4685_cast_fp16, y = var_4686_to_fp16)[name = string("op_4687_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_67_cast_fp16 = add(x = var_4687_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_67_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4689_cast_fp16 = softmax(axis = var_4612, x = scores_67_cast_fp16)[name = string("op_4689_cast_fp16")];
            bool out_33_transpose_x_0 = const()[name = string("out_33_transpose_x_0"), val = bool(false)];
            bool out_33_transpose_y_0 = const()[name = string("out_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_4658_cast_fp16)[name = string("transpose_348")];
            tensor<fp16, [1, 8, 188, 128]> out_33_cast_fp16 = matmul(transpose_x = out_33_transpose_x_0, transpose_y = out_33_transpose_y_0, x = var_4689_cast_fp16, y = v_33_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<int32, [4]> var_4691_perm_0 = const()[name = string("op_4691_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4692 = const()[name = string("op_4692"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_4691_cast_fp16 = transpose(perm = var_4691_perm_0, x = out_33_cast_fp16)[name = string("transpose_347")];
            tensor<fp16, [1, 188, 1024]> input_683_cast_fp16 = reshape(shape = var_4692, x = var_4691_cast_fp16)[name = string("input_683_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_16_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423759680))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424808320))))[name = string("blocks_16_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_185_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_16_o_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("linear_185_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_685_cast_fp16 = add(x = input_677_cast_fp16, y = linear_185_cast_fp16)[name = string("input_685_cast_fp16")];
            tensor<int32, [1]> input_687_axes_0 = const()[name = string("input_687_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_16_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424810432)))];
            tensor<fp16, [1024]> enc_layers_16_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424812544)))];
            tensor<fp16, [1, 188, 1024]> input_687_cast_fp16 = layer_norm(axes = input_687_axes_0, beta = enc_layers_16_norm_conv_bias_to_fp16, epsilon = var_4620_to_fp16, gamma = enc_layers_16_norm_conv_weight_to_fp16, x = input_685_cast_fp16)[name = string("input_687_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_16_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424814656))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426911872))))[name = string("blocks_16_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_186_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_16_pw1_weight_to_fp16_quantized, x = input_687_cast_fp16)[name = string("linear_186_cast_fp16")];
            int32 var_4701_split_num_splits_0 = const()[name = string("op_4701_split_num_splits_0"), val = int32(2)];
            int32 var_4701_split_axis_0 = const()[name = string("op_4701_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_4701_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_4701_split_cast_fp16_1 = split(axis = var_4701_split_axis_0, num_splits = var_4701_split_num_splits_0, x = linear_186_cast_fp16)[name = string("op_4701_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4701_split_1_sigmoid_cast_fp16 = sigmoid(x = var_4701_split_cast_fp16_1)[name = string("op_4701_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4701_cast_fp16 = mul(x = var_4701_split_cast_fp16_0, y = var_4701_split_1_sigmoid_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_65_cast_fp16 = mul(x = var_4701_cast_fp16, y = valid_cast_fp16)[name = string("h_65_cast_fp16")];
            tensor<int32, [3]> input_691_perm_0 = const()[name = string("input_691_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_693_pad_type_0 = const()[name = string("input_693_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_693_pad_0 = const()[name = string("input_693_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_693_groups_0 = const()[name = string("input_693_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_693_strides_0 = const()[name = string("input_693_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_693_dilations_0 = const()[name = string("input_693_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_70_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426916032))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426925312))))[name = string("const_70_to_fp16_quantized")];
            tensor<fp16, [1024]> const_71_to_fp16 = const()[name = string("const_71_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426927424)))];
            tensor<fp16, [1, 1024, 188]> input_691_cast_fp16 = transpose(perm = input_691_perm_0, x = h_65_cast_fp16)[name = string("transpose_346")];
            tensor<fp16, [1, 1024, 188]> input_695_cast_fp16 = conv(bias = const_71_to_fp16, dilations = input_693_dilations_0, groups = input_693_groups_0, pad = input_693_pad_0, pad_type = input_693_pad_type_0, strides = input_693_strides_0, weight = const_70_to_fp16_quantized, x = input_691_cast_fp16)[name = string("input_695_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_67_cast_fp16 = silu(x = input_695_cast_fp16)[name = string("h_67_cast_fp16")];
            tensor<int32, [3]> input_697_perm_0 = const()[name = string("input_697_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_16_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426929536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427978176))))[name = string("blocks_16_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_697_cast_fp16 = transpose(perm = input_697_perm_0, x = h_67_cast_fp16)[name = string("transpose_345")];
            tensor<fp16, [1, 188, 1024]> linear_187_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_16_pw2_weight_to_fp16_quantized, x = input_697_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_699_cast_fp16 = add(x = input_685_cast_fp16, y = linear_187_cast_fp16)[name = string("input_699_cast_fp16")];
            tensor<int32, [1]> input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427980288)))];
            tensor<fp16, [1024]> enc_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427982400)))];
            tensor<fp16, [1, 188, 1024]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = enc_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_4620_to_fp16, gamma = enc_layers_16_norm_feed_forward2_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_16_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427984512))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432178880))))[name = string("blocks_16_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_188_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_16_ff2_1_weight_to_fp16_quantized, x = input_701_cast_fp16)[name = string("linear_188_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_705_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_705_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_16_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432187136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436381504))))[name = string("blocks_16_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_189_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_16_ff2_2_weight_to_fp16_quantized, x = input_705_cast_fp16)[name = string("linear_189_cast_fp16")];
            fp16 var_4722_to_fp16 = const()[name = string("op_4722_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4723_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4722_to_fp16)[name = string("op_4723_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_707_cast_fp16 = add(x = input_699_cast_fp16, y = var_4723_cast_fp16)[name = string("input_707_cast_fp16")];
            tensor<int32, [1]> input_709_axes_0 = const()[name = string("input_709_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_16_norm_out_weight_to_fp16 = const()[name = string("enc_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436383616)))];
            tensor<fp16, [1024]> enc_layers_16_norm_out_bias_to_fp16 = const()[name = string("enc_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436385728)))];
            tensor<fp16, [1, 188, 1024]> input_709_cast_fp16 = layer_norm(axes = input_709_axes_0, beta = enc_layers_16_norm_out_bias_to_fp16, epsilon = var_4620_to_fp16, gamma = enc_layers_16_norm_out_weight_to_fp16, x = input_707_cast_fp16)[name = string("input_709_cast_fp16")];
            int32 var_4736 = const()[name = string("op_4736"), val = int32(-1)];
            tensor<int32, [1]> input_711_axes_0 = const()[name = string("input_711_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436387840)))];
            tensor<fp16, [1024]> enc_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436389952)))];
            fp16 var_4744_to_fp16 = const()[name = string("op_4744_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_711_cast_fp16 = layer_norm(axes = input_711_axes_0, beta = enc_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_4744_to_fp16, gamma = enc_layers_17_norm_feed_forward1_weight_to_fp16, x = input_709_cast_fp16)[name = string("input_711_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_17_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436392064))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440586432))))[name = string("blocks_17_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_190_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_17_ff1_1_weight_to_fp16_quantized, x = input_711_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_715_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_715_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_17_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440594688))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444789056))))[name = string("blocks_17_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_191_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_17_ff1_2_weight_to_fp16_quantized, x = input_715_cast_fp16)[name = string("linear_191_cast_fp16")];
            fp16 var_4764_to_fp16 = const()[name = string("op_4764_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4765_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4764_to_fp16)[name = string("op_4765_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_717_cast_fp16 = add(x = input_709_cast_fp16, y = var_4765_cast_fp16)[name = string("input_717_cast_fp16")];
            tensor<int32, [1]> input_719_axes_0 = const()[name = string("input_719_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444791168)))];
            tensor<fp16, [1024]> enc_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444793280)))];
            tensor<fp16, [1, 188, 1024]> input_719_cast_fp16 = layer_norm(axes = input_719_axes_0, beta = enc_layers_17_norm_self_att_bias_to_fp16, epsilon = var_4744_to_fp16, gamma = enc_layers_17_norm_self_att_weight_to_fp16, x = input_717_cast_fp16)[name = string("input_719_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_17_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444795392))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445844032))))[name = string("blocks_17_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_192_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_17_q_weight_to_fp16_quantized, x = input_719_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<int32, [4]> var_4771 = const()[name = string("op_4771"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4772_cast_fp16 = reshape(shape = var_4771, x = linear_192_cast_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_17_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445846144))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446894784))))[name = string("blocks_17_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_193_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_17_k_weight_to_fp16_quantized, x = input_719_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<int32, [4]> var_4776 = const()[name = string("op_4776"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4777_cast_fp16 = reshape(shape = var_4776, x = linear_193_cast_fp16)[name = string("op_4777_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_17_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446896896))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447945536))))[name = string("blocks_17_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_194_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_17_v_weight_to_fp16_quantized, x = input_719_cast_fp16)[name = string("linear_194_cast_fp16")];
            tensor<int32, [4]> var_4781 = const()[name = string("op_4781"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4782_cast_fp16 = reshape(shape = var_4781, x = linear_194_cast_fp16)[name = string("op_4782_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_120_to_fp16 = const()[name = string("const_120_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447947648)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_35_cast_fp16 = add(x = var_4772_cast_fp16, y = const_120_to_fp16)[name = string("q_u_35_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_121_to_fp16 = const()[name = string("const_121_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447949760)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_35_cast_fp16 = add(x = var_4772_cast_fp16, y = const_121_to_fp16)[name = string("q_v_35_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_17_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447951872))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449000512))))[name = string("blocks_17_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_195_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_17_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<int32, [4]> var_4792 = const()[name = string("op_4792"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_35_cast_fp16 = reshape(shape = var_4792, x = linear_195_cast_fp16)[name = string("p_35_cast_fp16")];
            bool input_721_transpose_x_0 = const()[name = string("input_721_transpose_x_0"), val = bool(false)];
            bool input_721_transpose_y_0 = const()[name = string("input_721_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_260_perm_0 = const()[name = string("transpose_260_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_261_perm_0 = const()[name = string("transpose_261_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_261 = transpose(perm = transpose_261_perm_0, x = p_35_cast_fp16)[name = string("transpose_343")];
            tensor<fp16, [1, 8, 188, 128]> transpose_260 = transpose(perm = transpose_260_perm_0, x = q_v_35_cast_fp16)[name = string("transpose_344")];
            tensor<fp16, [1, 8, 188, 375]> input_721_cast_fp16 = matmul(transpose_x = input_721_transpose_x_0, transpose_y = input_721_transpose_y_0, x = transpose_260, y = transpose_261)[name = string("input_721_cast_fp16")];
            tensor<int32, [8]> var_4798_pad_0 = const()[name = string("op_4798_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_4798_mode_0 = const()[name = string("op_4798_mode_0"), val = string("constant")];
            fp16 const_31_to_fp16 = const()[name = string("const_31_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_4798_cast_fp16 = pad(constant_val = const_31_to_fp16, mode = var_4798_mode_0, pad = var_4798_pad_0, x = input_721_cast_fp16)[name = string("op_4798_cast_fp16")];
            tensor<int32, [4]> var_4799 = const()[name = string("op_4799"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_69_cast_fp16 = reshape(shape = var_4799, x = var_4798_cast_fp16)[name = string("scores_69_cast_fp16")];
            tensor<int32, [4]> var_4803_begin_0 = const()[name = string("op_4803_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4803_end_0 = const()[name = string("op_4803_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_4803_end_mask_0 = const()[name = string("op_4803_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_4803_cast_fp16 = slice_by_index(begin = var_4803_begin_0, end = var_4803_end_0, end_mask = var_4803_end_mask_0, x = scores_69_cast_fp16)[name = string("op_4803_cast_fp16")];
            tensor<int32, [4]> var_4804 = const()[name = string("op_4804"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_4805_cast_fp16 = reshape(shape = var_4804, x = var_4803_cast_fp16)[name = string("op_4805_cast_fp16")];
            tensor<int32, [4]> bd_35_begin_0 = const()[name = string("bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_35_end_0 = const()[name = string("bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_35_end_mask_0 = const()[name = string("bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_35_cast_fp16 = slice_by_index(begin = bd_35_begin_0, end = bd_35_end_0, end_mask = bd_35_end_mask_0, x = var_4805_cast_fp16)[name = string("bd_35_cast_fp16")];
            bool var_4808_transpose_x_0 = const()[name = string("op_4808_transpose_x_0"), val = bool(false)];
            bool var_4808_transpose_y_0 = const()[name = string("op_4808_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_262_perm_0 = const()[name = string("transpose_262_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_263_perm_0 = const()[name = string("transpose_263_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_263 = transpose(perm = transpose_263_perm_0, x = var_4777_cast_fp16)[name = string("transpose_341")];
            tensor<fp16, [1, 8, 188, 128]> transpose_262 = transpose(perm = transpose_262_perm_0, x = q_u_35_cast_fp16)[name = string("transpose_342")];
            tensor<fp16, [1, 8, 188, 188]> var_4808_cast_fp16 = matmul(transpose_x = var_4808_transpose_x_0, transpose_y = var_4808_transpose_y_0, x = transpose_262, y = transpose_263)[name = string("op_4808_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4809_cast_fp16 = add(x = var_4808_cast_fp16, y = bd_35_cast_fp16)[name = string("op_4809_cast_fp16")];
            fp16 var_4810_to_fp16 = const()[name = string("op_4810_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_4811_cast_fp16 = mul(x = var_4809_cast_fp16, y = var_4810_to_fp16)[name = string("op_4811_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_71_cast_fp16 = add(x = var_4811_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_71_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4813_cast_fp16 = softmax(axis = var_4736, x = scores_71_cast_fp16)[name = string("op_4813_cast_fp16")];
            bool out_35_transpose_x_0 = const()[name = string("out_35_transpose_x_0"), val = bool(false)];
            bool out_35_transpose_y_0 = const()[name = string("out_35_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_4782_cast_fp16)[name = string("transpose_340")];
            tensor<fp16, [1, 8, 188, 128]> out_35_cast_fp16 = matmul(transpose_x = out_35_transpose_x_0, transpose_y = out_35_transpose_y_0, x = var_4813_cast_fp16, y = v_35_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<int32, [4]> var_4815_perm_0 = const()[name = string("op_4815_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4816 = const()[name = string("op_4816"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_4815_cast_fp16 = transpose(perm = var_4815_perm_0, x = out_35_cast_fp16)[name = string("transpose_339")];
            tensor<fp16, [1, 188, 1024]> input_723_cast_fp16 = reshape(shape = var_4816, x = var_4815_cast_fp16)[name = string("input_723_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_17_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449002624))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450051264))))[name = string("blocks_17_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_196_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_17_o_weight_to_fp16_quantized, x = input_723_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_725_cast_fp16 = add(x = input_717_cast_fp16, y = linear_196_cast_fp16)[name = string("input_725_cast_fp16")];
            tensor<int32, [1]> input_727_axes_0 = const()[name = string("input_727_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_17_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450053376)))];
            tensor<fp16, [1024]> enc_layers_17_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450055488)))];
            tensor<fp16, [1, 188, 1024]> input_727_cast_fp16 = layer_norm(axes = input_727_axes_0, beta = enc_layers_17_norm_conv_bias_to_fp16, epsilon = var_4744_to_fp16, gamma = enc_layers_17_norm_conv_weight_to_fp16, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_17_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450057600))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452154816))))[name = string("blocks_17_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_197_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_17_pw1_weight_to_fp16_quantized, x = input_727_cast_fp16)[name = string("linear_197_cast_fp16")];
            int32 var_4825_split_num_splits_0 = const()[name = string("op_4825_split_num_splits_0"), val = int32(2)];
            int32 var_4825_split_axis_0 = const()[name = string("op_4825_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_4825_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_4825_split_cast_fp16_1 = split(axis = var_4825_split_axis_0, num_splits = var_4825_split_num_splits_0, x = linear_197_cast_fp16)[name = string("op_4825_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4825_split_1_sigmoid_cast_fp16 = sigmoid(x = var_4825_split_cast_fp16_1)[name = string("op_4825_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4825_cast_fp16 = mul(x = var_4825_split_cast_fp16_0, y = var_4825_split_1_sigmoid_cast_fp16)[name = string("op_4825_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_69_cast_fp16 = mul(x = var_4825_cast_fp16, y = valid_cast_fp16)[name = string("h_69_cast_fp16")];
            tensor<int32, [3]> input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_733_pad_type_0 = const()[name = string("input_733_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_733_pad_0 = const()[name = string("input_733_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_733_groups_0 = const()[name = string("input_733_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_733_strides_0 = const()[name = string("input_733_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_733_dilations_0 = const()[name = string("input_733_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_72_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452158976))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452168256))))[name = string("const_72_to_fp16_quantized")];
            tensor<fp16, [1024]> const_73_to_fp16 = const()[name = string("const_73_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452170368)))];
            tensor<fp16, [1, 1024, 188]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = h_69_cast_fp16)[name = string("transpose_338")];
            tensor<fp16, [1, 1024, 188]> input_735_cast_fp16 = conv(bias = const_73_to_fp16, dilations = input_733_dilations_0, groups = input_733_groups_0, pad = input_733_pad_0, pad_type = input_733_pad_type_0, strides = input_733_strides_0, weight = const_72_to_fp16_quantized, x = input_731_cast_fp16)[name = string("input_735_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_71_cast_fp16 = silu(x = input_735_cast_fp16)[name = string("h_71_cast_fp16")];
            tensor<int32, [3]> input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_17_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452172480))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453221120))))[name = string("blocks_17_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = h_71_cast_fp16)[name = string("transpose_337")];
            tensor<fp16, [1, 188, 1024]> linear_198_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_17_pw2_weight_to_fp16_quantized, x = input_737_cast_fp16)[name = string("linear_198_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_739_cast_fp16 = add(x = input_725_cast_fp16, y = linear_198_cast_fp16)[name = string("input_739_cast_fp16")];
            tensor<int32, [1]> input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453223232)))];
            tensor<fp16, [1024]> enc_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453225344)))];
            tensor<fp16, [1, 188, 1024]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = enc_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_4744_to_fp16, gamma = enc_layers_17_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_17_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453227456))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457421824))))[name = string("blocks_17_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_199_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_17_ff2_1_weight_to_fp16_quantized, x = input_741_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_745_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_745_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_17_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457430080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461624448))))[name = string("blocks_17_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_200_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_17_ff2_2_weight_to_fp16_quantized, x = input_745_cast_fp16)[name = string("linear_200_cast_fp16")];
            fp16 var_4846_to_fp16 = const()[name = string("op_4846_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4847_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_4846_to_fp16)[name = string("op_4847_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_747_cast_fp16 = add(x = input_739_cast_fp16, y = var_4847_cast_fp16)[name = string("input_747_cast_fp16")];
            tensor<int32, [1]> input_749_axes_0 = const()[name = string("input_749_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_17_norm_out_weight_to_fp16 = const()[name = string("enc_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461626560)))];
            tensor<fp16, [1024]> enc_layers_17_norm_out_bias_to_fp16 = const()[name = string("enc_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461628672)))];
            tensor<fp16, [1, 188, 1024]> input_749_cast_fp16 = layer_norm(axes = input_749_axes_0, beta = enc_layers_17_norm_out_bias_to_fp16, epsilon = var_4744_to_fp16, gamma = enc_layers_17_norm_out_weight_to_fp16, x = input_747_cast_fp16)[name = string("input_749_cast_fp16")];
            int32 var_4860 = const()[name = string("op_4860"), val = int32(-1)];
            tensor<int32, [1]> input_751_axes_0 = const()[name = string("input_751_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461630784)))];
            tensor<fp16, [1024]> enc_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461632896)))];
            fp16 var_4868_to_fp16 = const()[name = string("op_4868_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_751_cast_fp16 = layer_norm(axes = input_751_axes_0, beta = enc_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_4868_to_fp16, gamma = enc_layers_18_norm_feed_forward1_weight_to_fp16, x = input_749_cast_fp16)[name = string("input_751_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_18_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461635008))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465829376))))[name = string("blocks_18_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_201_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_18_ff1_1_weight_to_fp16_quantized, x = input_751_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_755_cast_fp16 = silu(x = linear_201_cast_fp16)[name = string("input_755_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_18_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465837632))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470032000))))[name = string("blocks_18_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_202_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_18_ff1_2_weight_to_fp16_quantized, x = input_755_cast_fp16)[name = string("linear_202_cast_fp16")];
            fp16 var_4888_to_fp16 = const()[name = string("op_4888_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4889_cast_fp16 = mul(x = linear_202_cast_fp16, y = var_4888_to_fp16)[name = string("op_4889_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_757_cast_fp16 = add(x = input_749_cast_fp16, y = var_4889_cast_fp16)[name = string("input_757_cast_fp16")];
            tensor<int32, [1]> input_759_axes_0 = const()[name = string("input_759_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470034112)))];
            tensor<fp16, [1024]> enc_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470036224)))];
            tensor<fp16, [1, 188, 1024]> input_759_cast_fp16 = layer_norm(axes = input_759_axes_0, beta = enc_layers_18_norm_self_att_bias_to_fp16, epsilon = var_4868_to_fp16, gamma = enc_layers_18_norm_self_att_weight_to_fp16, x = input_757_cast_fp16)[name = string("input_759_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_18_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470038336))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471086976))))[name = string("blocks_18_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_203_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_18_q_weight_to_fp16_quantized, x = input_759_cast_fp16)[name = string("linear_203_cast_fp16")];
            tensor<int32, [4]> var_4895 = const()[name = string("op_4895"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4896_cast_fp16 = reshape(shape = var_4895, x = linear_203_cast_fp16)[name = string("op_4896_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_18_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471089088))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(472137728))))[name = string("blocks_18_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_204_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_18_k_weight_to_fp16_quantized, x = input_759_cast_fp16)[name = string("linear_204_cast_fp16")];
            tensor<int32, [4]> var_4900 = const()[name = string("op_4900"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4901_cast_fp16 = reshape(shape = var_4900, x = linear_204_cast_fp16)[name = string("op_4901_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_18_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(472139840))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473188480))))[name = string("blocks_18_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_205_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_18_v_weight_to_fp16_quantized, x = input_759_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<int32, [4]> var_4905 = const()[name = string("op_4905"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_4906_cast_fp16 = reshape(shape = var_4905, x = linear_205_cast_fp16)[name = string("op_4906_cast_fp16")];
            tensor<int32, [4]> v_37_perm_0 = const()[name = string("v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_122_to_fp16 = const()[name = string("const_122_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473190592)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_37_cast_fp16 = add(x = var_4896_cast_fp16, y = const_122_to_fp16)[name = string("q_u_37_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_123_to_fp16 = const()[name = string("const_123_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473192704)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_37_cast_fp16 = add(x = var_4896_cast_fp16, y = const_123_to_fp16)[name = string("q_v_37_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_18_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473194816))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474243456))))[name = string("blocks_18_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_206_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_18_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_206_cast_fp16")];
            tensor<int32, [4]> var_4916 = const()[name = string("op_4916"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_37_cast_fp16 = reshape(shape = var_4916, x = linear_206_cast_fp16)[name = string("p_37_cast_fp16")];
            bool input_761_transpose_x_0 = const()[name = string("input_761_transpose_x_0"), val = bool(false)];
            bool input_761_transpose_y_0 = const()[name = string("input_761_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_264_perm_0 = const()[name = string("transpose_264_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_265_perm_0 = const()[name = string("transpose_265_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_265 = transpose(perm = transpose_265_perm_0, x = p_37_cast_fp16)[name = string("transpose_335")];
            tensor<fp16, [1, 8, 188, 128]> transpose_264 = transpose(perm = transpose_264_perm_0, x = q_v_37_cast_fp16)[name = string("transpose_336")];
            tensor<fp16, [1, 8, 188, 375]> input_761_cast_fp16 = matmul(transpose_x = input_761_transpose_x_0, transpose_y = input_761_transpose_y_0, x = transpose_264, y = transpose_265)[name = string("input_761_cast_fp16")];
            tensor<int32, [8]> var_4922_pad_0 = const()[name = string("op_4922_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_4922_mode_0 = const()[name = string("op_4922_mode_0"), val = string("constant")];
            fp16 const_32_to_fp16 = const()[name = string("const_32_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_4922_cast_fp16 = pad(constant_val = const_32_to_fp16, mode = var_4922_mode_0, pad = var_4922_pad_0, x = input_761_cast_fp16)[name = string("op_4922_cast_fp16")];
            tensor<int32, [4]> var_4923 = const()[name = string("op_4923"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_73_cast_fp16 = reshape(shape = var_4923, x = var_4922_cast_fp16)[name = string("scores_73_cast_fp16")];
            tensor<int32, [4]> var_4927_begin_0 = const()[name = string("op_4927_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_4927_end_0 = const()[name = string("op_4927_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_4927_end_mask_0 = const()[name = string("op_4927_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_4927_cast_fp16 = slice_by_index(begin = var_4927_begin_0, end = var_4927_end_0, end_mask = var_4927_end_mask_0, x = scores_73_cast_fp16)[name = string("op_4927_cast_fp16")];
            tensor<int32, [4]> var_4928 = const()[name = string("op_4928"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_4929_cast_fp16 = reshape(shape = var_4928, x = var_4927_cast_fp16)[name = string("op_4929_cast_fp16")];
            tensor<int32, [4]> bd_37_begin_0 = const()[name = string("bd_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_37_end_0 = const()[name = string("bd_37_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_37_end_mask_0 = const()[name = string("bd_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_37_cast_fp16 = slice_by_index(begin = bd_37_begin_0, end = bd_37_end_0, end_mask = bd_37_end_mask_0, x = var_4929_cast_fp16)[name = string("bd_37_cast_fp16")];
            bool var_4932_transpose_x_0 = const()[name = string("op_4932_transpose_x_0"), val = bool(false)];
            bool var_4932_transpose_y_0 = const()[name = string("op_4932_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_266_perm_0 = const()[name = string("transpose_266_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_267_perm_0 = const()[name = string("transpose_267_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_267 = transpose(perm = transpose_267_perm_0, x = var_4901_cast_fp16)[name = string("transpose_333")];
            tensor<fp16, [1, 8, 188, 128]> transpose_266 = transpose(perm = transpose_266_perm_0, x = q_u_37_cast_fp16)[name = string("transpose_334")];
            tensor<fp16, [1, 8, 188, 188]> var_4932_cast_fp16 = matmul(transpose_x = var_4932_transpose_x_0, transpose_y = var_4932_transpose_y_0, x = transpose_266, y = transpose_267)[name = string("op_4932_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4933_cast_fp16 = add(x = var_4932_cast_fp16, y = bd_37_cast_fp16)[name = string("op_4933_cast_fp16")];
            fp16 var_4934_to_fp16 = const()[name = string("op_4934_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_4935_cast_fp16 = mul(x = var_4933_cast_fp16, y = var_4934_to_fp16)[name = string("op_4935_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_75_cast_fp16 = add(x = var_4935_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_75_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_4937_cast_fp16 = softmax(axis = var_4860, x = scores_75_cast_fp16)[name = string("op_4937_cast_fp16")];
            bool out_37_transpose_x_0 = const()[name = string("out_37_transpose_x_0"), val = bool(false)];
            bool out_37_transpose_y_0 = const()[name = string("out_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_4906_cast_fp16)[name = string("transpose_332")];
            tensor<fp16, [1, 8, 188, 128]> out_37_cast_fp16 = matmul(transpose_x = out_37_transpose_x_0, transpose_y = out_37_transpose_y_0, x = var_4937_cast_fp16, y = v_37_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<int32, [4]> var_4939_perm_0 = const()[name = string("op_4939_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4940 = const()[name = string("op_4940"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_4939_cast_fp16 = transpose(perm = var_4939_perm_0, x = out_37_cast_fp16)[name = string("transpose_331")];
            tensor<fp16, [1, 188, 1024]> input_763_cast_fp16 = reshape(shape = var_4940, x = var_4939_cast_fp16)[name = string("input_763_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_18_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474245568))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475294208))))[name = string("blocks_18_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_207_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_18_o_weight_to_fp16_quantized, x = input_763_cast_fp16)[name = string("linear_207_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_765_cast_fp16 = add(x = input_757_cast_fp16, y = linear_207_cast_fp16)[name = string("input_765_cast_fp16")];
            tensor<int32, [1]> input_767_axes_0 = const()[name = string("input_767_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_18_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475296320)))];
            tensor<fp16, [1024]> enc_layers_18_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475298432)))];
            tensor<fp16, [1, 188, 1024]> input_767_cast_fp16 = layer_norm(axes = input_767_axes_0, beta = enc_layers_18_norm_conv_bias_to_fp16, epsilon = var_4868_to_fp16, gamma = enc_layers_18_norm_conv_weight_to_fp16, x = input_765_cast_fp16)[name = string("input_767_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_18_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475300544))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477397760))))[name = string("blocks_18_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_208_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_18_pw1_weight_to_fp16_quantized, x = input_767_cast_fp16)[name = string("linear_208_cast_fp16")];
            int32 var_4949_split_num_splits_0 = const()[name = string("op_4949_split_num_splits_0"), val = int32(2)];
            int32 var_4949_split_axis_0 = const()[name = string("op_4949_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_4949_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_4949_split_cast_fp16_1 = split(axis = var_4949_split_axis_0, num_splits = var_4949_split_num_splits_0, x = linear_208_cast_fp16)[name = string("op_4949_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4949_split_1_sigmoid_cast_fp16 = sigmoid(x = var_4949_split_cast_fp16_1)[name = string("op_4949_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_4949_cast_fp16 = mul(x = var_4949_split_cast_fp16_0, y = var_4949_split_1_sigmoid_cast_fp16)[name = string("op_4949_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_73_cast_fp16 = mul(x = var_4949_cast_fp16, y = valid_cast_fp16)[name = string("h_73_cast_fp16")];
            tensor<int32, [3]> input_771_perm_0 = const()[name = string("input_771_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_773_pad_type_0 = const()[name = string("input_773_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_773_pad_0 = const()[name = string("input_773_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_773_groups_0 = const()[name = string("input_773_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_773_strides_0 = const()[name = string("input_773_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_773_dilations_0 = const()[name = string("input_773_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_74_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477401920))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477411200))))[name = string("const_74_to_fp16_quantized")];
            tensor<fp16, [1024]> const_75_to_fp16 = const()[name = string("const_75_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477413312)))];
            tensor<fp16, [1, 1024, 188]> input_771_cast_fp16 = transpose(perm = input_771_perm_0, x = h_73_cast_fp16)[name = string("transpose_330")];
            tensor<fp16, [1, 1024, 188]> input_775_cast_fp16 = conv(bias = const_75_to_fp16, dilations = input_773_dilations_0, groups = input_773_groups_0, pad = input_773_pad_0, pad_type = input_773_pad_type_0, strides = input_773_strides_0, weight = const_74_to_fp16_quantized, x = input_771_cast_fp16)[name = string("input_775_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_75_cast_fp16 = silu(x = input_775_cast_fp16)[name = string("h_75_cast_fp16")];
            tensor<int32, [3]> input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_18_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(477415424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478464064))))[name = string("blocks_18_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = h_75_cast_fp16)[name = string("transpose_329")];
            tensor<fp16, [1, 188, 1024]> linear_209_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_18_pw2_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("linear_209_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_779_cast_fp16 = add(x = input_765_cast_fp16, y = linear_209_cast_fp16)[name = string("input_779_cast_fp16")];
            tensor<int32, [1]> input_781_axes_0 = const()[name = string("input_781_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478466176)))];
            tensor<fp16, [1024]> enc_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478468288)))];
            tensor<fp16, [1, 188, 1024]> input_781_cast_fp16 = layer_norm(axes = input_781_axes_0, beta = enc_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_4868_to_fp16, gamma = enc_layers_18_norm_feed_forward2_weight_to_fp16, x = input_779_cast_fp16)[name = string("input_781_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_18_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478470400))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482664768))))[name = string("blocks_18_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_210_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_18_ff2_1_weight_to_fp16_quantized, x = input_781_cast_fp16)[name = string("linear_210_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_785_cast_fp16 = silu(x = linear_210_cast_fp16)[name = string("input_785_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_18_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482673024))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486867392))))[name = string("blocks_18_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_211_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_18_ff2_2_weight_to_fp16_quantized, x = input_785_cast_fp16)[name = string("linear_211_cast_fp16")];
            fp16 var_4970_to_fp16 = const()[name = string("op_4970_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_4971_cast_fp16 = mul(x = linear_211_cast_fp16, y = var_4970_to_fp16)[name = string("op_4971_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_787_cast_fp16 = add(x = input_779_cast_fp16, y = var_4971_cast_fp16)[name = string("input_787_cast_fp16")];
            tensor<int32, [1]> input_789_axes_0 = const()[name = string("input_789_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_18_norm_out_weight_to_fp16 = const()[name = string("enc_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486869504)))];
            tensor<fp16, [1024]> enc_layers_18_norm_out_bias_to_fp16 = const()[name = string("enc_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486871616)))];
            tensor<fp16, [1, 188, 1024]> input_789_cast_fp16 = layer_norm(axes = input_789_axes_0, beta = enc_layers_18_norm_out_bias_to_fp16, epsilon = var_4868_to_fp16, gamma = enc_layers_18_norm_out_weight_to_fp16, x = input_787_cast_fp16)[name = string("input_789_cast_fp16")];
            int32 var_4984 = const()[name = string("op_4984"), val = int32(-1)];
            tensor<int32, [1]> input_791_axes_0 = const()[name = string("input_791_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486873728)))];
            tensor<fp16, [1024]> enc_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486875840)))];
            fp16 var_4992_to_fp16 = const()[name = string("op_4992_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_791_cast_fp16 = layer_norm(axes = input_791_axes_0, beta = enc_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_4992_to_fp16, gamma = enc_layers_19_norm_feed_forward1_weight_to_fp16, x = input_789_cast_fp16)[name = string("input_791_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_19_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486877952))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491072320))))[name = string("blocks_19_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_212_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_19_ff1_1_weight_to_fp16_quantized, x = input_791_cast_fp16)[name = string("linear_212_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_795_cast_fp16 = silu(x = linear_212_cast_fp16)[name = string("input_795_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_19_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491080576))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495274944))))[name = string("blocks_19_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_213_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_19_ff1_2_weight_to_fp16_quantized, x = input_795_cast_fp16)[name = string("linear_213_cast_fp16")];
            fp16 var_5012_to_fp16 = const()[name = string("op_5012_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_5013_cast_fp16 = mul(x = linear_213_cast_fp16, y = var_5012_to_fp16)[name = string("op_5013_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_797_cast_fp16 = add(x = input_789_cast_fp16, y = var_5013_cast_fp16)[name = string("input_797_cast_fp16")];
            tensor<int32, [1]> input_799_axes_0 = const()[name = string("input_799_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495277056)))];
            tensor<fp16, [1024]> enc_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495279168)))];
            tensor<fp16, [1, 188, 1024]> input_799_cast_fp16 = layer_norm(axes = input_799_axes_0, beta = enc_layers_19_norm_self_att_bias_to_fp16, epsilon = var_4992_to_fp16, gamma = enc_layers_19_norm_self_att_weight_to_fp16, x = input_797_cast_fp16)[name = string("input_799_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_19_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495281280))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(496329920))))[name = string("blocks_19_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_214_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_19_q_weight_to_fp16_quantized, x = input_799_cast_fp16)[name = string("linear_214_cast_fp16")];
            tensor<int32, [4]> var_5019 = const()[name = string("op_5019"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5020_cast_fp16 = reshape(shape = var_5019, x = linear_214_cast_fp16)[name = string("op_5020_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_19_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(496332032))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497380672))))[name = string("blocks_19_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_215_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_19_k_weight_to_fp16_quantized, x = input_799_cast_fp16)[name = string("linear_215_cast_fp16")];
            tensor<int32, [4]> var_5024 = const()[name = string("op_5024"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5025_cast_fp16 = reshape(shape = var_5024, x = linear_215_cast_fp16)[name = string("op_5025_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_19_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497382784))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(498431424))))[name = string("blocks_19_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_216_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_19_v_weight_to_fp16_quantized, x = input_799_cast_fp16)[name = string("linear_216_cast_fp16")];
            tensor<int32, [4]> var_5029 = const()[name = string("op_5029"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5030_cast_fp16 = reshape(shape = var_5029, x = linear_216_cast_fp16)[name = string("op_5030_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_124_to_fp16 = const()[name = string("const_124_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(498433536)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_39_cast_fp16 = add(x = var_5020_cast_fp16, y = const_124_to_fp16)[name = string("q_u_39_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_125_to_fp16 = const()[name = string("const_125_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(498435648)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_39_cast_fp16 = add(x = var_5020_cast_fp16, y = const_125_to_fp16)[name = string("q_v_39_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_19_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(498437760))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(499486400))))[name = string("blocks_19_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_217_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_19_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_217_cast_fp16")];
            tensor<int32, [4]> var_5040 = const()[name = string("op_5040"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_39_cast_fp16 = reshape(shape = var_5040, x = linear_217_cast_fp16)[name = string("p_39_cast_fp16")];
            bool input_801_transpose_x_0 = const()[name = string("input_801_transpose_x_0"), val = bool(false)];
            bool input_801_transpose_y_0 = const()[name = string("input_801_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_268_perm_0 = const()[name = string("transpose_268_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_269_perm_0 = const()[name = string("transpose_269_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_269 = transpose(perm = transpose_269_perm_0, x = p_39_cast_fp16)[name = string("transpose_327")];
            tensor<fp16, [1, 8, 188, 128]> transpose_268 = transpose(perm = transpose_268_perm_0, x = q_v_39_cast_fp16)[name = string("transpose_328")];
            tensor<fp16, [1, 8, 188, 375]> input_801_cast_fp16 = matmul(transpose_x = input_801_transpose_x_0, transpose_y = input_801_transpose_y_0, x = transpose_268, y = transpose_269)[name = string("input_801_cast_fp16")];
            tensor<int32, [8]> var_5046_pad_0 = const()[name = string("op_5046_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_5046_mode_0 = const()[name = string("op_5046_mode_0"), val = string("constant")];
            fp16 const_33_to_fp16 = const()[name = string("const_33_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_5046_cast_fp16 = pad(constant_val = const_33_to_fp16, mode = var_5046_mode_0, pad = var_5046_pad_0, x = input_801_cast_fp16)[name = string("op_5046_cast_fp16")];
            tensor<int32, [4]> var_5047 = const()[name = string("op_5047"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_77_cast_fp16 = reshape(shape = var_5047, x = var_5046_cast_fp16)[name = string("scores_77_cast_fp16")];
            tensor<int32, [4]> var_5051_begin_0 = const()[name = string("op_5051_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5051_end_0 = const()[name = string("op_5051_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_5051_end_mask_0 = const()[name = string("op_5051_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_5051_cast_fp16 = slice_by_index(begin = var_5051_begin_0, end = var_5051_end_0, end_mask = var_5051_end_mask_0, x = scores_77_cast_fp16)[name = string("op_5051_cast_fp16")];
            tensor<int32, [4]> var_5052 = const()[name = string("op_5052"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_5053_cast_fp16 = reshape(shape = var_5052, x = var_5051_cast_fp16)[name = string("op_5053_cast_fp16")];
            tensor<int32, [4]> bd_39_begin_0 = const()[name = string("bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_39_end_0 = const()[name = string("bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_39_end_mask_0 = const()[name = string("bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_39_cast_fp16 = slice_by_index(begin = bd_39_begin_0, end = bd_39_end_0, end_mask = bd_39_end_mask_0, x = var_5053_cast_fp16)[name = string("bd_39_cast_fp16")];
            bool var_5056_transpose_x_0 = const()[name = string("op_5056_transpose_x_0"), val = bool(false)];
            bool var_5056_transpose_y_0 = const()[name = string("op_5056_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_270_perm_0 = const()[name = string("transpose_270_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_271_perm_0 = const()[name = string("transpose_271_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_271 = transpose(perm = transpose_271_perm_0, x = var_5025_cast_fp16)[name = string("transpose_325")];
            tensor<fp16, [1, 8, 188, 128]> transpose_270 = transpose(perm = transpose_270_perm_0, x = q_u_39_cast_fp16)[name = string("transpose_326")];
            tensor<fp16, [1, 8, 188, 188]> var_5056_cast_fp16 = matmul(transpose_x = var_5056_transpose_x_0, transpose_y = var_5056_transpose_y_0, x = transpose_270, y = transpose_271)[name = string("op_5056_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_5057_cast_fp16 = add(x = var_5056_cast_fp16, y = bd_39_cast_fp16)[name = string("op_5057_cast_fp16")];
            fp16 var_5058_to_fp16 = const()[name = string("op_5058_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_5059_cast_fp16 = mul(x = var_5057_cast_fp16, y = var_5058_to_fp16)[name = string("op_5059_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_79_cast_fp16 = add(x = var_5059_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_79_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_5061_cast_fp16 = softmax(axis = var_4984, x = scores_79_cast_fp16)[name = string("op_5061_cast_fp16")];
            bool out_39_transpose_x_0 = const()[name = string("out_39_transpose_x_0"), val = bool(false)];
            bool out_39_transpose_y_0 = const()[name = string("out_39_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_5030_cast_fp16)[name = string("transpose_324")];
            tensor<fp16, [1, 8, 188, 128]> out_39_cast_fp16 = matmul(transpose_x = out_39_transpose_x_0, transpose_y = out_39_transpose_y_0, x = var_5061_cast_fp16, y = v_39_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<int32, [4]> var_5063_perm_0 = const()[name = string("op_5063_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5064 = const()[name = string("op_5064"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_5063_cast_fp16 = transpose(perm = var_5063_perm_0, x = out_39_cast_fp16)[name = string("transpose_323")];
            tensor<fp16, [1, 188, 1024]> input_803_cast_fp16 = reshape(shape = var_5064, x = var_5063_cast_fp16)[name = string("input_803_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_19_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(499488512))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500537152))))[name = string("blocks_19_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_218_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_19_o_weight_to_fp16_quantized, x = input_803_cast_fp16)[name = string("linear_218_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_805_cast_fp16 = add(x = input_797_cast_fp16, y = linear_218_cast_fp16)[name = string("input_805_cast_fp16")];
            tensor<int32, [1]> input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_19_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500539264)))];
            tensor<fp16, [1024]> enc_layers_19_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500541376)))];
            tensor<fp16, [1, 188, 1024]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = enc_layers_19_norm_conv_bias_to_fp16, epsilon = var_4992_to_fp16, gamma = enc_layers_19_norm_conv_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_19_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500543488))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502640704))))[name = string("blocks_19_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_219_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_19_pw1_weight_to_fp16_quantized, x = input_807_cast_fp16)[name = string("linear_219_cast_fp16")];
            int32 var_5073_split_num_splits_0 = const()[name = string("op_5073_split_num_splits_0"), val = int32(2)];
            int32 var_5073_split_axis_0 = const()[name = string("op_5073_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_5073_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_5073_split_cast_fp16_1 = split(axis = var_5073_split_axis_0, num_splits = var_5073_split_num_splits_0, x = linear_219_cast_fp16)[name = string("op_5073_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_5073_split_1_sigmoid_cast_fp16 = sigmoid(x = var_5073_split_cast_fp16_1)[name = string("op_5073_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_5073_cast_fp16 = mul(x = var_5073_split_cast_fp16_0, y = var_5073_split_1_sigmoid_cast_fp16)[name = string("op_5073_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_77_cast_fp16 = mul(x = var_5073_cast_fp16, y = valid_cast_fp16)[name = string("h_77_cast_fp16")];
            tensor<int32, [3]> input_811_perm_0 = const()[name = string("input_811_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_813_pad_type_0 = const()[name = string("input_813_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_813_pad_0 = const()[name = string("input_813_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_813_groups_0 = const()[name = string("input_813_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_813_strides_0 = const()[name = string("input_813_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_813_dilations_0 = const()[name = string("input_813_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_76_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502644864))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502654144))))[name = string("const_76_to_fp16_quantized")];
            tensor<fp16, [1024]> const_77_to_fp16 = const()[name = string("const_77_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502656256)))];
            tensor<fp16, [1, 1024, 188]> input_811_cast_fp16 = transpose(perm = input_811_perm_0, x = h_77_cast_fp16)[name = string("transpose_322")];
            tensor<fp16, [1, 1024, 188]> input_815_cast_fp16 = conv(bias = const_77_to_fp16, dilations = input_813_dilations_0, groups = input_813_groups_0, pad = input_813_pad_0, pad_type = input_813_pad_type_0, strides = input_813_strides_0, weight = const_76_to_fp16_quantized, x = input_811_cast_fp16)[name = string("input_815_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_79_cast_fp16 = silu(x = input_815_cast_fp16)[name = string("h_79_cast_fp16")];
            tensor<int32, [3]> input_817_perm_0 = const()[name = string("input_817_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_19_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502658368))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503707008))))[name = string("blocks_19_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_817_cast_fp16 = transpose(perm = input_817_perm_0, x = h_79_cast_fp16)[name = string("transpose_321")];
            tensor<fp16, [1, 188, 1024]> linear_220_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_19_pw2_weight_to_fp16_quantized, x = input_817_cast_fp16)[name = string("linear_220_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_819_cast_fp16 = add(x = input_805_cast_fp16, y = linear_220_cast_fp16)[name = string("input_819_cast_fp16")];
            tensor<int32, [1]> input_821_axes_0 = const()[name = string("input_821_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503709120)))];
            tensor<fp16, [1024]> enc_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503711232)))];
            tensor<fp16, [1, 188, 1024]> input_821_cast_fp16 = layer_norm(axes = input_821_axes_0, beta = enc_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_4992_to_fp16, gamma = enc_layers_19_norm_feed_forward2_weight_to_fp16, x = input_819_cast_fp16)[name = string("input_821_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_19_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503713344))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(507907712))))[name = string("blocks_19_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_221_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_19_ff2_1_weight_to_fp16_quantized, x = input_821_cast_fp16)[name = string("linear_221_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_825_cast_fp16 = silu(x = linear_221_cast_fp16)[name = string("input_825_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_19_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(507915968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512110336))))[name = string("blocks_19_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_222_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_19_ff2_2_weight_to_fp16_quantized, x = input_825_cast_fp16)[name = string("linear_222_cast_fp16")];
            fp16 var_5094_to_fp16 = const()[name = string("op_5094_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_5095_cast_fp16 = mul(x = linear_222_cast_fp16, y = var_5094_to_fp16)[name = string("op_5095_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_827_cast_fp16 = add(x = input_819_cast_fp16, y = var_5095_cast_fp16)[name = string("input_827_cast_fp16")];
            tensor<int32, [1]> input_829_axes_0 = const()[name = string("input_829_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_19_norm_out_weight_to_fp16 = const()[name = string("enc_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512112448)))];
            tensor<fp16, [1024]> enc_layers_19_norm_out_bias_to_fp16 = const()[name = string("enc_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512114560)))];
            tensor<fp16, [1, 188, 1024]> input_829_cast_fp16 = layer_norm(axes = input_829_axes_0, beta = enc_layers_19_norm_out_bias_to_fp16, epsilon = var_4992_to_fp16, gamma = enc_layers_19_norm_out_weight_to_fp16, x = input_827_cast_fp16)[name = string("input_829_cast_fp16")];
            int32 var_5108 = const()[name = string("op_5108"), val = int32(-1)];
            tensor<int32, [1]> input_831_axes_0 = const()[name = string("input_831_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512116672)))];
            tensor<fp16, [1024]> enc_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512118784)))];
            fp16 var_5116_to_fp16 = const()[name = string("op_5116_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_831_cast_fp16 = layer_norm(axes = input_831_axes_0, beta = enc_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_5116_to_fp16, gamma = enc_layers_20_norm_feed_forward1_weight_to_fp16, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_20_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512120896))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516315264))))[name = string("blocks_20_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_223_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_20_ff1_1_weight_to_fp16_quantized, x = input_831_cast_fp16)[name = string("linear_223_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_835_cast_fp16 = silu(x = linear_223_cast_fp16)[name = string("input_835_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_20_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516323520))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520517888))))[name = string("blocks_20_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_224_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_20_ff1_2_weight_to_fp16_quantized, x = input_835_cast_fp16)[name = string("linear_224_cast_fp16")];
            fp16 var_5136_to_fp16 = const()[name = string("op_5136_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_5137_cast_fp16 = mul(x = linear_224_cast_fp16, y = var_5136_to_fp16)[name = string("op_5137_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_837_cast_fp16 = add(x = input_829_cast_fp16, y = var_5137_cast_fp16)[name = string("input_837_cast_fp16")];
            tensor<int32, [1]> input_839_axes_0 = const()[name = string("input_839_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520520000)))];
            tensor<fp16, [1024]> enc_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520522112)))];
            tensor<fp16, [1, 188, 1024]> input_839_cast_fp16 = layer_norm(axes = input_839_axes_0, beta = enc_layers_20_norm_self_att_bias_to_fp16, epsilon = var_5116_to_fp16, gamma = enc_layers_20_norm_self_att_weight_to_fp16, x = input_837_cast_fp16)[name = string("input_839_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_20_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520524224))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(521572864))))[name = string("blocks_20_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_225_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_20_q_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("linear_225_cast_fp16")];
            tensor<int32, [4]> var_5143 = const()[name = string("op_5143"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5144_cast_fp16 = reshape(shape = var_5143, x = linear_225_cast_fp16)[name = string("op_5144_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_20_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(521574976))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522623616))))[name = string("blocks_20_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_226_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_20_k_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("linear_226_cast_fp16")];
            tensor<int32, [4]> var_5148 = const()[name = string("op_5148"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5149_cast_fp16 = reshape(shape = var_5148, x = linear_226_cast_fp16)[name = string("op_5149_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_20_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522625728))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(523674368))))[name = string("blocks_20_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_227_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_20_v_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("linear_227_cast_fp16")];
            tensor<int32, [4]> var_5153 = const()[name = string("op_5153"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5154_cast_fp16 = reshape(shape = var_5153, x = linear_227_cast_fp16)[name = string("op_5154_cast_fp16")];
            tensor<int32, [4]> v_41_perm_0 = const()[name = string("v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_126_to_fp16 = const()[name = string("const_126_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(523676480)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_41_cast_fp16 = add(x = var_5144_cast_fp16, y = const_126_to_fp16)[name = string("q_u_41_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_127_to_fp16 = const()[name = string("const_127_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(523678592)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_41_cast_fp16 = add(x = var_5144_cast_fp16, y = const_127_to_fp16)[name = string("q_v_41_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_20_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(523680704))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524729344))))[name = string("blocks_20_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_228_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_20_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_228_cast_fp16")];
            tensor<int32, [4]> var_5164 = const()[name = string("op_5164"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_41_cast_fp16 = reshape(shape = var_5164, x = linear_228_cast_fp16)[name = string("p_41_cast_fp16")];
            bool input_841_transpose_x_0 = const()[name = string("input_841_transpose_x_0"), val = bool(false)];
            bool input_841_transpose_y_0 = const()[name = string("input_841_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_272_perm_0 = const()[name = string("transpose_272_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_273_perm_0 = const()[name = string("transpose_273_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_273 = transpose(perm = transpose_273_perm_0, x = p_41_cast_fp16)[name = string("transpose_319")];
            tensor<fp16, [1, 8, 188, 128]> transpose_272 = transpose(perm = transpose_272_perm_0, x = q_v_41_cast_fp16)[name = string("transpose_320")];
            tensor<fp16, [1, 8, 188, 375]> input_841_cast_fp16 = matmul(transpose_x = input_841_transpose_x_0, transpose_y = input_841_transpose_y_0, x = transpose_272, y = transpose_273)[name = string("input_841_cast_fp16")];
            tensor<int32, [8]> var_5170_pad_0 = const()[name = string("op_5170_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_5170_mode_0 = const()[name = string("op_5170_mode_0"), val = string("constant")];
            fp16 const_34_to_fp16 = const()[name = string("const_34_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_5170_cast_fp16 = pad(constant_val = const_34_to_fp16, mode = var_5170_mode_0, pad = var_5170_pad_0, x = input_841_cast_fp16)[name = string("op_5170_cast_fp16")];
            tensor<int32, [4]> var_5171 = const()[name = string("op_5171"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_81_cast_fp16 = reshape(shape = var_5171, x = var_5170_cast_fp16)[name = string("scores_81_cast_fp16")];
            tensor<int32, [4]> var_5175_begin_0 = const()[name = string("op_5175_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5175_end_0 = const()[name = string("op_5175_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_5175_end_mask_0 = const()[name = string("op_5175_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_5175_cast_fp16 = slice_by_index(begin = var_5175_begin_0, end = var_5175_end_0, end_mask = var_5175_end_mask_0, x = scores_81_cast_fp16)[name = string("op_5175_cast_fp16")];
            tensor<int32, [4]> var_5176 = const()[name = string("op_5176"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_5177_cast_fp16 = reshape(shape = var_5176, x = var_5175_cast_fp16)[name = string("op_5177_cast_fp16")];
            tensor<int32, [4]> bd_41_begin_0 = const()[name = string("bd_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_41_end_0 = const()[name = string("bd_41_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_41_end_mask_0 = const()[name = string("bd_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_41_cast_fp16 = slice_by_index(begin = bd_41_begin_0, end = bd_41_end_0, end_mask = bd_41_end_mask_0, x = var_5177_cast_fp16)[name = string("bd_41_cast_fp16")];
            bool var_5180_transpose_x_0 = const()[name = string("op_5180_transpose_x_0"), val = bool(false)];
            bool var_5180_transpose_y_0 = const()[name = string("op_5180_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_274_perm_0 = const()[name = string("transpose_274_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_275_perm_0 = const()[name = string("transpose_275_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_275 = transpose(perm = transpose_275_perm_0, x = var_5149_cast_fp16)[name = string("transpose_317")];
            tensor<fp16, [1, 8, 188, 128]> transpose_274 = transpose(perm = transpose_274_perm_0, x = q_u_41_cast_fp16)[name = string("transpose_318")];
            tensor<fp16, [1, 8, 188, 188]> var_5180_cast_fp16 = matmul(transpose_x = var_5180_transpose_x_0, transpose_y = var_5180_transpose_y_0, x = transpose_274, y = transpose_275)[name = string("op_5180_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_5181_cast_fp16 = add(x = var_5180_cast_fp16, y = bd_41_cast_fp16)[name = string("op_5181_cast_fp16")];
            fp16 var_5182_to_fp16 = const()[name = string("op_5182_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_5183_cast_fp16 = mul(x = var_5181_cast_fp16, y = var_5182_to_fp16)[name = string("op_5183_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_83_cast_fp16 = add(x = var_5183_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_83_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_5185_cast_fp16 = softmax(axis = var_5108, x = scores_83_cast_fp16)[name = string("op_5185_cast_fp16")];
            bool out_41_transpose_x_0 = const()[name = string("out_41_transpose_x_0"), val = bool(false)];
            bool out_41_transpose_y_0 = const()[name = string("out_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = var_5154_cast_fp16)[name = string("transpose_316")];
            tensor<fp16, [1, 8, 188, 128]> out_41_cast_fp16 = matmul(transpose_x = out_41_transpose_x_0, transpose_y = out_41_transpose_y_0, x = var_5185_cast_fp16, y = v_41_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<int32, [4]> var_5187_perm_0 = const()[name = string("op_5187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5188 = const()[name = string("op_5188"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_5187_cast_fp16 = transpose(perm = var_5187_perm_0, x = out_41_cast_fp16)[name = string("transpose_315")];
            tensor<fp16, [1, 188, 1024]> input_843_cast_fp16 = reshape(shape = var_5188, x = var_5187_cast_fp16)[name = string("input_843_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_20_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(524731456))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525780096))))[name = string("blocks_20_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_229_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_20_o_weight_to_fp16_quantized, x = input_843_cast_fp16)[name = string("linear_229_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_845_cast_fp16 = add(x = input_837_cast_fp16, y = linear_229_cast_fp16)[name = string("input_845_cast_fp16")];
            tensor<int32, [1]> input_847_axes_0 = const()[name = string("input_847_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_20_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525782208)))];
            tensor<fp16, [1024]> enc_layers_20_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525784320)))];
            tensor<fp16, [1, 188, 1024]> input_847_cast_fp16 = layer_norm(axes = input_847_axes_0, beta = enc_layers_20_norm_conv_bias_to_fp16, epsilon = var_5116_to_fp16, gamma = enc_layers_20_norm_conv_weight_to_fp16, x = input_845_cast_fp16)[name = string("input_847_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_20_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525786432))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527883648))))[name = string("blocks_20_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_230_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_20_pw1_weight_to_fp16_quantized, x = input_847_cast_fp16)[name = string("linear_230_cast_fp16")];
            int32 var_5197_split_num_splits_0 = const()[name = string("op_5197_split_num_splits_0"), val = int32(2)];
            int32 var_5197_split_axis_0 = const()[name = string("op_5197_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_5197_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_5197_split_cast_fp16_1 = split(axis = var_5197_split_axis_0, num_splits = var_5197_split_num_splits_0, x = linear_230_cast_fp16)[name = string("op_5197_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_5197_split_1_sigmoid_cast_fp16 = sigmoid(x = var_5197_split_cast_fp16_1)[name = string("op_5197_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_5197_cast_fp16 = mul(x = var_5197_split_cast_fp16_0, y = var_5197_split_1_sigmoid_cast_fp16)[name = string("op_5197_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_81_cast_fp16 = mul(x = var_5197_cast_fp16, y = valid_cast_fp16)[name = string("h_81_cast_fp16")];
            tensor<int32, [3]> input_851_perm_0 = const()[name = string("input_851_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_853_pad_type_0 = const()[name = string("input_853_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_853_pad_0 = const()[name = string("input_853_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_853_groups_0 = const()[name = string("input_853_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_853_strides_0 = const()[name = string("input_853_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_853_dilations_0 = const()[name = string("input_853_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_78_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527887808))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527897088))))[name = string("const_78_to_fp16_quantized")];
            tensor<fp16, [1024]> const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527899200)))];
            tensor<fp16, [1, 1024, 188]> input_851_cast_fp16 = transpose(perm = input_851_perm_0, x = h_81_cast_fp16)[name = string("transpose_314")];
            tensor<fp16, [1, 1024, 188]> input_855_cast_fp16 = conv(bias = const_79_to_fp16, dilations = input_853_dilations_0, groups = input_853_groups_0, pad = input_853_pad_0, pad_type = input_853_pad_type_0, strides = input_853_strides_0, weight = const_78_to_fp16_quantized, x = input_851_cast_fp16)[name = string("input_855_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_83_cast_fp16 = silu(x = input_855_cast_fp16)[name = string("h_83_cast_fp16")];
            tensor<int32, [3]> input_857_perm_0 = const()[name = string("input_857_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_20_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527901312))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528949952))))[name = string("blocks_20_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_857_cast_fp16 = transpose(perm = input_857_perm_0, x = h_83_cast_fp16)[name = string("transpose_313")];
            tensor<fp16, [1, 188, 1024]> linear_231_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_20_pw2_weight_to_fp16_quantized, x = input_857_cast_fp16)[name = string("linear_231_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_859_cast_fp16 = add(x = input_845_cast_fp16, y = linear_231_cast_fp16)[name = string("input_859_cast_fp16")];
            tensor<int32, [1]> input_861_axes_0 = const()[name = string("input_861_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528952064)))];
            tensor<fp16, [1024]> enc_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528954176)))];
            tensor<fp16, [1, 188, 1024]> input_861_cast_fp16 = layer_norm(axes = input_861_axes_0, beta = enc_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_5116_to_fp16, gamma = enc_layers_20_norm_feed_forward2_weight_to_fp16, x = input_859_cast_fp16)[name = string("input_861_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_20_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528956288))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533150656))))[name = string("blocks_20_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_232_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_20_ff2_1_weight_to_fp16_quantized, x = input_861_cast_fp16)[name = string("linear_232_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_865_cast_fp16 = silu(x = linear_232_cast_fp16)[name = string("input_865_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_20_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533158912))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537353280))))[name = string("blocks_20_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_233_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_20_ff2_2_weight_to_fp16_quantized, x = input_865_cast_fp16)[name = string("linear_233_cast_fp16")];
            fp16 var_5218_to_fp16 = const()[name = string("op_5218_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_5219_cast_fp16 = mul(x = linear_233_cast_fp16, y = var_5218_to_fp16)[name = string("op_5219_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_867_cast_fp16 = add(x = input_859_cast_fp16, y = var_5219_cast_fp16)[name = string("input_867_cast_fp16")];
            tensor<int32, [1]> input_869_axes_0 = const()[name = string("input_869_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_20_norm_out_weight_to_fp16 = const()[name = string("enc_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537355392)))];
            tensor<fp16, [1024]> enc_layers_20_norm_out_bias_to_fp16 = const()[name = string("enc_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537357504)))];
            tensor<fp16, [1, 188, 1024]> input_869_cast_fp16 = layer_norm(axes = input_869_axes_0, beta = enc_layers_20_norm_out_bias_to_fp16, epsilon = var_5116_to_fp16, gamma = enc_layers_20_norm_out_weight_to_fp16, x = input_867_cast_fp16)[name = string("input_869_cast_fp16")];
            int32 var_5232 = const()[name = string("op_5232"), val = int32(-1)];
            tensor<int32, [1]> input_871_axes_0 = const()[name = string("input_871_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537359616)))];
            tensor<fp16, [1024]> enc_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537361728)))];
            fp16 var_5240_to_fp16 = const()[name = string("op_5240_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_871_cast_fp16 = layer_norm(axes = input_871_axes_0, beta = enc_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_5240_to_fp16, gamma = enc_layers_21_norm_feed_forward1_weight_to_fp16, x = input_869_cast_fp16)[name = string("input_871_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_21_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537363840))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541558208))))[name = string("blocks_21_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_234_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_21_ff1_1_weight_to_fp16_quantized, x = input_871_cast_fp16)[name = string("linear_234_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_875_cast_fp16 = silu(x = linear_234_cast_fp16)[name = string("input_875_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_21_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541566464))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545760832))))[name = string("blocks_21_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_235_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_21_ff1_2_weight_to_fp16_quantized, x = input_875_cast_fp16)[name = string("linear_235_cast_fp16")];
            fp16 var_5260_to_fp16 = const()[name = string("op_5260_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_5261_cast_fp16 = mul(x = linear_235_cast_fp16, y = var_5260_to_fp16)[name = string("op_5261_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_877_cast_fp16 = add(x = input_869_cast_fp16, y = var_5261_cast_fp16)[name = string("input_877_cast_fp16")];
            tensor<int32, [1]> input_879_axes_0 = const()[name = string("input_879_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545762944)))];
            tensor<fp16, [1024]> enc_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545765056)))];
            tensor<fp16, [1, 188, 1024]> input_879_cast_fp16 = layer_norm(axes = input_879_axes_0, beta = enc_layers_21_norm_self_att_bias_to_fp16, epsilon = var_5240_to_fp16, gamma = enc_layers_21_norm_self_att_weight_to_fp16, x = input_877_cast_fp16)[name = string("input_879_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_21_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545767168))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546815808))))[name = string("blocks_21_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_236_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_21_q_weight_to_fp16_quantized, x = input_879_cast_fp16)[name = string("linear_236_cast_fp16")];
            tensor<int32, [4]> var_5267 = const()[name = string("op_5267"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5268_cast_fp16 = reshape(shape = var_5267, x = linear_236_cast_fp16)[name = string("op_5268_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_21_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546817920))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547866560))))[name = string("blocks_21_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_237_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_21_k_weight_to_fp16_quantized, x = input_879_cast_fp16)[name = string("linear_237_cast_fp16")];
            tensor<int32, [4]> var_5272 = const()[name = string("op_5272"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5273_cast_fp16 = reshape(shape = var_5272, x = linear_237_cast_fp16)[name = string("op_5273_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_21_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547868672))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548917312))))[name = string("blocks_21_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_238_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_21_v_weight_to_fp16_quantized, x = input_879_cast_fp16)[name = string("linear_238_cast_fp16")];
            tensor<int32, [4]> var_5277 = const()[name = string("op_5277"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5278_cast_fp16 = reshape(shape = var_5277, x = linear_238_cast_fp16)[name = string("op_5278_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_128_to_fp16 = const()[name = string("const_128_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548919424)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_43_cast_fp16 = add(x = var_5268_cast_fp16, y = const_128_to_fp16)[name = string("q_u_43_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_129_to_fp16 = const()[name = string("const_129_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548921536)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_43_cast_fp16 = add(x = var_5268_cast_fp16, y = const_129_to_fp16)[name = string("q_v_43_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_21_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(548923648))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549972288))))[name = string("blocks_21_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_239_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_21_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_239_cast_fp16")];
            tensor<int32, [4]> var_5288 = const()[name = string("op_5288"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_43_cast_fp16 = reshape(shape = var_5288, x = linear_239_cast_fp16)[name = string("p_43_cast_fp16")];
            bool input_881_transpose_x_0 = const()[name = string("input_881_transpose_x_0"), val = bool(false)];
            bool input_881_transpose_y_0 = const()[name = string("input_881_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_276_perm_0 = const()[name = string("transpose_276_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_277_perm_0 = const()[name = string("transpose_277_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_277 = transpose(perm = transpose_277_perm_0, x = p_43_cast_fp16)[name = string("transpose_311")];
            tensor<fp16, [1, 8, 188, 128]> transpose_276 = transpose(perm = transpose_276_perm_0, x = q_v_43_cast_fp16)[name = string("transpose_312")];
            tensor<fp16, [1, 8, 188, 375]> input_881_cast_fp16 = matmul(transpose_x = input_881_transpose_x_0, transpose_y = input_881_transpose_y_0, x = transpose_276, y = transpose_277)[name = string("input_881_cast_fp16")];
            tensor<int32, [8]> var_5294_pad_0 = const()[name = string("op_5294_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_5294_mode_0 = const()[name = string("op_5294_mode_0"), val = string("constant")];
            fp16 const_35_to_fp16 = const()[name = string("const_35_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_5294_cast_fp16 = pad(constant_val = const_35_to_fp16, mode = var_5294_mode_0, pad = var_5294_pad_0, x = input_881_cast_fp16)[name = string("op_5294_cast_fp16")];
            tensor<int32, [4]> var_5295 = const()[name = string("op_5295"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_85_cast_fp16 = reshape(shape = var_5295, x = var_5294_cast_fp16)[name = string("scores_85_cast_fp16")];
            tensor<int32, [4]> var_5299_begin_0 = const()[name = string("op_5299_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5299_end_0 = const()[name = string("op_5299_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_5299_end_mask_0 = const()[name = string("op_5299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_5299_cast_fp16 = slice_by_index(begin = var_5299_begin_0, end = var_5299_end_0, end_mask = var_5299_end_mask_0, x = scores_85_cast_fp16)[name = string("op_5299_cast_fp16")];
            tensor<int32, [4]> var_5300 = const()[name = string("op_5300"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_5301_cast_fp16 = reshape(shape = var_5300, x = var_5299_cast_fp16)[name = string("op_5301_cast_fp16")];
            tensor<int32, [4]> bd_43_begin_0 = const()[name = string("bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_43_end_0 = const()[name = string("bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_43_end_mask_0 = const()[name = string("bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_43_cast_fp16 = slice_by_index(begin = bd_43_begin_0, end = bd_43_end_0, end_mask = bd_43_end_mask_0, x = var_5301_cast_fp16)[name = string("bd_43_cast_fp16")];
            bool var_5304_transpose_x_0 = const()[name = string("op_5304_transpose_x_0"), val = bool(false)];
            bool var_5304_transpose_y_0 = const()[name = string("op_5304_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_278_perm_0 = const()[name = string("transpose_278_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_279_perm_0 = const()[name = string("transpose_279_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_279 = transpose(perm = transpose_279_perm_0, x = var_5273_cast_fp16)[name = string("transpose_309")];
            tensor<fp16, [1, 8, 188, 128]> transpose_278 = transpose(perm = transpose_278_perm_0, x = q_u_43_cast_fp16)[name = string("transpose_310")];
            tensor<fp16, [1, 8, 188, 188]> var_5304_cast_fp16 = matmul(transpose_x = var_5304_transpose_x_0, transpose_y = var_5304_transpose_y_0, x = transpose_278, y = transpose_279)[name = string("op_5304_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_5305_cast_fp16 = add(x = var_5304_cast_fp16, y = bd_43_cast_fp16)[name = string("op_5305_cast_fp16")];
            fp16 var_5306_to_fp16 = const()[name = string("op_5306_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_5307_cast_fp16 = mul(x = var_5305_cast_fp16, y = var_5306_to_fp16)[name = string("op_5307_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_87_cast_fp16 = add(x = var_5307_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_87_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_5309_cast_fp16 = softmax(axis = var_5232, x = scores_87_cast_fp16)[name = string("op_5309_cast_fp16")];
            bool out_43_transpose_x_0 = const()[name = string("out_43_transpose_x_0"), val = bool(false)];
            bool out_43_transpose_y_0 = const()[name = string("out_43_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_5278_cast_fp16)[name = string("transpose_308")];
            tensor<fp16, [1, 8, 188, 128]> out_43_cast_fp16 = matmul(transpose_x = out_43_transpose_x_0, transpose_y = out_43_transpose_y_0, x = var_5309_cast_fp16, y = v_43_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<int32, [4]> var_5311_perm_0 = const()[name = string("op_5311_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5312 = const()[name = string("op_5312"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_5311_cast_fp16 = transpose(perm = var_5311_perm_0, x = out_43_cast_fp16)[name = string("transpose_307")];
            tensor<fp16, [1, 188, 1024]> input_883_cast_fp16 = reshape(shape = var_5312, x = var_5311_cast_fp16)[name = string("input_883_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_21_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549974400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551023040))))[name = string("blocks_21_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_240_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_21_o_weight_to_fp16_quantized, x = input_883_cast_fp16)[name = string("linear_240_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_885_cast_fp16 = add(x = input_877_cast_fp16, y = linear_240_cast_fp16)[name = string("input_885_cast_fp16")];
            tensor<int32, [1]> input_887_axes_0 = const()[name = string("input_887_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_21_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551025152)))];
            tensor<fp16, [1024]> enc_layers_21_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551027264)))];
            tensor<fp16, [1, 188, 1024]> input_887_cast_fp16 = layer_norm(axes = input_887_axes_0, beta = enc_layers_21_norm_conv_bias_to_fp16, epsilon = var_5240_to_fp16, gamma = enc_layers_21_norm_conv_weight_to_fp16, x = input_885_cast_fp16)[name = string("input_887_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_21_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551029376))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553126592))))[name = string("blocks_21_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_241_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_21_pw1_weight_to_fp16_quantized, x = input_887_cast_fp16)[name = string("linear_241_cast_fp16")];
            int32 var_5321_split_num_splits_0 = const()[name = string("op_5321_split_num_splits_0"), val = int32(2)];
            int32 var_5321_split_axis_0 = const()[name = string("op_5321_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_5321_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_5321_split_cast_fp16_1 = split(axis = var_5321_split_axis_0, num_splits = var_5321_split_num_splits_0, x = linear_241_cast_fp16)[name = string("op_5321_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_5321_split_1_sigmoid_cast_fp16 = sigmoid(x = var_5321_split_cast_fp16_1)[name = string("op_5321_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_5321_cast_fp16 = mul(x = var_5321_split_cast_fp16_0, y = var_5321_split_1_sigmoid_cast_fp16)[name = string("op_5321_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_85_cast_fp16 = mul(x = var_5321_cast_fp16, y = valid_cast_fp16)[name = string("h_85_cast_fp16")];
            tensor<int32, [3]> input_891_perm_0 = const()[name = string("input_891_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_893_pad_type_0 = const()[name = string("input_893_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_893_pad_0 = const()[name = string("input_893_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_893_groups_0 = const()[name = string("input_893_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_893_strides_0 = const()[name = string("input_893_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_893_dilations_0 = const()[name = string("input_893_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_80_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553130752))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553140032))))[name = string("const_80_to_fp16_quantized")];
            tensor<fp16, [1024]> const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553142144)))];
            tensor<fp16, [1, 1024, 188]> input_891_cast_fp16 = transpose(perm = input_891_perm_0, x = h_85_cast_fp16)[name = string("transpose_306")];
            tensor<fp16, [1, 1024, 188]> input_895_cast_fp16 = conv(bias = const_81_to_fp16, dilations = input_893_dilations_0, groups = input_893_groups_0, pad = input_893_pad_0, pad_type = input_893_pad_type_0, strides = input_893_strides_0, weight = const_80_to_fp16_quantized, x = input_891_cast_fp16)[name = string("input_895_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_87_cast_fp16 = silu(x = input_895_cast_fp16)[name = string("h_87_cast_fp16")];
            tensor<int32, [3]> input_897_perm_0 = const()[name = string("input_897_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_21_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553144256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554192896))))[name = string("blocks_21_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_897_cast_fp16 = transpose(perm = input_897_perm_0, x = h_87_cast_fp16)[name = string("transpose_305")];
            tensor<fp16, [1, 188, 1024]> linear_242_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_21_pw2_weight_to_fp16_quantized, x = input_897_cast_fp16)[name = string("linear_242_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_899_cast_fp16 = add(x = input_885_cast_fp16, y = linear_242_cast_fp16)[name = string("input_899_cast_fp16")];
            tensor<int32, [1]> input_901_axes_0 = const()[name = string("input_901_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554195008)))];
            tensor<fp16, [1024]> enc_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554197120)))];
            tensor<fp16, [1, 188, 1024]> input_901_cast_fp16 = layer_norm(axes = input_901_axes_0, beta = enc_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_5240_to_fp16, gamma = enc_layers_21_norm_feed_forward2_weight_to_fp16, x = input_899_cast_fp16)[name = string("input_901_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_21_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554199232))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(558393600))))[name = string("blocks_21_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_243_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_21_ff2_1_weight_to_fp16_quantized, x = input_901_cast_fp16)[name = string("linear_243_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_905_cast_fp16 = silu(x = linear_243_cast_fp16)[name = string("input_905_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_21_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(558401856))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562596224))))[name = string("blocks_21_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_244_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_21_ff2_2_weight_to_fp16_quantized, x = input_905_cast_fp16)[name = string("linear_244_cast_fp16")];
            fp16 var_5342_to_fp16 = const()[name = string("op_5342_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_5343_cast_fp16 = mul(x = linear_244_cast_fp16, y = var_5342_to_fp16)[name = string("op_5343_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_907_cast_fp16 = add(x = input_899_cast_fp16, y = var_5343_cast_fp16)[name = string("input_907_cast_fp16")];
            tensor<int32, [1]> input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_21_norm_out_weight_to_fp16 = const()[name = string("enc_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562598336)))];
            tensor<fp16, [1024]> enc_layers_21_norm_out_bias_to_fp16 = const()[name = string("enc_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562600448)))];
            tensor<fp16, [1, 188, 1024]> input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = enc_layers_21_norm_out_bias_to_fp16, epsilon = var_5240_to_fp16, gamma = enc_layers_21_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")];
            int32 var_5356 = const()[name = string("op_5356"), val = int32(-1)];
            tensor<int32, [1]> input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562602560)))];
            tensor<fp16, [1024]> enc_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562604672)))];
            fp16 var_5364_to_fp16 = const()[name = string("op_5364_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = enc_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_5364_to_fp16, gamma = enc_layers_22_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_22_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562606784))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566801152))))[name = string("blocks_22_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_245_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_22_ff1_1_weight_to_fp16_quantized, x = input_911_cast_fp16)[name = string("linear_245_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_915_cast_fp16 = silu(x = linear_245_cast_fp16)[name = string("input_915_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_22_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566809408))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571003776))))[name = string("blocks_22_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_246_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_22_ff1_2_weight_to_fp16_quantized, x = input_915_cast_fp16)[name = string("linear_246_cast_fp16")];
            fp16 var_5384_to_fp16 = const()[name = string("op_5384_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_5385_cast_fp16 = mul(x = linear_246_cast_fp16, y = var_5384_to_fp16)[name = string("op_5385_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_917_cast_fp16 = add(x = input_909_cast_fp16, y = var_5385_cast_fp16)[name = string("input_917_cast_fp16")];
            tensor<int32, [1]> input_919_axes_0 = const()[name = string("input_919_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571005888)))];
            tensor<fp16, [1024]> enc_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571008000)))];
            tensor<fp16, [1, 188, 1024]> input_919_cast_fp16 = layer_norm(axes = input_919_axes_0, beta = enc_layers_22_norm_self_att_bias_to_fp16, epsilon = var_5364_to_fp16, gamma = enc_layers_22_norm_self_att_weight_to_fp16, x = input_917_cast_fp16)[name = string("input_919_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_22_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571010112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572058752))))[name = string("blocks_22_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_247_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_22_q_weight_to_fp16_quantized, x = input_919_cast_fp16)[name = string("linear_247_cast_fp16")];
            tensor<int32, [4]> var_5391 = const()[name = string("op_5391"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5392_cast_fp16 = reshape(shape = var_5391, x = linear_247_cast_fp16)[name = string("op_5392_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_22_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572060864))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573109504))))[name = string("blocks_22_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_248_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_22_k_weight_to_fp16_quantized, x = input_919_cast_fp16)[name = string("linear_248_cast_fp16")];
            tensor<int32, [4]> var_5396 = const()[name = string("op_5396"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5397_cast_fp16 = reshape(shape = var_5396, x = linear_248_cast_fp16)[name = string("op_5397_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_22_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573111616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574160256))))[name = string("blocks_22_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_249_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_22_v_weight_to_fp16_quantized, x = input_919_cast_fp16)[name = string("linear_249_cast_fp16")];
            tensor<int32, [4]> var_5401 = const()[name = string("op_5401"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5402_cast_fp16 = reshape(shape = var_5401, x = linear_249_cast_fp16)[name = string("op_5402_cast_fp16")];
            tensor<int32, [4]> v_45_perm_0 = const()[name = string("v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_130_to_fp16 = const()[name = string("const_130_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574162368)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_45_cast_fp16 = add(x = var_5392_cast_fp16, y = const_130_to_fp16)[name = string("q_u_45_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574164480)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_45_cast_fp16 = add(x = var_5392_cast_fp16, y = const_131_to_fp16)[name = string("q_v_45_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_22_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(574166592))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575215232))))[name = string("blocks_22_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_250_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_22_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_250_cast_fp16")];
            tensor<int32, [4]> var_5412 = const()[name = string("op_5412"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_45_cast_fp16 = reshape(shape = var_5412, x = linear_250_cast_fp16)[name = string("p_45_cast_fp16")];
            bool input_921_transpose_x_0 = const()[name = string("input_921_transpose_x_0"), val = bool(false)];
            bool input_921_transpose_y_0 = const()[name = string("input_921_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_280_perm_0 = const()[name = string("transpose_280_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_281_perm_0 = const()[name = string("transpose_281_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_281 = transpose(perm = transpose_281_perm_0, x = p_45_cast_fp16)[name = string("transpose_303")];
            tensor<fp16, [1, 8, 188, 128]> transpose_280 = transpose(perm = transpose_280_perm_0, x = q_v_45_cast_fp16)[name = string("transpose_304")];
            tensor<fp16, [1, 8, 188, 375]> input_921_cast_fp16 = matmul(transpose_x = input_921_transpose_x_0, transpose_y = input_921_transpose_y_0, x = transpose_280, y = transpose_281)[name = string("input_921_cast_fp16")];
            tensor<int32, [8]> var_5418_pad_0 = const()[name = string("op_5418_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_5418_mode_0 = const()[name = string("op_5418_mode_0"), val = string("constant")];
            fp16 const_36_to_fp16 = const()[name = string("const_36_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_5418_cast_fp16 = pad(constant_val = const_36_to_fp16, mode = var_5418_mode_0, pad = var_5418_pad_0, x = input_921_cast_fp16)[name = string("op_5418_cast_fp16")];
            tensor<int32, [4]> var_5419 = const()[name = string("op_5419"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_89_cast_fp16 = reshape(shape = var_5419, x = var_5418_cast_fp16)[name = string("scores_89_cast_fp16")];
            tensor<int32, [4]> var_5423_begin_0 = const()[name = string("op_5423_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5423_end_0 = const()[name = string("op_5423_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_5423_end_mask_0 = const()[name = string("op_5423_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_5423_cast_fp16 = slice_by_index(begin = var_5423_begin_0, end = var_5423_end_0, end_mask = var_5423_end_mask_0, x = scores_89_cast_fp16)[name = string("op_5423_cast_fp16")];
            tensor<int32, [4]> var_5424 = const()[name = string("op_5424"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_5425_cast_fp16 = reshape(shape = var_5424, x = var_5423_cast_fp16)[name = string("op_5425_cast_fp16")];
            tensor<int32, [4]> bd_45_begin_0 = const()[name = string("bd_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_45_end_0 = const()[name = string("bd_45_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_45_end_mask_0 = const()[name = string("bd_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_45_cast_fp16 = slice_by_index(begin = bd_45_begin_0, end = bd_45_end_0, end_mask = bd_45_end_mask_0, x = var_5425_cast_fp16)[name = string("bd_45_cast_fp16")];
            bool var_5428_transpose_x_0 = const()[name = string("op_5428_transpose_x_0"), val = bool(false)];
            bool var_5428_transpose_y_0 = const()[name = string("op_5428_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_282_perm_0 = const()[name = string("transpose_282_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_283_perm_0 = const()[name = string("transpose_283_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_283 = transpose(perm = transpose_283_perm_0, x = var_5397_cast_fp16)[name = string("transpose_301")];
            tensor<fp16, [1, 8, 188, 128]> transpose_282 = transpose(perm = transpose_282_perm_0, x = q_u_45_cast_fp16)[name = string("transpose_302")];
            tensor<fp16, [1, 8, 188, 188]> var_5428_cast_fp16 = matmul(transpose_x = var_5428_transpose_x_0, transpose_y = var_5428_transpose_y_0, x = transpose_282, y = transpose_283)[name = string("op_5428_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_5429_cast_fp16 = add(x = var_5428_cast_fp16, y = bd_45_cast_fp16)[name = string("op_5429_cast_fp16")];
            fp16 var_5430_to_fp16 = const()[name = string("op_5430_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_5431_cast_fp16 = mul(x = var_5429_cast_fp16, y = var_5430_to_fp16)[name = string("op_5431_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_91_cast_fp16 = add(x = var_5431_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_91_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_5433_cast_fp16 = softmax(axis = var_5356, x = scores_91_cast_fp16)[name = string("op_5433_cast_fp16")];
            bool out_45_transpose_x_0 = const()[name = string("out_45_transpose_x_0"), val = bool(false)];
            bool out_45_transpose_y_0 = const()[name = string("out_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_5402_cast_fp16)[name = string("transpose_300")];
            tensor<fp16, [1, 8, 188, 128]> out_45_cast_fp16 = matmul(transpose_x = out_45_transpose_x_0, transpose_y = out_45_transpose_y_0, x = var_5433_cast_fp16, y = v_45_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<int32, [4]> var_5435_perm_0 = const()[name = string("op_5435_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5436 = const()[name = string("op_5436"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_5435_cast_fp16 = transpose(perm = var_5435_perm_0, x = out_45_cast_fp16)[name = string("transpose_299")];
            tensor<fp16, [1, 188, 1024]> input_923_cast_fp16 = reshape(shape = var_5436, x = var_5435_cast_fp16)[name = string("input_923_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_22_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575217344))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576265984))))[name = string("blocks_22_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_251_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_22_o_weight_to_fp16_quantized, x = input_923_cast_fp16)[name = string("linear_251_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_925_cast_fp16 = add(x = input_917_cast_fp16, y = linear_251_cast_fp16)[name = string("input_925_cast_fp16")];
            tensor<int32, [1]> input_927_axes_0 = const()[name = string("input_927_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_22_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576268096)))];
            tensor<fp16, [1024]> enc_layers_22_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576270208)))];
            tensor<fp16, [1, 188, 1024]> input_927_cast_fp16 = layer_norm(axes = input_927_axes_0, beta = enc_layers_22_norm_conv_bias_to_fp16, epsilon = var_5364_to_fp16, gamma = enc_layers_22_norm_conv_weight_to_fp16, x = input_925_cast_fp16)[name = string("input_927_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_22_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576272320))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578369536))))[name = string("blocks_22_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_252_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_22_pw1_weight_to_fp16_quantized, x = input_927_cast_fp16)[name = string("linear_252_cast_fp16")];
            int32 var_5445_split_num_splits_0 = const()[name = string("op_5445_split_num_splits_0"), val = int32(2)];
            int32 var_5445_split_axis_0 = const()[name = string("op_5445_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_5445_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_5445_split_cast_fp16_1 = split(axis = var_5445_split_axis_0, num_splits = var_5445_split_num_splits_0, x = linear_252_cast_fp16)[name = string("op_5445_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_5445_split_1_sigmoid_cast_fp16 = sigmoid(x = var_5445_split_cast_fp16_1)[name = string("op_5445_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_5445_cast_fp16 = mul(x = var_5445_split_cast_fp16_0, y = var_5445_split_1_sigmoid_cast_fp16)[name = string("op_5445_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_89_cast_fp16 = mul(x = var_5445_cast_fp16, y = valid_cast_fp16)[name = string("h_89_cast_fp16")];
            tensor<int32, [3]> input_931_perm_0 = const()[name = string("input_931_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_933_pad_type_0 = const()[name = string("input_933_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_933_pad_0 = const()[name = string("input_933_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_933_groups_0 = const()[name = string("input_933_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_933_strides_0 = const()[name = string("input_933_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_933_dilations_0 = const()[name = string("input_933_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_82_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578373696))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578382976))))[name = string("const_82_to_fp16_quantized")];
            tensor<fp16, [1024]> const_83_to_fp16 = const()[name = string("const_83_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578385088)))];
            tensor<fp16, [1, 1024, 188]> input_931_cast_fp16 = transpose(perm = input_931_perm_0, x = h_89_cast_fp16)[name = string("transpose_298")];
            tensor<fp16, [1, 1024, 188]> input_935_cast_fp16 = conv(bias = const_83_to_fp16, dilations = input_933_dilations_0, groups = input_933_groups_0, pad = input_933_pad_0, pad_type = input_933_pad_type_0, strides = input_933_strides_0, weight = const_82_to_fp16_quantized, x = input_931_cast_fp16)[name = string("input_935_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_91_cast_fp16 = silu(x = input_935_cast_fp16)[name = string("h_91_cast_fp16")];
            tensor<int32, [3]> input_937_perm_0 = const()[name = string("input_937_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_22_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578387200))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579435840))))[name = string("blocks_22_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_937_cast_fp16 = transpose(perm = input_937_perm_0, x = h_91_cast_fp16)[name = string("transpose_297")];
            tensor<fp16, [1, 188, 1024]> linear_253_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_22_pw2_weight_to_fp16_quantized, x = input_937_cast_fp16)[name = string("linear_253_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_939_cast_fp16 = add(x = input_925_cast_fp16, y = linear_253_cast_fp16)[name = string("input_939_cast_fp16")];
            tensor<int32, [1]> input_941_axes_0 = const()[name = string("input_941_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579437952)))];
            tensor<fp16, [1024]> enc_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579440064)))];
            tensor<fp16, [1, 188, 1024]> input_941_cast_fp16 = layer_norm(axes = input_941_axes_0, beta = enc_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_5364_to_fp16, gamma = enc_layers_22_norm_feed_forward2_weight_to_fp16, x = input_939_cast_fp16)[name = string("input_941_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_22_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579442176))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583636544))))[name = string("blocks_22_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_254_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_22_ff2_1_weight_to_fp16_quantized, x = input_941_cast_fp16)[name = string("linear_254_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_945_cast_fp16 = silu(x = linear_254_cast_fp16)[name = string("input_945_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_22_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583644800))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587839168))))[name = string("blocks_22_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_255_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_22_ff2_2_weight_to_fp16_quantized, x = input_945_cast_fp16)[name = string("linear_255_cast_fp16")];
            fp16 var_5466_to_fp16 = const()[name = string("op_5466_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_5467_cast_fp16 = mul(x = linear_255_cast_fp16, y = var_5466_to_fp16)[name = string("op_5467_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_947_cast_fp16 = add(x = input_939_cast_fp16, y = var_5467_cast_fp16)[name = string("input_947_cast_fp16")];
            tensor<int32, [1]> input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_22_norm_out_weight_to_fp16 = const()[name = string("enc_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587841280)))];
            tensor<fp16, [1024]> enc_layers_22_norm_out_bias_to_fp16 = const()[name = string("enc_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587843392)))];
            tensor<fp16, [1, 188, 1024]> input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = enc_layers_22_norm_out_bias_to_fp16, epsilon = var_5364_to_fp16, gamma = enc_layers_22_norm_out_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")];
            int32 var_5480 = const()[name = string("op_5480"), val = int32(-1)];
            tensor<int32, [1]> input_951_axes_0 = const()[name = string("input_951_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("enc_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587845504)))];
            tensor<fp16, [1024]> enc_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("enc_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587847616)))];
            fp16 var_5488_to_fp16 = const()[name = string("op_5488_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 188, 1024]> input_951_cast_fp16 = layer_norm(axes = input_951_axes_0, beta = enc_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_5488_to_fp16, gamma = enc_layers_23_norm_feed_forward1_weight_to_fp16, x = input_949_cast_fp16)[name = string("input_951_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_23_ff1_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587849728))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592044096))))[name = string("blocks_23_ff1_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_256_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_23_ff1_1_weight_to_fp16_quantized, x = input_951_cast_fp16)[name = string("linear_256_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_955_cast_fp16 = silu(x = linear_256_cast_fp16)[name = string("input_955_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_23_ff1_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592052352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596246720))))[name = string("blocks_23_ff1_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_257_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_23_ff1_2_weight_to_fp16_quantized, x = input_955_cast_fp16)[name = string("linear_257_cast_fp16")];
            fp16 var_5508_to_fp16 = const()[name = string("op_5508_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_5509_cast_fp16 = mul(x = linear_257_cast_fp16, y = var_5508_to_fp16)[name = string("op_5509_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_957_cast_fp16 = add(x = input_949_cast_fp16, y = var_5509_cast_fp16)[name = string("input_957_cast_fp16")];
            tensor<int32, [1]> input_959_axes_0 = const()[name = string("input_959_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("enc_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596248832)))];
            tensor<fp16, [1024]> enc_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("enc_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596250944)))];
            tensor<fp16, [1, 188, 1024]> input_959_cast_fp16 = layer_norm(axes = input_959_axes_0, beta = enc_layers_23_norm_self_att_bias_to_fp16, epsilon = var_5488_to_fp16, gamma = enc_layers_23_norm_self_att_weight_to_fp16, x = input_957_cast_fp16)[name = string("input_959_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_23_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596253056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597301696))))[name = string("blocks_23_q_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_258_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_23_q_weight_to_fp16_quantized, x = input_959_cast_fp16)[name = string("linear_258_cast_fp16")];
            tensor<int32, [4]> var_5515 = const()[name = string("op_5515"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5516_cast_fp16 = reshape(shape = var_5515, x = linear_258_cast_fp16)[name = string("op_5516_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_23_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597303808))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598352448))))[name = string("blocks_23_k_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_259_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_23_k_weight_to_fp16_quantized, x = input_959_cast_fp16)[name = string("linear_259_cast_fp16")];
            tensor<int32, [4]> var_5520 = const()[name = string("op_5520"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5521_cast_fp16 = reshape(shape = var_5520, x = linear_259_cast_fp16)[name = string("op_5521_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_23_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598354560))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599403200))))[name = string("blocks_23_v_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_260_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_23_v_weight_to_fp16_quantized, x = input_959_cast_fp16)[name = string("linear_260_cast_fp16")];
            tensor<int32, [4]> var_5525 = const()[name = string("op_5525"), val = tensor<int32, [4]>([1, 188, 8, 128])];
            tensor<fp16, [1, 188, 8, 128]> var_5526_cast_fp16 = reshape(shape = var_5525, x = linear_260_cast_fp16)[name = string("op_5526_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<fp16, [1, 1, 8, 128]> const_132_to_fp16 = const()[name = string("const_132_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599405312)))];
            tensor<fp16, [1, 188, 8, 128]> q_u_cast_fp16 = add(x = var_5516_cast_fp16, y = const_132_to_fp16)[name = string("q_u_cast_fp16")];
            tensor<fp16, [1, 1, 8, 128]> const_133_to_fp16 = const()[name = string("const_133_to_fp16"), val = tensor<fp16, [1, 1, 8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599407424)))];
            tensor<fp16, [1, 188, 8, 128]> q_v_cast_fp16 = add(x = var_5516_cast_fp16, y = const_133_to_fp16)[name = string("q_v_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_23_pos_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599409536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600458176))))[name = string("blocks_23_pos_weight_to_fp16_quantized")];
            tensor<fp16, [1, 375, 1024]> linear_261_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_23_pos_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_261_cast_fp16")];
            tensor<int32, [4]> var_5536 = const()[name = string("op_5536"), val = tensor<int32, [4]>([1, -1, 8, 128])];
            tensor<fp16, [1, 375, 8, 128]> p_cast_fp16 = reshape(shape = var_5536, x = linear_261_cast_fp16)[name = string("p_cast_fp16")];
            bool input_961_transpose_x_0 = const()[name = string("input_961_transpose_x_0"), val = bool(false)];
            bool input_961_transpose_y_0 = const()[name = string("input_961_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_284_perm_0 = const()[name = string("transpose_284_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_285_perm_0 = const()[name = string("transpose_285_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 375]> transpose_285 = transpose(perm = transpose_285_perm_0, x = p_cast_fp16)[name = string("transpose_295")];
            tensor<fp16, [1, 8, 188, 128]> transpose_284 = transpose(perm = transpose_284_perm_0, x = q_v_cast_fp16)[name = string("transpose_296")];
            tensor<fp16, [1, 8, 188, 375]> input_961_cast_fp16 = matmul(transpose_x = input_961_transpose_x_0, transpose_y = input_961_transpose_y_0, x = transpose_284, y = transpose_285)[name = string("input_961_cast_fp16")];
            tensor<int32, [8]> var_5542_pad_0 = const()[name = string("op_5542_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
            string var_5542_mode_0 = const()[name = string("op_5542_mode_0"), val = string("constant")];
            fp16 const_37_to_fp16 = const()[name = string("const_37_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 8, 188, 376]> var_5542_cast_fp16 = pad(constant_val = const_37_to_fp16, mode = var_5542_mode_0, pad = var_5542_pad_0, x = input_961_cast_fp16)[name = string("op_5542_cast_fp16")];
            tensor<int32, [4]> var_5543 = const()[name = string("op_5543"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<fp16, [1, 8, 376, 188]> scores_93_cast_fp16 = reshape(shape = var_5543, x = var_5542_cast_fp16)[name = string("scores_93_cast_fp16")];
            tensor<int32, [4]> var_5547_begin_0 = const()[name = string("op_5547_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
            tensor<int32, [4]> var_5547_end_0 = const()[name = string("op_5547_end_0"), val = tensor<int32, [4]>([1, 8, 376, 188])];
            tensor<bool, [4]> var_5547_end_mask_0 = const()[name = string("op_5547_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 375, 188]> var_5547_cast_fp16 = slice_by_index(begin = var_5547_begin_0, end = var_5547_end_0, end_mask = var_5547_end_mask_0, x = scores_93_cast_fp16)[name = string("op_5547_cast_fp16")];
            tensor<int32, [4]> var_5548 = const()[name = string("op_5548"), val = tensor<int32, [4]>([1, 8, 188, 375])];
            tensor<fp16, [1, 8, 188, 375]> var_5549_cast_fp16 = reshape(shape = var_5548, x = var_5547_cast_fp16)[name = string("op_5549_cast_fp16")];
            tensor<int32, [4]> bd_begin_0 = const()[name = string("bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> bd_end_0 = const()[name = string("bd_end_0"), val = tensor<int32, [4]>([1, 8, 188, 188])];
            tensor<bool, [4]> bd_end_mask_0 = const()[name = string("bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 188, 188]> bd_cast_fp16 = slice_by_index(begin = bd_begin_0, end = bd_end_0, end_mask = bd_end_mask_0, x = var_5549_cast_fp16)[name = string("bd_cast_fp16")];
            bool var_5552_transpose_x_0 = const()[name = string("op_5552_transpose_x_0"), val = bool(false)];
            bool var_5552_transpose_y_0 = const()[name = string("op_5552_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_286_perm_0 = const()[name = string("transpose_286_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_287_perm_0 = const()[name = string("transpose_287_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 8, 128, 188]> transpose_287 = transpose(perm = transpose_287_perm_0, x = var_5521_cast_fp16)[name = string("transpose_293")];
            tensor<fp16, [1, 8, 188, 128]> transpose_286 = transpose(perm = transpose_286_perm_0, x = q_u_cast_fp16)[name = string("transpose_294")];
            tensor<fp16, [1, 8, 188, 188]> var_5552_cast_fp16 = matmul(transpose_x = var_5552_transpose_x_0, transpose_y = var_5552_transpose_y_0, x = transpose_286, y = transpose_287)[name = string("op_5552_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_5553_cast_fp16 = add(x = var_5552_cast_fp16, y = bd_cast_fp16)[name = string("op_5553_cast_fp16")];
            fp16 var_5554_to_fp16 = const()[name = string("op_5554_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 188, 188]> var_5555_cast_fp16 = mul(x = var_5553_cast_fp16, y = var_5554_to_fp16)[name = string("op_5555_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> scores_cast_fp16 = add(x = var_5555_cast_fp16, y = key_bias_cast_fp16)[name = string("scores_cast_fp16")];
            tensor<fp16, [1, 8, 188, 188]> var_5557_cast_fp16 = softmax(axis = var_5480, x = scores_cast_fp16)[name = string("op_5557_cast_fp16")];
            bool out_transpose_x_0 = const()[name = string("out_transpose_x_0"), val = bool(false)];
            bool out_transpose_y_0 = const()[name = string("out_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 188, 128]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = var_5526_cast_fp16)[name = string("transpose_292")];
            tensor<fp16, [1, 8, 188, 128]> out_cast_fp16 = matmul(transpose_x = out_transpose_x_0, transpose_y = out_transpose_y_0, x = var_5557_cast_fp16, y = v_47_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [4]> var_5559_perm_0 = const()[name = string("op_5559_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5560 = const()[name = string("op_5560"), val = tensor<int32, [3]>([1, 188, -1])];
            tensor<fp16, [1, 188, 8, 128]> var_5559_cast_fp16 = transpose(perm = var_5559_perm_0, x = out_cast_fp16)[name = string("transpose_291")];
            tensor<fp16, [1, 188, 1024]> input_963_cast_fp16 = reshape(shape = var_5560, x = var_5559_cast_fp16)[name = string("input_963_cast_fp16")];
            tensor<fp16, [1024, 1024]> blocks_23_o_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(600460288))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601508928))))[name = string("blocks_23_o_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_262_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_23_o_weight_to_fp16_quantized, x = input_963_cast_fp16)[name = string("linear_262_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_965_cast_fp16 = add(x = input_957_cast_fp16, y = linear_262_cast_fp16)[name = string("input_965_cast_fp16")];
            tensor<int32, [1]> input_967_axes_0 = const()[name = string("input_967_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_23_norm_conv_weight_to_fp16 = const()[name = string("enc_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601511040)))];
            tensor<fp16, [1024]> enc_layers_23_norm_conv_bias_to_fp16 = const()[name = string("enc_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601513152)))];
            tensor<fp16, [1, 188, 1024]> input_967_cast_fp16 = layer_norm(axes = input_967_axes_0, beta = enc_layers_23_norm_conv_bias_to_fp16, epsilon = var_5488_to_fp16, gamma = enc_layers_23_norm_conv_weight_to_fp16, x = input_965_cast_fp16)[name = string("input_967_cast_fp16")];
            tensor<fp16, [2048, 1024]> blocks_23_pw1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601515264))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603612480))))[name = string("blocks_23_pw1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 2048]> linear_263_cast_fp16 = linear(bias = linear_10_bias_0_to_fp16, weight = blocks_23_pw1_weight_to_fp16_quantized, x = input_967_cast_fp16)[name = string("linear_263_cast_fp16")];
            int32 var_5569_split_num_splits_0 = const()[name = string("op_5569_split_num_splits_0"), val = int32(2)];
            int32 var_5569_split_axis_0 = const()[name = string("op_5569_split_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 188, 1024]> var_5569_split_cast_fp16_0, tensor<fp16, [1, 188, 1024]> var_5569_split_cast_fp16_1 = split(axis = var_5569_split_axis_0, num_splits = var_5569_split_num_splits_0, x = linear_263_cast_fp16)[name = string("op_5569_split_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_5569_split_1_sigmoid_cast_fp16 = sigmoid(x = var_5569_split_cast_fp16_1)[name = string("op_5569_split_1_sigmoid_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> var_5569_cast_fp16 = mul(x = var_5569_split_cast_fp16_0, y = var_5569_split_1_sigmoid_cast_fp16)[name = string("op_5569_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> h_93_cast_fp16 = mul(x = var_5569_cast_fp16, y = valid_cast_fp16)[name = string("h_93_cast_fp16")];
            tensor<int32, [3]> input_971_perm_0 = const()[name = string("input_971_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string input_973_pad_type_0 = const()[name = string("input_973_pad_type_0"), val = string("custom")];
            tensor<int32, [2]> input_973_pad_0 = const()[name = string("input_973_pad_0"), val = tensor<int32, [2]>([4, 4])];
            int32 input_973_groups_0 = const()[name = string("input_973_groups_0"), val = int32(1024)];
            tensor<int32, [1]> input_973_strides_0 = const()[name = string("input_973_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> input_973_dilations_0 = const()[name = string("input_973_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1024, 1, 9]> const_84_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603616640))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603625920))))[name = string("const_84_to_fp16_quantized")];
            tensor<fp16, [1024]> const_85_to_fp16 = const()[name = string("const_85_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603628032)))];
            tensor<fp16, [1, 1024, 188]> input_971_cast_fp16 = transpose(perm = input_971_perm_0, x = h_93_cast_fp16)[name = string("transpose_290")];
            tensor<fp16, [1, 1024, 188]> input_975_cast_fp16 = conv(bias = const_85_to_fp16, dilations = input_973_dilations_0, groups = input_973_groups_0, pad = input_973_pad_0, pad_type = input_973_pad_type_0, strides = input_973_strides_0, weight = const_84_to_fp16_quantized, x = input_971_cast_fp16)[name = string("input_975_cast_fp16")];
            tensor<fp16, [1, 1024, 188]> h_cast_fp16 = silu(x = input_975_cast_fp16)[name = string("h_cast_fp16")];
            tensor<int32, [3]> input_977_perm_0 = const()[name = string("input_977_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 1024]> blocks_23_pw2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(603630144))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604678784))))[name = string("blocks_23_pw2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> input_977_cast_fp16 = transpose(perm = input_977_perm_0, x = h_cast_fp16)[name = string("transpose_289")];
            tensor<fp16, [1, 188, 1024]> linear_264_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_23_pw2_weight_to_fp16_quantized, x = input_977_cast_fp16)[name = string("linear_264_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_979_cast_fp16 = add(x = input_965_cast_fp16, y = linear_264_cast_fp16)[name = string("input_979_cast_fp16")];
            tensor<int32, [1]> input_981_axes_0 = const()[name = string("input_981_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("enc_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604680896)))];
            tensor<fp16, [1024]> enc_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("enc_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604683008)))];
            tensor<fp16, [1, 188, 1024]> input_981_cast_fp16 = layer_norm(axes = input_981_axes_0, beta = enc_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_5488_to_fp16, gamma = enc_layers_23_norm_feed_forward2_weight_to_fp16, x = input_979_cast_fp16)[name = string("input_981_cast_fp16")];
            tensor<fp16, [4096, 1024]> blocks_23_ff2_1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604685120))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608879488))))[name = string("blocks_23_ff2_1_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 4096]> linear_265_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = blocks_23_ff2_1_weight_to_fp16_quantized, x = input_981_cast_fp16)[name = string("linear_265_cast_fp16")];
            tensor<fp16, [1, 188, 4096]> input_985_cast_fp16 = silu(x = linear_265_cast_fp16)[name = string("input_985_cast_fp16")];
            tensor<fp16, [1024, 4096]> blocks_23_ff2_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(608887744))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613082112))))[name = string("blocks_23_ff2_2_weight_to_fp16_quantized")];
            tensor<fp16, [1, 188, 1024]> linear_266_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = blocks_23_ff2_2_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("linear_266_cast_fp16")];
            fp16 var_5590_to_fp16 = const()[name = string("op_5590_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 188, 1024]> var_5591_cast_fp16 = mul(x = linear_266_cast_fp16, y = var_5590_to_fp16)[name = string("op_5591_cast_fp16")];
            tensor<fp16, [1, 188, 1024]> input_cast_fp16 = add(x = input_979_cast_fp16, y = var_5591_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> x_axes_0 = const()[name = string("x_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1024]> enc_layers_23_norm_out_weight_to_fp16 = const()[name = string("enc_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613084224)))];
            tensor<fp16, [1024]> enc_layers_23_norm_out_bias_to_fp16 = const()[name = string("enc_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613086336)))];
            tensor<fp16, [1, 188, 1024]> x_cast_fp16 = layer_norm(axes = x_axes_0, beta = enc_layers_23_norm_out_bias_to_fp16, epsilon = var_5488_to_fp16, gamma = enc_layers_23_norm_out_weight_to_fp16, x = input_cast_fp16)[name = string("x_cast_fp16")];
            tensor<int32, [3]> var_5597_perm_0 = const()[name = string("op_5597_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            string var_5597_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5597_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            string cast_243_dtype_0 = const()[name = string("cast_243_dtype_0"), val = string("int32")];
            tensor<fp16, [1, 1024, 188]> var_5597_cast_fp16 = transpose(perm = var_5597_perm_0, x = x_cast_fp16)[name = string("transpose_288")];
            tensor<fp32, [1, 1024, 188]> encoder = cast(dtype = var_5597_cast_fp16_to_fp32_dtype_0, x = var_5597_cast_fp16)[name = string("cast_0")];
            tensor<int32, [1]> encoder_length = cast(dtype = cast_243_dtype_0, x = length_cast_fp16)[name = string("cast_1")];
        } -> (encoder, encoder_length);
}