program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios17>(tensor<fp32, [1]> guidance_scale, tensor<fp32, [1, 1024]> padding_mask, tensor<fp32, [1, 1024, 100]> speech_condition, tensor<fp32, [1]> t, tensor<fp32, [1, 1024, 100]> text_condition, tensor<fp32, [1, 1024, 100]> x) {
            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(2)];
            tensor<bool, []> xt_interleave_0 = const()[name = tensor<string, []>("xt_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> x_to_fp16_dtype_0 = const()[name = tensor<string, []>("x_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<string, []> text_condition_to_fp16_dtype_0 = const()[name = tensor<string, []>("text_condition_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<string, []> speech_condition_to_fp16_dtype_0 = const()[name = tensor<string, []>("speech_condition_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1024, 100]> speech_condition_to_fp16 = cast(dtype = speech_condition_to_fp16_dtype_0, x = speech_condition)[name = tensor<string, []>("cast_135")];
            tensor<fp16, [1, 1024, 100]> text_condition_to_fp16 = cast(dtype = text_condition_to_fp16_dtype_0, x = text_condition)[name = tensor<string, []>("cast_136")];
            tensor<fp16, [1, 1024, 100]> x_to_fp16 = cast(dtype = x_to_fp16_dtype_0, x = x)[name = tensor<string, []>("cast_137")];
            tensor<fp16, [1, 1024, 300]> xt_cast_fp16 = concat(axis = var_9, interleave = xt_interleave_0, values = (x_to_fp16, text_condition_to_fp16, speech_condition_to_fp16))[name = tensor<string, []>("xt_cast_fp16")];
            tensor<int32, [3]> var_14 = const()[name = tensor<string, []>("op_14"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 300, 1024]> var_15_cast_fp16 = transpose(perm = var_14, x = xt_cast_fp16)[name = tensor<string, []>("transpose_321")];
            tensor<fp16, [1, 300, 1, 1024]> input_7_cast_fp16 = expand_dims(axes = input_7_axes_0, x = var_15_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<int32, [4]> var_22 = const()[name = tensor<string, []>("op_22"), val = tensor<int32, [4]>([1, 1, 1, 1024])];
            tensor<string, []> padding_mask_to_fp16_dtype_0 = const()[name = tensor<string, []>("padding_mask_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1024]> padding_mask_to_fp16 = cast(dtype = padding_mask_to_fp16_dtype_0, x = padding_mask)[name = tensor<string, []>("cast_134")];
            tensor<fp16, [1, 1, 1, 1024]> mask_cast_fp16 = reshape(shape = var_22, x = padding_mask_to_fp16)[name = tensor<string, []>("mask_cast_fp16")];
            tensor<int32, []> var_41 = const()[name = tensor<string, []>("op_41"), val = tensor<int32, []>(-1)];
            tensor<int32, []> var_51 = const()[name = tensor<string, []>("op_51"), val = tensor<int32, []>(1)];
            tensor<int32, [4]> var_71 = const()[name = tensor<string, []>("op_71"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<string, []> t_to_fp16_dtype_0 = const()[name = tensor<string, []>("t_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1]> t_to_fp16 = cast(dtype = t_to_fp16_dtype_0, x = t)[name = tensor<string, []>("cast_133")];
            tensor<fp16, [1, 1, 1, 1]> var_72_cast_fp16 = reshape(shape = var_71, x = t_to_fp16)[name = tensor<string, []>("op_72_cast_fp16")];
            tensor<fp16, [1, 96, 1, 1]> core_freqs_to_fp16 = const()[name = tensor<string, []>("core_freqs_to_fp16"), val = tensor<fp16, [1, 96, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [1, 96, 1, 1]> targs_cast_fp16 = mul(x = var_72_cast_fp16, y = core_freqs_to_fp16)[name = tensor<string, []>("targs_cast_fp16")];
            tensor<fp16, [1, 96, 1, 1]> var_74_cast_fp16 = cos(x = targs_cast_fp16)[name = tensor<string, []>("op_74_cast_fp16")];
            tensor<fp16, [1, 96, 1, 1]> var_75_cast_fp16 = sin(x = targs_cast_fp16)[name = tensor<string, []>("op_75_cast_fp16")];
            tensor<bool, []> te_interleave_0 = const()[name = tensor<string, []>("te_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 192, 1, 1]> te_cast_fp16 = concat(axis = var_51, interleave = te_interleave_0, values = (var_74_cast_fp16, var_75_cast_fp16))[name = tensor<string, []>("te_cast_fp16")];
            tensor<int32, [4]> var_78 = const()[name = tensor<string, []>("op_78"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<string, []> guidance_scale_to_fp16_dtype_0 = const()[name = tensor<string, []>("guidance_scale_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1]> guidance_scale_to_fp16 = cast(dtype = guidance_scale_to_fp16_dtype_0, x = guidance_scale)[name = tensor<string, []>("cast_132")];
            tensor<fp16, [1, 1, 1, 1]> var_79_cast_fp16 = reshape(shape = var_78, x = guidance_scale_to_fp16)[name = tensor<string, []>("op_79_cast_fp16")];
            tensor<fp16, [1, 96, 1, 1]> gargs_cast_fp16 = mul(x = var_79_cast_fp16, y = core_freqs_to_fp16)[name = tensor<string, []>("gargs_cast_fp16")];
            tensor<fp16, [1, 96, 1, 1]> var_81_cast_fp16 = cos(x = gargs_cast_fp16)[name = tensor<string, []>("op_81_cast_fp16")];
            tensor<fp16, [1, 96, 1, 1]> var_82_cast_fp16 = sin(x = gargs_cast_fp16)[name = tensor<string, []>("op_82_cast_fp16")];
            tensor<bool, []> input_1_interleave_0 = const()[name = tensor<string, []>("input_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 192, 1, 1]> input_1_cast_fp16 = concat(axis = var_51, interleave = input_1_interleave_0, values = (var_81_cast_fp16, var_82_cast_fp16))[name = tensor<string, []>("input_1_cast_fp16")];
            tensor<string, []> var_90_pad_type_0 = const()[name = tensor<string, []>("op_90_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_90_strides_0 = const()[name = tensor<string, []>("op_90_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_90_pad_0 = const()[name = tensor<string, []>("op_90_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_90_dilations_0 = const()[name = tensor<string, []>("op_90_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_90_groups_0 = const()[name = tensor<string, []>("op_90_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [192, 192, 1, 1]> core_guid_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_guid_proj_weight_to_fp16"), val = tensor<fp16, [192, 192, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320)))];
            tensor<fp16, [1, 192, 1, 1]> var_90_cast_fp16 = conv(dilations = var_90_dilations_0, groups = var_90_groups_0, pad = var_90_pad_0, pad_type = var_90_pad_type_0, strides = var_90_strides_0, weight = core_guid_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("op_90_cast_fp16")];
            tensor<fp16, [1, 192, 1, 1]> input_3_cast_fp16 = add(x = te_cast_fp16, y = var_90_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<string, []> x_3_pad_type_0 = const()[name = tensor<string, []>("x_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_3_strides_0 = const()[name = tensor<string, []>("x_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_3_pad_0 = const()[name = tensor<string, []>("x_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_3_dilations_0 = const()[name = tensor<string, []>("x_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_3_groups_0 = const()[name = tensor<string, []>("x_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 192, 1, 1]> core_time_lin1_weight_to_fp16 = const()[name = tensor<string, []>("core_time_lin1_weight_to_fp16"), val = tensor<fp16, [384, 192, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74112)))];
            tensor<fp16, [384]> core_time_lin1_bias_to_fp16 = const()[name = tensor<string, []>("core_time_lin1_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(221632)))];
            tensor<fp16, [1, 384, 1, 1]> x_3_cast_fp16 = conv(bias = core_time_lin1_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 = core_time_lin1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
            tensor<fp16, []> var_99_to_fp16 = const()[name = tensor<string, []>("op_99_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 384, 1, 1]> x_5_cast_fp16 = sub(x = x_3_cast_fp16, y = var_99_to_fp16)[name = tensor<string, []>("x_5_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1]> var_101_cast_fp16 = relu(x = x_5_cast_fp16)[name = tensor<string, []>("op_101_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1]> var_102_cast_fp16 = abs(x = x_5_cast_fp16)[name = tensor<string, []>("op_102_cast_fp16")];
            tensor<fp16, []> const_0_promoted_to_fp16 = const()[name = tensor<string, []>("const_0_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 384, 1, 1]> var_103_cast_fp16 = mul(x = var_102_cast_fp16, y = const_0_promoted_to_fp16)[name = tensor<string, []>("op_103_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1]> var_104_cast_fp16 = exp(x = var_103_cast_fp16)[name = tensor<string, []>("op_104_cast_fp16")];
            tensor<fp32, []> var_105_epsilon_0 = const()[name = tensor<string, []>("op_105_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 384, 1, 1]> var_105_cast_fp16 = log(epsilon = var_105_epsilon_0, x = var_104_cast_fp16)[name = tensor<string, []>("op_105_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1]> var_106_cast_fp16 = add(x = var_101_cast_fp16, y = var_105_cast_fp16)[name = tensor<string, []>("op_106_cast_fp16")];
            tensor<fp16, []> var_107_to_fp16 = const()[name = tensor<string, []>("op_107_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 384, 1, 1]> var_108_cast_fp16 = mul(x = x_3_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_108_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1]> var_109_cast_fp16 = sub(x = var_106_cast_fp16, y = var_108_cast_fp16)[name = tensor<string, []>("op_109_cast_fp16")];
            tensor<fp16, []> var_110_to_fp16 = const()[name = tensor<string, []>("op_110_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 384, 1, 1]> input_5_cast_fp16 = sub(x = var_109_cast_fp16, y = var_110_to_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<string, []> x_7_pad_type_0 = const()[name = tensor<string, []>("x_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_7_strides_0 = const()[name = tensor<string, []>("x_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_7_pad_0 = const()[name = tensor<string, []>("x_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_7_dilations_0 = const()[name = tensor<string, []>("x_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_7_groups_0 = const()[name = tensor<string, []>("x_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [192, 384, 1, 1]> core_time_lin2_weight_to_fp16 = const()[name = tensor<string, []>("core_time_lin2_weight_to_fp16"), val = tensor<fp16, [192, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222464)))];
            tensor<fp16, [192]> core_time_lin2_bias_to_fp16 = const()[name = tensor<string, []>("core_time_lin2_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369984)))];
            tensor<fp16, [1, 192, 1, 1]> x_7_cast_fp16 = conv(bias = core_time_lin2_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 = core_time_lin2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
            tensor<fp16, []> var_119_to_fp16 = const()[name = tensor<string, []>("op_119_to_fp16"), val = tensor<fp16, []>(-0x1.f4p+9)];
            tensor<fp16, [1, 1, 1, 1024]> attn_bias_1_cast_fp16 = mul(x = mask_cast_fp16, y = var_119_to_fp16)[name = tensor<string, []>("attn_bias_1_cast_fp16")];
            tensor<fp16, []> var_46_to_fp16 = const()[name = tensor<string, []>("op_46_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 1, 1, 1024]> conv_keep_1_cast_fp16 = sub(x = var_46_to_fp16, y = mask_cast_fp16)[name = tensor<string, []>("conv_keep_1_cast_fp16")];
            tensor<int32, [4]> attn_bias_3_begin_0 = const()[name = tensor<string, []>("attn_bias_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> attn_bias_3_end_0 = const()[name = tensor<string, []>("attn_bias_3_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1024])];
            tensor<bool, [4]> attn_bias_3_end_mask_0 = const()[name = tensor<string, []>("attn_bias_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> attn_bias_3_stride_0 = const()[name = tensor<string, []>("attn_bias_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 1, 512]> attn_bias_3_cast_fp16 = slice_by_index(begin = attn_bias_3_begin_0, end = attn_bias_3_end_0, end_mask = attn_bias_3_end_mask_0, stride = attn_bias_3_stride_0, x = attn_bias_1_cast_fp16)[name = tensor<string, []>("attn_bias_3_cast_fp16")];
            tensor<int32, [4]> conv_keep_3_begin_0 = const()[name = tensor<string, []>("conv_keep_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> conv_keep_3_end_0 = const()[name = tensor<string, []>("conv_keep_3_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1024])];
            tensor<bool, [4]> conv_keep_3_end_mask_0 = const()[name = tensor<string, []>("conv_keep_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> conv_keep_3_stride_0 = const()[name = tensor<string, []>("conv_keep_3_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 2])];
            tensor<fp16, [1, 1, 1, 512]> conv_keep_3_cast_fp16 = slice_by_index(begin = conv_keep_3_begin_0, end = conv_keep_3_end_0, end_mask = conv_keep_3_end_mask_0, stride = conv_keep_3_stride_0, x = conv_keep_1_cast_fp16)[name = tensor<string, []>("conv_keep_3_cast_fp16")];
            tensor<int32, [4]> attn_bias_begin_0 = const()[name = tensor<string, []>("attn_bias_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> attn_bias_end_0 = const()[name = tensor<string, []>("attn_bias_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1024])];
            tensor<bool, [4]> attn_bias_end_mask_0 = const()[name = tensor<string, []>("attn_bias_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> attn_bias_stride_0 = const()[name = tensor<string, []>("attn_bias_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 4])];
            tensor<fp16, [1, 1, 1, 256]> attn_bias_cast_fp16 = slice_by_index(begin = attn_bias_begin_0, end = attn_bias_end_0, end_mask = attn_bias_end_mask_0, stride = attn_bias_stride_0, x = attn_bias_1_cast_fp16)[name = tensor<string, []>("attn_bias_cast_fp16")];
            tensor<int32, [4]> conv_keep_begin_0 = const()[name = tensor<string, []>("conv_keep_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> conv_keep_end_0 = const()[name = tensor<string, []>("conv_keep_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1024])];
            tensor<bool, [4]> conv_keep_end_mask_0 = const()[name = tensor<string, []>("conv_keep_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> conv_keep_stride_0 = const()[name = tensor<string, []>("conv_keep_stride_0"), val = tensor<int32, [4]>([1, 1, 1, 4])];
            tensor<fp16, [1, 1, 1, 256]> conv_keep_cast_fp16 = slice_by_index(begin = conv_keep_begin_0, end = conv_keep_end_0, end_mask = conv_keep_end_mask_0, stride = conv_keep_stride_0, x = conv_keep_1_cast_fp16)[name = tensor<string, []>("conv_keep_cast_fp16")];
            tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_11_strides_0 = const()[name = tensor<string, []>("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_11_pad_0 = const()[name = tensor<string, []>("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_11_dilations_0 = const()[name = tensor<string, []>("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> input_11_groups_0 = const()[name = tensor<string, []>("input_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 300, 1, 1]> core_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_in_proj_weight_to_fp16"), val = tensor<fp16, [512, 300, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370432)))];
            tensor<fp16, [512]> core_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_in_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(677696)))];
            tensor<fp16, [1, 512, 1, 1024]> input_11_cast_fp16 = conv(bias = core_in_proj_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = core_in_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<fp16, []> var_138_to_fp16 = const()[name = tensor<string, []>("op_138_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 192, 1, 1]> x_9_cast_fp16 = sub(x = x_7_cast_fp16, y = var_138_to_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
            tensor<fp16, [1, 192, 1, 1]> var_140_cast_fp16 = relu(x = x_9_cast_fp16)[name = tensor<string, []>("op_140_cast_fp16")];
            tensor<fp16, [1, 192, 1, 1]> var_141_cast_fp16 = abs(x = x_9_cast_fp16)[name = tensor<string, []>("op_141_cast_fp16")];
            tensor<fp16, []> const_1_promoted_to_fp16 = const()[name = tensor<string, []>("const_1_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 192, 1, 1]> var_142_cast_fp16 = mul(x = var_141_cast_fp16, y = const_1_promoted_to_fp16)[name = tensor<string, []>("op_142_cast_fp16")];
            tensor<fp16, [1, 192, 1, 1]> var_143_cast_fp16 = exp(x = var_142_cast_fp16)[name = tensor<string, []>("op_143_cast_fp16")];
            tensor<fp32, []> var_144_epsilon_0 = const()[name = tensor<string, []>("op_144_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 192, 1, 1]> var_144_cast_fp16 = log(epsilon = var_144_epsilon_0, x = var_143_cast_fp16)[name = tensor<string, []>("op_144_cast_fp16")];
            tensor<fp16, [1, 192, 1, 1]> var_145_cast_fp16 = add(x = var_140_cast_fp16, y = var_144_cast_fp16)[name = tensor<string, []>("op_145_cast_fp16")];
            tensor<fp16, []> var_146_to_fp16 = const()[name = tensor<string, []>("op_146_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 192, 1, 1]> var_147_cast_fp16 = mul(x = x_7_cast_fp16, y = var_146_to_fp16)[name = tensor<string, []>("op_147_cast_fp16")];
            tensor<fp16, [1, 192, 1, 1]> var_148_cast_fp16 = sub(x = var_145_cast_fp16, y = var_147_cast_fp16)[name = tensor<string, []>("op_148_cast_fp16")];
            tensor<fp16, []> var_149_to_fp16 = const()[name = tensor<string, []>("op_149_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 192, 1, 1]> input_9_cast_fp16 = sub(x = var_148_cast_fp16, y = var_149_to_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<string, []> time_emb_1_pad_type_0 = const()[name = tensor<string, []>("time_emb_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> time_emb_1_strides_0 = const()[name = tensor<string, []>("time_emb_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> time_emb_1_pad_0 = const()[name = tensor<string, []>("time_emb_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> time_emb_1_dilations_0 = const()[name = tensor<string, []>("time_emb_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> time_emb_1_groups_0 = const()[name = tensor<string, []>("time_emb_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 192, 1, 1]> core_stacks_0_time_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_time_proj_weight_to_fp16"), val = tensor<fp16, [512, 192, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(678784)))];
            tensor<fp16, [512]> core_stacks_0_time_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_time_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(875456)))];
            tensor<fp16, [1, 512, 1, 1]> time_emb_1_cast_fp16 = conv(bias = core_stacks_0_time_proj_bias_to_fp16, dilations = time_emb_1_dilations_0, groups = time_emb_1_groups_0, pad = time_emb_1_pad_0, pad_type = time_emb_1_pad_type_0, strides = time_emb_1_strides_0, weight = core_stacks_0_time_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("time_emb_1_cast_fp16")];
            tensor<string, []> proj_1_pad_type_0 = const()[name = tensor<string, []>("proj_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_1_strides_0 = const()[name = tensor<string, []>("proj_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_1_pad_0 = const()[name = tensor<string, []>("proj_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_1_dilations_0 = const()[name = tensor<string, []>("proj_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_1_groups_0 = const()[name = tensor<string, []>("proj_1_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_0_layers_0_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(876544)))];
            tensor<fp16, [384]> core_stacks_0_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1269824)))];
            tensor<fp16, [1, 384, 1, 1024]> proj_1_cast_fp16 = conv(bias = core_stacks_0_layers_0_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_1_dilations_0, groups = proj_1_groups_0, pad = proj_1_pad_0, pad_type = proj_1_pad_type_0, strides = proj_1_strides_0, weight = core_stacks_0_layers_0_self_attn_weights_in_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("proj_1_cast_fp16")];
            tensor<int32, [4]> var_179_begin_0 = const()[name = tensor<string, []>("op_179_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_179_end_0 = const()[name = tensor<string, []>("op_179_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1024])];
            tensor<bool, [4]> var_179_end_mask_0 = const()[name = tensor<string, []>("op_179_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_179_cast_fp16 = slice_by_index(begin = var_179_begin_0, end = var_179_end_0, end_mask = var_179_end_mask_0, x = proj_1_cast_fp16)[name = tensor<string, []>("op_179_cast_fp16")];
            tensor<int32, [4]> var_180 = const()[name = tensor<string, []>("op_180"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_181_cast_fp16 = reshape(shape = var_180, x = var_179_cast_fp16)[name = tensor<string, []>("op_181_cast_fp16")];
            tensor<int32, [4]> var_183_begin_0 = const()[name = tensor<string, []>("op_183_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_183_end_0 = const()[name = tensor<string, []>("op_183_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1024])];
            tensor<bool, [4]> var_183_end_mask_0 = const()[name = tensor<string, []>("op_183_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_183_cast_fp16 = slice_by_index(begin = var_183_begin_0, end = var_183_end_0, end_mask = var_183_end_mask_0, x = proj_1_cast_fp16)[name = tensor<string, []>("op_183_cast_fp16")];
            tensor<int32, [4]> var_184 = const()[name = tensor<string, []>("op_184"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_185_cast_fp16 = reshape(shape = var_184, x = var_183_cast_fp16)[name = tensor<string, []>("op_185_cast_fp16")];
            tensor<int32, [4]> var_187_begin_0 = const()[name = tensor<string, []>("op_187_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_187_end_0 = const()[name = tensor<string, []>("op_187_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_187_end_mask_0 = const()[name = tensor<string, []>("op_187_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_187_cast_fp16 = slice_by_index(begin = var_187_begin_0, end = var_187_end_0, end_mask = var_187_end_mask_0, x = proj_1_cast_fp16)[name = tensor<string, []>("op_187_cast_fp16")];
            tensor<int32, [4]> var_188 = const()[name = tensor<string, []>("op_188"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> pt_1_cast_fp16 = reshape(shape = var_188, x = var_187_cast_fp16)[name = tensor<string, []>("pt_1_cast_fp16")];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<int32, [3]>([4, 1024, 32])];
            tensor<fp16, [4, 1, 1024, 32]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = var_181_cast_fp16)[name = tensor<string, []>("transpose_320")];
            tensor<fp16, [4, 1024, 32]> reshape_0_cast_fp16 = reshape(shape = concat_4, x = transpose_0_cast_fp16)[name = tensor<string, []>("reshape_0_cast_fp16")];
            tensor<int32, [4]> transpose_1_perm_0 = const()[name = tensor<string, []>("transpose_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_5 = const()[name = tensor<string, []>("concat_5"), val = tensor<int32, [3]>([4, 32, 1024])];
            tensor<fp16, [4, 1, 32, 1024]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = var_185_cast_fp16)[name = tensor<string, []>("transpose_319")];
            tensor<fp16, [4, 32, 1024]> reshape_1_cast_fp16 = reshape(shape = concat_5, x = transpose_1_cast_fp16)[name = tensor<string, []>("reshape_1_cast_fp16")];
            tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 1024]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = reshape_0_cast_fp16, y = reshape_1_cast_fp16)[name = tensor<string, []>("matmul_0_cast_fp16")];
            tensor<int32, [4]> concat_9 = const()[name = tensor<string, []>("concat_9"), val = tensor<int32, [4]>([4, 1, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> reshape_2_cast_fp16 = reshape(shape = concat_9, x = matmul_0_cast_fp16)[name = tensor<string, []>("reshape_2_cast_fp16")];
            tensor<int32, [4]> attn_scores_1_perm_0 = const()[name = tensor<string, []>("attn_scores_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_192 = const()[name = tensor<string, []>("op_192"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 32, 1024, 1024]> core_pos_basis_1024_to_fp16 = const()[name = tensor<string, []>("core_pos_basis_1024_to_fp16"), val = tensor<fp16, [1, 32, 1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1270656)))];
            tensor<fp16, [4, 32, 1024, 1]> var_193_cast_fp16 = transpose(perm = var_192, x = pt_1_cast_fp16)[name = tensor<string, []>("transpose_317")];
            tensor<fp16, [4, 32, 1024, 1024]> var_194_cast_fp16 = mul(x = var_193_cast_fp16, y = core_pos_basis_1024_to_fp16)[name = tensor<string, []>("op_194_cast_fp16")];
            tensor<int32, [1]> pos_scores_1_axes_0 = const()[name = tensor<string, []>("pos_scores_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_1_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_1_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 1024, 1024]> pos_scores_1_cast_fp16 = reduce_sum(axes = pos_scores_1_axes_0, keep_dims = pos_scores_1_keep_dims_0, x = var_194_cast_fp16)[name = tensor<string, []>("pos_scores_1_cast_fp16")];
            tensor<int32, [4]> var_197 = const()[name = tensor<string, []>("op_197"), val = tensor<int32, [4]>([4, 1024, 1, 1024])];
            tensor<fp16, [4, 1024, 1, 1024]> var_198_cast_fp16 = reshape(shape = var_197, x = pos_scores_1_cast_fp16)[name = tensor<string, []>("op_198_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_1_cast_fp16 = transpose(perm = attn_scores_1_perm_0, x = reshape_2_cast_fp16)[name = tensor<string, []>("transpose_318")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_3_cast_fp16 = add(x = attn_scores_1_cast_fp16, y = var_198_cast_fp16)[name = tensor<string, []>("attn_scores_3_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_5_cast_fp16 = add(x = attn_scores_3_cast_fp16, y = attn_bias_1_cast_fp16)[name = tensor<string, []>("attn_scores_5_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_weights_1_cast_fp16 = softmax(axis = var_41, x = attn_scores_5_cast_fp16)[name = tensor<string, []>("attn_weights_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_13_cast_fp16 = add(x = input_11_cast_fp16, y = time_emb_1_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
            tensor<string, []> x_11_pad_type_0 = const()[name = tensor<string, []>("x_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_11_strides_0 = const()[name = tensor<string, []>("x_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_11_pad_0 = const()[name = tensor<string, []>("x_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_11_dilations_0 = const()[name = tensor<string, []>("x_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_11_groups_0 = const()[name = tensor<string, []>("x_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_0_layers_0_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68379584)))];
            tensor<fp16, [1152]> core_stacks_0_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69559296)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_11_cast_fp16 = conv(bias = core_stacks_0_layers_0_feed_forward1_in_proj_bias_to_fp16, dilations = x_11_dilations_0, groups = x_11_groups_0, pad = x_11_pad_0, pad_type = x_11_pad_type_0, strides = x_11_strides_0, weight = core_stacks_0_layers_0_feed_forward1_in_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
            tensor<fp16, []> var_212_to_fp16 = const()[name = tensor<string, []>("op_212_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 1024]> x_13_cast_fp16 = sub(x = x_11_cast_fp16, y = var_212_to_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_214_cast_fp16 = relu(x = x_13_cast_fp16)[name = tensor<string, []>("op_214_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_215_cast_fp16 = abs(x = x_13_cast_fp16)[name = tensor<string, []>("op_215_cast_fp16")];
            tensor<fp16, []> const_2_promoted_to_fp16 = const()[name = tensor<string, []>("const_2_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_216_cast_fp16 = mul(x = var_215_cast_fp16, y = const_2_promoted_to_fp16)[name = tensor<string, []>("op_216_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_217_cast_fp16 = exp(x = var_216_cast_fp16)[name = tensor<string, []>("op_217_cast_fp16")];
            tensor<fp32, []> var_218_epsilon_0 = const()[name = tensor<string, []>("op_218_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_218_cast_fp16 = log(epsilon = var_218_epsilon_0, x = var_217_cast_fp16)[name = tensor<string, []>("op_218_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_219_cast_fp16 = add(x = var_214_cast_fp16, y = var_218_cast_fp16)[name = tensor<string, []>("op_219_cast_fp16")];
            tensor<fp16, []> var_220_to_fp16 = const()[name = tensor<string, []>("op_220_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 1024]> var_221_cast_fp16 = mul(x = x_11_cast_fp16, y = var_220_to_fp16)[name = tensor<string, []>("op_221_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_222_cast_fp16 = sub(x = var_219_cast_fp16, y = var_221_cast_fp16)[name = tensor<string, []>("op_222_cast_fp16")];
            tensor<fp16, []> var_223_to_fp16 = const()[name = tensor<string, []>("op_223_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 1024]> input_15_cast_fp16 = sub(x = var_222_cast_fp16, y = var_223_to_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<string, []> var_231_pad_type_0 = const()[name = tensor<string, []>("op_231_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_231_strides_0 = const()[name = tensor<string, []>("op_231_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_231_pad_0 = const()[name = tensor<string, []>("op_231_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_231_dilations_0 = const()[name = tensor<string, []>("op_231_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_231_groups_0 = const()[name = tensor<string, []>("op_231_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_0_layers_0_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69561664)))];
            tensor<fp16, [512]> core_stacks_0_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70741376)))];
            tensor<fp16, [1, 512, 1, 1024]> var_231_cast_fp16 = conv(bias = core_stacks_0_layers_0_feed_forward1_out_proj_bias_to_fp16, dilations = var_231_dilations_0, groups = var_231_groups_0, pad = var_231_pad_0, pad_type = var_231_pad_type_0, strides = var_231_strides_0, weight = core_stacks_0_layers_0_feed_forward1_out_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("op_231_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_17_cast_fp16 = add(x = input_13_cast_fp16, y = var_231_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<int32, [4]> var_233_begin_0 = const()[name = tensor<string, []>("op_233_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_233_end_0 = const()[name = tensor<string, []>("op_233_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_233_end_mask_0 = const()[name = tensor<string, []>("op_233_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> var_233_cast_fp16 = slice_by_index(begin = var_233_begin_0, end = var_233_end_0, end_mask = var_233_end_mask_0, x = attn_weights_1_cast_fp16)[name = tensor<string, []>("op_233_cast_fp16")];
            tensor<string, []> x_15_pad_type_0 = const()[name = tensor<string, []>("x_15_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_15_strides_0 = const()[name = tensor<string, []>("x_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_15_pad_0 = const()[name = tensor<string, []>("x_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_15_dilations_0 = const()[name = tensor<string, []>("x_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_15_groups_0 = const()[name = tensor<string, []>("x_15_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_0_layers_0_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(70742464)))];
            tensor<fp16, [1152]> core_stacks_0_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71922176)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_15_cast_fp16 = conv(bias = core_stacks_0_layers_0_nonlin_attention_in_proj_bias_to_fp16, dilations = x_15_dilations_0, groups = x_15_groups_0, pad = x_15_pad_0, pad_type = x_15_pad_type_0, strides = x_15_strides_0, weight = core_stacks_0_layers_0_nonlin_attention_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")];
            tensor<int32, [4]> var_244_begin_0 = const()[name = tensor<string, []>("op_244_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_244_end_0 = const()[name = tensor<string, []>("op_244_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_244_end_mask_0 = const()[name = tensor<string, []>("op_244_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_244_cast_fp16 = slice_by_index(begin = var_244_begin_0, end = var_244_end_0, end_mask = var_244_end_mask_0, x = x_15_cast_fp16)[name = tensor<string, []>("op_244_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> s_gate_1_cast_fp16 = tanh(x = var_244_cast_fp16)[name = tensor<string, []>("s_gate_1_cast_fp16")];
            tensor<int32, [4]> var_247_begin_0 = const()[name = tensor<string, []>("op_247_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_247_end_0 = const()[name = tensor<string, []>("op_247_end_0"), val = tensor<int32, [4]>([1, 768, 1, 1024])];
            tensor<bool, [4]> var_247_end_mask_0 = const()[name = tensor<string, []>("op_247_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_247_cast_fp16 = slice_by_index(begin = var_247_begin_0, end = var_247_end_0, end_mask = var_247_end_mask_0, x = x_15_cast_fp16)[name = tensor<string, []>("op_247_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> var_248_cast_fp16 = mul(x = var_247_cast_fp16, y = s_gate_1_cast_fp16)[name = tensor<string, []>("op_248_cast_fp16")];
            tensor<int32, [4]> y_1_begin_0 = const()[name = tensor<string, []>("y_1_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_1_end_0 = const()[name = tensor<string, []>("y_1_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 1024])];
            tensor<bool, [4]> y_1_end_mask_0 = const()[name = tensor<string, []>("y_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> y_1_cast_fp16 = slice_by_index(begin = y_1_begin_0, end = y_1_end_0, end_mask = y_1_end_mask_0, x = x_15_cast_fp16)[name = tensor<string, []>("y_1_cast_fp16")];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = tensor<string, []>("transpose_2_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_14 = const()[name = tensor<string, []>("concat_14"), val = tensor<int32, [3]>([1, 1024, 1024])];
            tensor<fp16, [1, 1, 1024, 1024]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = var_233_cast_fp16)[name = tensor<string, []>("transpose_316")];
            tensor<fp16, [1, 1024, 1024]> reshape_3_cast_fp16 = reshape(shape = concat_14, x = transpose_2_cast_fp16)[name = tensor<string, []>("reshape_3_cast_fp16")];
            tensor<int32, [4]> transpose_3_perm_0 = const()[name = tensor<string, []>("transpose_3_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_15 = const()[name = tensor<string, []>("concat_15"), val = tensor<int32, [3]>([1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = var_248_cast_fp16)[name = tensor<string, []>("transpose_315")];
            tensor<fp16, [1, 1024, 384]> reshape_4_cast_fp16 = reshape(shape = concat_15, x = transpose_3_cast_fp16)[name = tensor<string, []>("reshape_4_cast_fp16")];
            tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1024, 384]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = reshape_3_cast_fp16, y = reshape_4_cast_fp16)[name = tensor<string, []>("matmul_1_cast_fp16")];
            tensor<int32, [4]> concat_19 = const()[name = tensor<string, []>("concat_19"), val = tensor<int32, [4]>([1, 1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> reshape_5_cast_fp16 = reshape(shape = concat_19, x = matmul_1_cast_fp16)[name = tensor<string, []>("reshape_5_cast_fp16")];
            tensor<int32, [4]> v_1_perm_0 = const()[name = tensor<string, []>("v_1_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 1024]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = reshape_5_cast_fp16)[name = tensor<string, []>("transpose_314")];
            tensor<fp16, [1, 384, 1, 1024]> input_19_cast_fp16 = mul(x = v_1_cast_fp16, y = y_1_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<string, []> var_260_pad_type_0 = const()[name = tensor<string, []>("op_260_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_260_strides_0 = const()[name = tensor<string, []>("op_260_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_260_pad_0 = const()[name = tensor<string, []>("op_260_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_260_dilations_0 = const()[name = tensor<string, []>("op_260_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_260_groups_0 = const()[name = tensor<string, []>("op_260_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_0_layers_0_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71924544)))];
            tensor<fp16, [512]> core_stacks_0_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72317824)))];
            tensor<fp16, [1, 512, 1, 1024]> var_260_cast_fp16 = conv(bias = core_stacks_0_layers_0_nonlin_attention_out_proj_bias_to_fp16, dilations = var_260_dilations_0, groups = var_260_groups_0, pad = var_260_pad_0, pad_type = var_260_pad_type_0, strides = var_260_strides_0, weight = core_stacks_0_layers_0_nonlin_attention_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("op_260_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_17_cast_fp16 = add(x = input_17_cast_fp16, y = var_260_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
            tensor<string, []> var_271_pad_type_0 = const()[name = tensor<string, []>("op_271_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_271_strides_0 = const()[name = tensor<string, []>("op_271_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_271_pad_0 = const()[name = tensor<string, []>("op_271_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_271_dilations_0 = const()[name = tensor<string, []>("op_271_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_271_groups_0 = const()[name = tensor<string, []>("op_271_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_0_layers_0_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72318912)))];
            tensor<fp16, [48]> core_stacks_0_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72368128)))];
            tensor<fp16, [1, 48, 1, 1024]> var_271_cast_fp16 = conv(bias = core_stacks_0_layers_0_self_attn1_in_proj_bias_to_fp16, dilations = var_271_dilations_0, groups = var_271_groups_0, pad = var_271_pad_0, pad_type = var_271_pad_type_0, strides = var_271_strides_0, weight = core_stacks_0_layers_0_self_attn1_in_proj_weight_to_fp16, x = x_17_cast_fp16)[name = tensor<string, []>("op_271_cast_fp16")];
            tensor<int32, [4]> var_272 = const()[name = tensor<string, []>("op_272"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_273_cast_fp16 = reshape(shape = var_272, x = var_271_cast_fp16)[name = tensor<string, []>("op_273_cast_fp16")];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = tensor<string, []>("transpose_4_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_24 = const()[name = tensor<string, []>("concat_24"), val = tensor<int32, [3]>([4, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = attn_weights_1_cast_fp16)[name = tensor<string, []>("transpose_313")];
            tensor<fp16, [4, 1024, 1024]> reshape_6_cast_fp16 = reshape(shape = concat_24, x = transpose_4_cast_fp16)[name = tensor<string, []>("reshape_6_cast_fp16")];
            tensor<int32, [4]> transpose_5_perm_0 = const()[name = tensor<string, []>("transpose_5_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_25 = const()[name = tensor<string, []>("concat_25"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = var_273_cast_fp16)[name = tensor<string, []>("transpose_312")];
            tensor<fp16, [4, 1024, 12]> reshape_7_cast_fp16 = reshape(shape = concat_25, x = transpose_5_cast_fp16)[name = tensor<string, []>("reshape_7_cast_fp16")];
            tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_6_cast_fp16, y = reshape_7_cast_fp16)[name = tensor<string, []>("matmul_2_cast_fp16")];
            tensor<int32, [4]> concat_29 = const()[name = tensor<string, []>("concat_29"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_8_cast_fp16 = reshape(shape = concat_29, x = matmul_2_cast_fp16)[name = tensor<string, []>("reshape_8_cast_fp16")];
            tensor<int32, [4]> out_1_perm_0 = const()[name = tensor<string, []>("out_1_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_276 = const()[name = tensor<string, []>("op_276"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_1_cast_fp16 = transpose(perm = out_1_perm_0, x = reshape_8_cast_fp16)[name = tensor<string, []>("transpose_311")];
            tensor<fp16, [1, 48, 1, 1024]> input_21_cast_fp16 = reshape(shape = var_276, x = out_1_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<string, []> var_284_pad_type_0 = const()[name = tensor<string, []>("op_284_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_284_strides_0 = const()[name = tensor<string, []>("op_284_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_284_pad_0 = const()[name = tensor<string, []>("op_284_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_284_dilations_0 = const()[name = tensor<string, []>("op_284_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_284_groups_0 = const()[name = tensor<string, []>("op_284_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_0_layers_0_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72368320)))];
            tensor<fp16, [512]> core_stacks_0_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72417536)))];
            tensor<fp16, [1, 512, 1, 1024]> var_284_cast_fp16 = conv(bias = core_stacks_0_layers_0_self_attn1_out_proj_bias_to_fp16, dilations = var_284_dilations_0, groups = var_284_groups_0, pad = var_284_pad_0, pad_type = var_284_pad_type_0, strides = var_284_strides_0, weight = core_stacks_0_layers_0_self_attn1_out_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_284_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = var_284_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = time_emb_1_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
            tensor<string, []> x_23_pad_type_0 = const()[name = tensor<string, []>("x_23_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_23_strides_0 = const()[name = tensor<string, []>("x_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_23_pad_0 = const()[name = tensor<string, []>("x_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_23_dilations_0 = const()[name = tensor<string, []>("x_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_23_groups_0 = const()[name = tensor<string, []>("x_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_0_layers_0_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72418624)))];
            tensor<fp16, [1024]> core_stacks_0_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73467264)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_23_cast_fp16 = conv(bias = core_stacks_0_layers_0_conv_module1_in_proj_bias_to_fp16, dilations = x_23_dilations_0, groups = x_23_groups_0, pad = x_23_pad_0, pad_type = x_23_pad_type_0, strides = x_23_strides_0, weight = core_stacks_0_layers_0_conv_module1_in_proj_weight_to_fp16, x = x_21_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
            tensor<int32, [4]> var_302_begin_0 = const()[name = tensor<string, []>("op_302_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_302_end_0 = const()[name = tensor<string, []>("op_302_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_302_end_mask_0 = const()[name = tensor<string, []>("op_302_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_302_cast_fp16 = slice_by_index(begin = var_302_begin_0, end = var_302_end_0, end_mask = var_302_end_mask_0, x = x_23_cast_fp16)[name = tensor<string, []>("op_302_cast_fp16")];
            tensor<int32, [4]> var_304_begin_0 = const()[name = tensor<string, []>("op_304_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_304_end_0 = const()[name = tensor<string, []>("op_304_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_304_end_mask_0 = const()[name = tensor<string, []>("op_304_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_304_cast_fp16 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = x_23_cast_fp16)[name = tensor<string, []>("op_304_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_305_cast_fp16 = sigmoid(x = var_304_cast_fp16)[name = tensor<string, []>("op_305_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_25_cast_fp16 = mul(x = var_302_cast_fp16, y = var_305_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_23_cast_fp16 = mul(x = x_25_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<string, []> var_314_pad_type_0 = const()[name = tensor<string, []>("op_314_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_314_pad_0 = const()[name = tensor<string, []>("op_314_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_314_groups_0 = const()[name = tensor<string, []>("op_314_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_314_strides_0 = const()[name = tensor<string, []>("op_314_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_314_dilations_0 = const()[name = tensor<string, []>("op_314_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_0_layers_0_conv_module1_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module1_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73469376)))];
            tensor<fp16, [512]> core_stacks_0_layers_0_conv_module1_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module1_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73484800)))];
            tensor<fp16, [1, 512, 1, 1040]> var_314_cast_fp16 = conv(bias = core_stacks_0_layers_0_conv_module1_dw_a_bias_to_fp16, dilations = var_314_dilations_0, groups = var_314_groups_0, pad = var_314_pad_0, pad_type = var_314_pad_type_0, strides = var_314_strides_0, weight = core_stacks_0_layers_0_conv_module1_dw_a_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("op_314_cast_fp16")];
            tensor<int32, [4]> var_315_begin_0 = const()[name = tensor<string, []>("op_315_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_315_end_0 = const()[name = tensor<string, []>("op_315_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_315_end_mask_0 = const()[name = tensor<string, []>("op_315_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_315_cast_fp16 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = var_314_cast_fp16)[name = tensor<string, []>("op_315_cast_fp16")];
            tensor<string, []> var_321_pad_type_0 = const()[name = tensor<string, []>("op_321_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_321_pad_0 = const()[name = tensor<string, []>("op_321_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_321_groups_0 = const()[name = tensor<string, []>("op_321_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_321_strides_0 = const()[name = tensor<string, []>("op_321_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_321_dilations_0 = const()[name = tensor<string, []>("op_321_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_0_layers_0_conv_module1_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module1_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73485888)))];
            tensor<fp16, [1, 512, 1, 1040]> var_321_cast_fp16 = conv(dilations = var_321_dilations_0, groups = var_321_groups_0, pad = var_321_pad_0, pad_type = var_321_pad_type_0, strides = var_321_strides_0, weight = core_stacks_0_layers_0_conv_module1_dw_b_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("op_321_cast_fp16")];
            tensor<int32, [4]> var_325_begin_0 = const()[name = tensor<string, []>("op_325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_325_end_0 = const()[name = tensor<string, []>("op_325_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_325_end_mask_0 = const()[name = tensor<string, []>("op_325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_325_cast_fp16 = slice_by_index(begin = var_325_begin_0, end = var_325_end_0, end_mask = var_325_end_mask_0, x = var_321_cast_fp16)[name = tensor<string, []>("op_325_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_326_cast_fp16 = add(x = var_315_cast_fp16, y = var_325_cast_fp16)[name = tensor<string, []>("op_326_cast_fp16")];
            tensor<string, []> var_332_pad_type_0 = const()[name = tensor<string, []>("op_332_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_332_pad_0 = const()[name = tensor<string, []>("op_332_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_332_groups_0 = const()[name = tensor<string, []>("op_332_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_332_strides_0 = const()[name = tensor<string, []>("op_332_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_332_dilations_0 = const()[name = tensor<string, []>("op_332_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_0_layers_0_conv_module1_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module1_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73501312)))];
            tensor<fp16, [1, 512, 1, 1054]> var_332_cast_fp16 = conv(dilations = var_332_dilations_0, groups = var_332_groups_0, pad = var_332_pad_0, pad_type = var_332_pad_type_0, strides = var_332_strides_0, weight = core_stacks_0_layers_0_conv_module1_dw_c_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("op_332_cast_fp16")];
            tensor<int32, [4]> var_336_begin_0 = const()[name = tensor<string, []>("op_336_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_336_end_0 = const()[name = tensor<string, []>("op_336_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_336_end_mask_0 = const()[name = tensor<string, []>("op_336_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_336_cast_fp16 = slice_by_index(begin = var_336_begin_0, end = var_336_end_0, end_mask = var_336_end_mask_0, x = var_332_cast_fp16)[name = tensor<string, []>("op_336_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_27_cast_fp16 = add(x = var_326_cast_fp16, y = var_336_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
            tensor<fp16, []> var_338_to_fp16 = const()[name = tensor<string, []>("op_338_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_29_cast_fp16 = sub(x = x_27_cast_fp16, y = var_338_to_fp16)[name = tensor<string, []>("x_29_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_340_cast_fp16 = relu(x = x_29_cast_fp16)[name = tensor<string, []>("op_340_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_341_cast_fp16 = abs(x = x_29_cast_fp16)[name = tensor<string, []>("op_341_cast_fp16")];
            tensor<fp16, []> const_5_promoted_to_fp16 = const()[name = tensor<string, []>("const_5_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_342_cast_fp16 = mul(x = var_341_cast_fp16, y = const_5_promoted_to_fp16)[name = tensor<string, []>("op_342_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_343_cast_fp16 = exp(x = var_342_cast_fp16)[name = tensor<string, []>("op_343_cast_fp16")];
            tensor<fp32, []> var_344_epsilon_0 = const()[name = tensor<string, []>("op_344_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_344_cast_fp16 = log(epsilon = var_344_epsilon_0, x = var_343_cast_fp16)[name = tensor<string, []>("op_344_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_345_cast_fp16 = add(x = var_340_cast_fp16, y = var_344_cast_fp16)[name = tensor<string, []>("op_345_cast_fp16")];
            tensor<fp16, []> var_346_to_fp16 = const()[name = tensor<string, []>("op_346_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_347_cast_fp16 = mul(x = x_27_cast_fp16, y = var_346_to_fp16)[name = tensor<string, []>("op_347_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_348_cast_fp16 = sub(x = var_345_cast_fp16, y = var_347_cast_fp16)[name = tensor<string, []>("op_348_cast_fp16")];
            tensor<fp16, []> var_349_to_fp16 = const()[name = tensor<string, []>("op_349_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_25_cast_fp16 = sub(x = var_348_cast_fp16, y = var_349_to_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<string, []> var_357_pad_type_0 = const()[name = tensor<string, []>("op_357_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_357_strides_0 = const()[name = tensor<string, []>("op_357_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_357_pad_0 = const()[name = tensor<string, []>("op_357_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_357_dilations_0 = const()[name = tensor<string, []>("op_357_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_357_groups_0 = const()[name = tensor<string, []>("op_357_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_0_layers_0_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73502400)))];
            tensor<fp16, [512]> core_stacks_0_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74026752)))];
            tensor<fp16, [1, 512, 1, 1024]> var_357_cast_fp16 = conv(bias = core_stacks_0_layers_0_conv_module1_out_proj_bias_to_fp16, dilations = var_357_dilations_0, groups = var_357_groups_0, pad = var_357_pad_0, pad_type = var_357_pad_type_0, strides = var_357_strides_0, weight = core_stacks_0_layers_0_conv_module1_out_proj_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("op_357_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_27_cast_fp16 = add(x = x_21_cast_fp16, y = var_357_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<string, []> x_31_pad_type_0 = const()[name = tensor<string, []>("x_31_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_31_strides_0 = const()[name = tensor<string, []>("x_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_31_pad_0 = const()[name = tensor<string, []>("x_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_31_dilations_0 = const()[name = tensor<string, []>("x_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_31_groups_0 = const()[name = tensor<string, []>("x_31_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_0_layers_0_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(74027840)))];
            tensor<fp16, [1536]> core_stacks_0_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75600768)))];
            tensor<fp16, [1, 1536, 1, 1024]> x_31_cast_fp16 = conv(bias = core_stacks_0_layers_0_feed_forward2_in_proj_bias_to_fp16, dilations = x_31_dilations_0, groups = x_31_groups_0, pad = x_31_pad_0, pad_type = x_31_pad_type_0, strides = x_31_strides_0, weight = core_stacks_0_layers_0_feed_forward2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")];
            tensor<fp16, []> var_368_to_fp16 = const()[name = tensor<string, []>("op_368_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 1024]> x_33_cast_fp16 = sub(x = x_31_cast_fp16, y = var_368_to_fp16)[name = tensor<string, []>("x_33_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_370_cast_fp16 = relu(x = x_33_cast_fp16)[name = tensor<string, []>("op_370_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_371_cast_fp16 = abs(x = x_33_cast_fp16)[name = tensor<string, []>("op_371_cast_fp16")];
            tensor<fp16, []> const_6_promoted_to_fp16 = const()[name = tensor<string, []>("const_6_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_372_cast_fp16 = mul(x = var_371_cast_fp16, y = const_6_promoted_to_fp16)[name = tensor<string, []>("op_372_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_373_cast_fp16 = exp(x = var_372_cast_fp16)[name = tensor<string, []>("op_373_cast_fp16")];
            tensor<fp32, []> var_374_epsilon_0 = const()[name = tensor<string, []>("op_374_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_374_cast_fp16 = log(epsilon = var_374_epsilon_0, x = var_373_cast_fp16)[name = tensor<string, []>("op_374_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_375_cast_fp16 = add(x = var_370_cast_fp16, y = var_374_cast_fp16)[name = tensor<string, []>("op_375_cast_fp16")];
            tensor<fp16, []> var_376_to_fp16 = const()[name = tensor<string, []>("op_376_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 1024]> var_377_cast_fp16 = mul(x = x_31_cast_fp16, y = var_376_to_fp16)[name = tensor<string, []>("op_377_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_378_cast_fp16 = sub(x = var_375_cast_fp16, y = var_377_cast_fp16)[name = tensor<string, []>("op_378_cast_fp16")];
            tensor<fp16, []> var_379_to_fp16 = const()[name = tensor<string, []>("op_379_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 1024]> input_29_cast_fp16 = sub(x = var_378_cast_fp16, y = var_379_to_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<string, []> var_387_pad_type_0 = const()[name = tensor<string, []>("op_387_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_387_strides_0 = const()[name = tensor<string, []>("op_387_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_387_pad_0 = const()[name = tensor<string, []>("op_387_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_387_dilations_0 = const()[name = tensor<string, []>("op_387_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_387_groups_0 = const()[name = tensor<string, []>("op_387_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_0_layers_0_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75603904)))];
            tensor<fp16, [512]> core_stacks_0_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77176832)))];
            tensor<fp16, [1, 512, 1, 1024]> var_387_cast_fp16 = conv(bias = core_stacks_0_layers_0_feed_forward2_out_proj_bias_to_fp16, dilations = var_387_dilations_0, groups = var_387_groups_0, pad = var_387_pad_0, pad_type = var_387_pad_type_0, strides = var_387_strides_0, weight = core_stacks_0_layers_0_feed_forward2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("op_387_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> src_1_cast_fp16 = add(x = input_27_cast_fp16, y = var_387_cast_fp16)[name = tensor<string, []>("src_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_390_cast_fp16 = sub(x = src_1_cast_fp16, y = input_11_cast_fp16)[name = tensor<string, []>("op_390_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_0_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77177920)))];
            tensor<fp16, [1, 512, 1, 1024]> var_391_cast_fp16 = mul(x = var_390_cast_fp16, y = core_stacks_0_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_391_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_35_cast_fp16 = add(x = input_11_cast_fp16, y = var_391_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
            tensor<string, []> var_402_pad_type_0 = const()[name = tensor<string, []>("op_402_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_402_strides_0 = const()[name = tensor<string, []>("op_402_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_402_pad_0 = const()[name = tensor<string, []>("op_402_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_402_dilations_0 = const()[name = tensor<string, []>("op_402_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_402_groups_0 = const()[name = tensor<string, []>("op_402_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_0_layers_0_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77179008)))];
            tensor<fp16, [48]> core_stacks_0_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77228224)))];
            tensor<fp16, [1, 48, 1, 1024]> var_402_cast_fp16 = conv(bias = core_stacks_0_layers_0_self_attn2_in_proj_bias_to_fp16, dilations = var_402_dilations_0, groups = var_402_groups_0, pad = var_402_pad_0, pad_type = var_402_pad_type_0, strides = var_402_strides_0, weight = core_stacks_0_layers_0_self_attn2_in_proj_weight_to_fp16, x = x_35_cast_fp16)[name = tensor<string, []>("op_402_cast_fp16")];
            tensor<int32, [4]> var_403 = const()[name = tensor<string, []>("op_403"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_404_cast_fp16 = reshape(shape = var_403, x = var_402_cast_fp16)[name = tensor<string, []>("op_404_cast_fp16")];
            tensor<int32, [4]> transpose_7_perm_0 = const()[name = tensor<string, []>("transpose_7_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_35 = const()[name = tensor<string, []>("concat_35"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = var_404_cast_fp16)[name = tensor<string, []>("transpose_310")];
            tensor<fp16, [4, 1024, 12]> reshape_10_cast_fp16 = reshape(shape = concat_35, x = transpose_7_cast_fp16)[name = tensor<string, []>("reshape_10_cast_fp16")];
            tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = reshape_6_cast_fp16, y = reshape_10_cast_fp16)[name = tensor<string, []>("matmul_3_cast_fp16")];
            tensor<int32, [4]> concat_39 = const()[name = tensor<string, []>("concat_39"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_11_cast_fp16 = reshape(shape = concat_39, x = matmul_3_cast_fp16)[name = tensor<string, []>("reshape_11_cast_fp16")];
            tensor<int32, [4]> out_3_perm_0 = const()[name = tensor<string, []>("out_3_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_407 = const()[name = tensor<string, []>("op_407"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_3_cast_fp16 = transpose(perm = out_3_perm_0, x = reshape_11_cast_fp16)[name = tensor<string, []>("transpose_309")];
            tensor<fp16, [1, 48, 1, 1024]> input_31_cast_fp16 = reshape(shape = var_407, x = out_3_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<string, []> var_415_pad_type_0 = const()[name = tensor<string, []>("op_415_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_415_strides_0 = const()[name = tensor<string, []>("op_415_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_415_pad_0 = const()[name = tensor<string, []>("op_415_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_415_dilations_0 = const()[name = tensor<string, []>("op_415_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_415_groups_0 = const()[name = tensor<string, []>("op_415_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_0_layers_0_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77228416)))];
            tensor<fp16, [512]> core_stacks_0_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77277632)))];
            tensor<fp16, [1, 512, 1, 1024]> var_415_cast_fp16 = conv(bias = core_stacks_0_layers_0_self_attn2_out_proj_bias_to_fp16, dilations = var_415_dilations_0, groups = var_415_groups_0, pad = var_415_pad_0, pad_type = var_415_pad_type_0, strides = var_415_strides_0, weight = core_stacks_0_layers_0_self_attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("op_415_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = var_415_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = time_emb_1_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
            tensor<string, []> x_41_pad_type_0 = const()[name = tensor<string, []>("x_41_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_41_strides_0 = const()[name = tensor<string, []>("x_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_41_pad_0 = const()[name = tensor<string, []>("x_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_41_dilations_0 = const()[name = tensor<string, []>("x_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_41_groups_0 = const()[name = tensor<string, []>("x_41_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_0_layers_0_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77278720)))];
            tensor<fp16, [1024]> core_stacks_0_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78327360)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_41_cast_fp16 = conv(bias = core_stacks_0_layers_0_conv_module2_in_proj_bias_to_fp16, dilations = x_41_dilations_0, groups = x_41_groups_0, pad = x_41_pad_0, pad_type = x_41_pad_type_0, strides = x_41_strides_0, weight = core_stacks_0_layers_0_conv_module2_in_proj_weight_to_fp16, x = x_39_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")];
            tensor<int32, [4]> var_433_begin_0 = const()[name = tensor<string, []>("op_433_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_433_end_0 = const()[name = tensor<string, []>("op_433_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_433_end_mask_0 = const()[name = tensor<string, []>("op_433_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_433_cast_fp16 = slice_by_index(begin = var_433_begin_0, end = var_433_end_0, end_mask = var_433_end_mask_0, x = x_41_cast_fp16)[name = tensor<string, []>("op_433_cast_fp16")];
            tensor<int32, [4]> var_435_begin_0 = const()[name = tensor<string, []>("op_435_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_435_end_0 = const()[name = tensor<string, []>("op_435_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_435_end_mask_0 = const()[name = tensor<string, []>("op_435_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_435_cast_fp16 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_41_cast_fp16)[name = tensor<string, []>("op_435_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_436_cast_fp16 = sigmoid(x = var_435_cast_fp16)[name = tensor<string, []>("op_436_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_43_cast_fp16 = mul(x = var_433_cast_fp16, y = var_436_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_33_cast_fp16 = mul(x = x_43_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
            tensor<string, []> var_445_pad_type_0 = const()[name = tensor<string, []>("op_445_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_445_pad_0 = const()[name = tensor<string, []>("op_445_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_445_groups_0 = const()[name = tensor<string, []>("op_445_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_445_strides_0 = const()[name = tensor<string, []>("op_445_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_445_dilations_0 = const()[name = tensor<string, []>("op_445_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_0_layers_0_conv_module2_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module2_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78329472)))];
            tensor<fp16, [512]> core_stacks_0_layers_0_conv_module2_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module2_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78344896)))];
            tensor<fp16, [1, 512, 1, 1040]> var_445_cast_fp16 = conv(bias = core_stacks_0_layers_0_conv_module2_dw_a_bias_to_fp16, dilations = var_445_dilations_0, groups = var_445_groups_0, pad = var_445_pad_0, pad_type = var_445_pad_type_0, strides = var_445_strides_0, weight = core_stacks_0_layers_0_conv_module2_dw_a_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("op_445_cast_fp16")];
            tensor<int32, [4]> var_446_begin_0 = const()[name = tensor<string, []>("op_446_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_446_end_0 = const()[name = tensor<string, []>("op_446_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_446_end_mask_0 = const()[name = tensor<string, []>("op_446_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_446_cast_fp16 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = var_445_cast_fp16)[name = tensor<string, []>("op_446_cast_fp16")];
            tensor<string, []> var_452_pad_type_0 = const()[name = tensor<string, []>("op_452_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_452_pad_0 = const()[name = tensor<string, []>("op_452_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_452_groups_0 = const()[name = tensor<string, []>("op_452_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_452_strides_0 = const()[name = tensor<string, []>("op_452_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_452_dilations_0 = const()[name = tensor<string, []>("op_452_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_0_layers_0_conv_module2_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module2_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78345984)))];
            tensor<fp16, [1, 512, 1, 1040]> var_452_cast_fp16 = conv(dilations = var_452_dilations_0, groups = var_452_groups_0, pad = var_452_pad_0, pad_type = var_452_pad_type_0, strides = var_452_strides_0, weight = core_stacks_0_layers_0_conv_module2_dw_b_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("op_452_cast_fp16")];
            tensor<int32, [4]> var_456_begin_0 = const()[name = tensor<string, []>("op_456_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_456_end_0 = const()[name = tensor<string, []>("op_456_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_456_end_mask_0 = const()[name = tensor<string, []>("op_456_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_456_cast_fp16 = slice_by_index(begin = var_456_begin_0, end = var_456_end_0, end_mask = var_456_end_mask_0, x = var_452_cast_fp16)[name = tensor<string, []>("op_456_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_457_cast_fp16 = add(x = var_446_cast_fp16, y = var_456_cast_fp16)[name = tensor<string, []>("op_457_cast_fp16")];
            tensor<string, []> var_463_pad_type_0 = const()[name = tensor<string, []>("op_463_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_463_pad_0 = const()[name = tensor<string, []>("op_463_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_463_groups_0 = const()[name = tensor<string, []>("op_463_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_463_strides_0 = const()[name = tensor<string, []>("op_463_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_463_dilations_0 = const()[name = tensor<string, []>("op_463_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_0_layers_0_conv_module2_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module2_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78361408)))];
            tensor<fp16, [1, 512, 1, 1054]> var_463_cast_fp16 = conv(dilations = var_463_dilations_0, groups = var_463_groups_0, pad = var_463_pad_0, pad_type = var_463_pad_type_0, strides = var_463_strides_0, weight = core_stacks_0_layers_0_conv_module2_dw_c_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("op_463_cast_fp16")];
            tensor<int32, [4]> var_467_begin_0 = const()[name = tensor<string, []>("op_467_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_467_end_0 = const()[name = tensor<string, []>("op_467_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_467_end_mask_0 = const()[name = tensor<string, []>("op_467_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_467_cast_fp16 = slice_by_index(begin = var_467_begin_0, end = var_467_end_0, end_mask = var_467_end_mask_0, x = var_463_cast_fp16)[name = tensor<string, []>("op_467_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_45_cast_fp16 = add(x = var_457_cast_fp16, y = var_467_cast_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
            tensor<fp16, []> var_469_to_fp16 = const()[name = tensor<string, []>("op_469_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_47_cast_fp16 = sub(x = x_45_cast_fp16, y = var_469_to_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_471_cast_fp16 = relu(x = x_47_cast_fp16)[name = tensor<string, []>("op_471_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_472_cast_fp16 = abs(x = x_47_cast_fp16)[name = tensor<string, []>("op_472_cast_fp16")];
            tensor<fp16, []> const_9_promoted_to_fp16 = const()[name = tensor<string, []>("const_9_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_473_cast_fp16 = mul(x = var_472_cast_fp16, y = const_9_promoted_to_fp16)[name = tensor<string, []>("op_473_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_474_cast_fp16 = exp(x = var_473_cast_fp16)[name = tensor<string, []>("op_474_cast_fp16")];
            tensor<fp32, []> var_475_epsilon_0 = const()[name = tensor<string, []>("op_475_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_475_cast_fp16 = log(epsilon = var_475_epsilon_0, x = var_474_cast_fp16)[name = tensor<string, []>("op_475_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_476_cast_fp16 = add(x = var_471_cast_fp16, y = var_475_cast_fp16)[name = tensor<string, []>("op_476_cast_fp16")];
            tensor<fp16, []> var_477_to_fp16 = const()[name = tensor<string, []>("op_477_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_478_cast_fp16 = mul(x = x_45_cast_fp16, y = var_477_to_fp16)[name = tensor<string, []>("op_478_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_479_cast_fp16 = sub(x = var_476_cast_fp16, y = var_478_cast_fp16)[name = tensor<string, []>("op_479_cast_fp16")];
            tensor<fp16, []> var_480_to_fp16 = const()[name = tensor<string, []>("op_480_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_35_cast_fp16 = sub(x = var_479_cast_fp16, y = var_480_to_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<string, []> var_488_pad_type_0 = const()[name = tensor<string, []>("op_488_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_488_strides_0 = const()[name = tensor<string, []>("op_488_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_488_pad_0 = const()[name = tensor<string, []>("op_488_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_488_dilations_0 = const()[name = tensor<string, []>("op_488_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_488_groups_0 = const()[name = tensor<string, []>("op_488_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_0_layers_0_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78362496)))];
            tensor<fp16, [512]> core_stacks_0_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78886848)))];
            tensor<fp16, [1, 512, 1, 1024]> var_488_cast_fp16 = conv(bias = core_stacks_0_layers_0_conv_module2_out_proj_bias_to_fp16, dilations = var_488_dilations_0, groups = var_488_groups_0, pad = var_488_pad_0, pad_type = var_488_pad_type_0, strides = var_488_strides_0, weight = core_stacks_0_layers_0_conv_module2_out_proj_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("op_488_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_37_cast_fp16 = add(x = x_39_cast_fp16, y = var_488_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<string, []> x_49_pad_type_0 = const()[name = tensor<string, []>("x_49_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_49_strides_0 = const()[name = tensor<string, []>("x_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_49_pad_0 = const()[name = tensor<string, []>("x_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_49_dilations_0 = const()[name = tensor<string, []>("x_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_49_groups_0 = const()[name = tensor<string, []>("x_49_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_0_layers_0_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78887936)))];
            tensor<fp16, [1920]> core_stacks_0_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80854080)))];
            tensor<fp16, [1, 1920, 1, 1024]> x_49_cast_fp16 = conv(bias = core_stacks_0_layers_0_feed_forward3_in_proj_bias_to_fp16, dilations = x_49_dilations_0, groups = x_49_groups_0, pad = x_49_pad_0, pad_type = x_49_pad_type_0, strides = x_49_strides_0, weight = core_stacks_0_layers_0_feed_forward3_in_proj_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")];
            tensor<fp16, []> var_499_to_fp16 = const()[name = tensor<string, []>("op_499_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 1024]> x_51_cast_fp16 = sub(x = x_49_cast_fp16, y = var_499_to_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_501_cast_fp16 = relu(x = x_51_cast_fp16)[name = tensor<string, []>("op_501_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_502_cast_fp16 = abs(x = x_51_cast_fp16)[name = tensor<string, []>("op_502_cast_fp16")];
            tensor<fp16, []> const_10_promoted_to_fp16 = const()[name = tensor<string, []>("const_10_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_503_cast_fp16 = mul(x = var_502_cast_fp16, y = const_10_promoted_to_fp16)[name = tensor<string, []>("op_503_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_504_cast_fp16 = exp(x = var_503_cast_fp16)[name = tensor<string, []>("op_504_cast_fp16")];
            tensor<fp32, []> var_505_epsilon_0 = const()[name = tensor<string, []>("op_505_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_505_cast_fp16 = log(epsilon = var_505_epsilon_0, x = var_504_cast_fp16)[name = tensor<string, []>("op_505_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_506_cast_fp16 = add(x = var_501_cast_fp16, y = var_505_cast_fp16)[name = tensor<string, []>("op_506_cast_fp16")];
            tensor<fp16, []> var_507_to_fp16 = const()[name = tensor<string, []>("op_507_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 1024]> var_508_cast_fp16 = mul(x = x_49_cast_fp16, y = var_507_to_fp16)[name = tensor<string, []>("op_508_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_509_cast_fp16 = sub(x = var_506_cast_fp16, y = var_508_cast_fp16)[name = tensor<string, []>("op_509_cast_fp16")];
            tensor<fp16, []> var_510_to_fp16 = const()[name = tensor<string, []>("op_510_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 1024]> input_39_cast_fp16 = sub(x = var_509_cast_fp16, y = var_510_to_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<string, []> var_518_pad_type_0 = const()[name = tensor<string, []>("op_518_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_518_strides_0 = const()[name = tensor<string, []>("op_518_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_518_pad_0 = const()[name = tensor<string, []>("op_518_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_518_dilations_0 = const()[name = tensor<string, []>("op_518_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_518_groups_0 = const()[name = tensor<string, []>("op_518_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_0_layers_0_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80857984)))];
            tensor<fp16, [512]> core_stacks_0_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82824128)))];
            tensor<fp16, [1, 512, 1, 1024]> var_518_cast_fp16 = conv(bias = core_stacks_0_layers_0_feed_forward3_out_proj_bias_to_fp16, dilations = var_518_dilations_0, groups = var_518_groups_0, pad = var_518_pad_0, pad_type = var_518_pad_type_0, strides = var_518_strides_0, weight = core_stacks_0_layers_0_feed_forward3_out_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("op_518_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_53_cast_fp16 = add(x = input_37_cast_fp16, y = var_518_cast_fp16)[name = tensor<string, []>("x_53_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_0_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82825216)))];
            tensor<fp16, [1, 512, 1, 1024]> d_1_cast_fp16 = sub(x = x_53_cast_fp16, y = core_stacks_0_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("d_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_523_cast_fp16 = mul(x = d_1_cast_fp16, y = d_1_cast_fp16)[name = tensor<string, []>("op_523_cast_fp16")];
            tensor<int32, [1]> var_525_axes_0 = const()[name = tensor<string, []>("op_525_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_525_keep_dims_0 = const()[name = tensor<string, []>("op_525_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 1024]> var_525_cast_fp16 = reduce_mean(axes = var_525_axes_0, keep_dims = var_525_keep_dims_0, x = var_523_cast_fp16)[name = tensor<string, []>("op_525_cast_fp16")];
            tensor<fp32, []> scales_1_epsilon_0 = const()[name = tensor<string, []>("scales_1_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1024]> scales_1_cast_fp16 = rsqrt(epsilon = scales_1_epsilon_0, x = var_525_cast_fp16)[name = tensor<string, []>("scales_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_527_cast_fp16 = mul(x = x_53_cast_fp16, y = scales_1_cast_fp16)[name = tensor<string, []>("op_527_cast_fp16")];
            tensor<fp16, []> core_stacks_0_layers_0_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.51p+0)];
            tensor<fp16, [1, 512, 1, 1024]> src_3_cast_fp16 = mul(x = var_527_cast_fp16, y = core_stacks_0_layers_0_norm_scale_to_fp16)[name = tensor<string, []>("src_3_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_530_cast_fp16 = sub(x = src_3_cast_fp16, y = input_11_cast_fp16)[name = tensor<string, []>("op_530_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_0_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82826304)))];
            tensor<fp16, [1, 512, 1, 1024]> var_531_cast_fp16 = mul(x = var_530_cast_fp16, y = core_stacks_0_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_531_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_41_cast_fp16 = add(x = input_11_cast_fp16, y = var_531_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
            tensor<string, []> proj_3_pad_type_0 = const()[name = tensor<string, []>("proj_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_3_strides_0 = const()[name = tensor<string, []>("proj_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_3_pad_0 = const()[name = tensor<string, []>("proj_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_3_dilations_0 = const()[name = tensor<string, []>("proj_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_3_groups_0 = const()[name = tensor<string, []>("proj_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_0_layers_1_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82827392)))];
            tensor<fp16, [384]> core_stacks_0_layers_1_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83220672)))];
            tensor<fp16, [1, 384, 1, 1024]> proj_3_cast_fp16 = conv(bias = core_stacks_0_layers_1_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_3_dilations_0, groups = proj_3_groups_0, pad = proj_3_pad_0, pad_type = proj_3_pad_type_0, strides = proj_3_strides_0, weight = core_stacks_0_layers_1_self_attn_weights_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("proj_3_cast_fp16")];
            tensor<int32, [4]> var_554_begin_0 = const()[name = tensor<string, []>("op_554_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_554_end_0 = const()[name = tensor<string, []>("op_554_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1024])];
            tensor<bool, [4]> var_554_end_mask_0 = const()[name = tensor<string, []>("op_554_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_554_cast_fp16 = slice_by_index(begin = var_554_begin_0, end = var_554_end_0, end_mask = var_554_end_mask_0, x = proj_3_cast_fp16)[name = tensor<string, []>("op_554_cast_fp16")];
            tensor<int32, [4]> var_555 = const()[name = tensor<string, []>("op_555"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_556_cast_fp16 = reshape(shape = var_555, x = var_554_cast_fp16)[name = tensor<string, []>("op_556_cast_fp16")];
            tensor<int32, [4]> var_558_begin_0 = const()[name = tensor<string, []>("op_558_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_558_end_0 = const()[name = tensor<string, []>("op_558_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1024])];
            tensor<bool, [4]> var_558_end_mask_0 = const()[name = tensor<string, []>("op_558_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_558_cast_fp16 = slice_by_index(begin = var_558_begin_0, end = var_558_end_0, end_mask = var_558_end_mask_0, x = proj_3_cast_fp16)[name = tensor<string, []>("op_558_cast_fp16")];
            tensor<int32, [4]> var_559 = const()[name = tensor<string, []>("op_559"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_560_cast_fp16 = reshape(shape = var_559, x = var_558_cast_fp16)[name = tensor<string, []>("op_560_cast_fp16")];
            tensor<int32, [4]> var_562_begin_0 = const()[name = tensor<string, []>("op_562_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_562_end_0 = const()[name = tensor<string, []>("op_562_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_562_end_mask_0 = const()[name = tensor<string, []>("op_562_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_562_cast_fp16 = slice_by_index(begin = var_562_begin_0, end = var_562_end_0, end_mask = var_562_end_mask_0, x = proj_3_cast_fp16)[name = tensor<string, []>("op_562_cast_fp16")];
            tensor<int32, [4]> var_563 = const()[name = tensor<string, []>("op_563"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> pt_3_cast_fp16 = reshape(shape = var_563, x = var_562_cast_fp16)[name = tensor<string, []>("pt_3_cast_fp16")];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = tensor<string, []>("transpose_8_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_44 = const()[name = tensor<string, []>("concat_44"), val = tensor<int32, [3]>([4, 1024, 32])];
            tensor<fp16, [4, 1, 1024, 32]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = var_556_cast_fp16)[name = tensor<string, []>("transpose_308")];
            tensor<fp16, [4, 1024, 32]> reshape_12_cast_fp16 = reshape(shape = concat_44, x = transpose_8_cast_fp16)[name = tensor<string, []>("reshape_12_cast_fp16")];
            tensor<int32, [4]> transpose_9_perm_0 = const()[name = tensor<string, []>("transpose_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_45 = const()[name = tensor<string, []>("concat_45"), val = tensor<int32, [3]>([4, 32, 1024])];
            tensor<fp16, [4, 1, 32, 1024]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = var_560_cast_fp16)[name = tensor<string, []>("transpose_307")];
            tensor<fp16, [4, 32, 1024]> reshape_13_cast_fp16 = reshape(shape = concat_45, x = transpose_9_cast_fp16)[name = tensor<string, []>("reshape_13_cast_fp16")];
            tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 1024]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = reshape_12_cast_fp16, y = reshape_13_cast_fp16)[name = tensor<string, []>("matmul_4_cast_fp16")];
            tensor<int32, [4]> concat_49 = const()[name = tensor<string, []>("concat_49"), val = tensor<int32, [4]>([4, 1, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> reshape_14_cast_fp16 = reshape(shape = concat_49, x = matmul_4_cast_fp16)[name = tensor<string, []>("reshape_14_cast_fp16")];
            tensor<int32, [4]> attn_scores_7_perm_0 = const()[name = tensor<string, []>("attn_scores_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_567 = const()[name = tensor<string, []>("op_567"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 1024, 1]> var_568_cast_fp16 = transpose(perm = var_567, x = pt_3_cast_fp16)[name = tensor<string, []>("transpose_305")];
            tensor<fp16, [4, 32, 1024, 1024]> var_569_cast_fp16 = mul(x = var_568_cast_fp16, y = core_pos_basis_1024_to_fp16)[name = tensor<string, []>("op_569_cast_fp16")];
            tensor<int32, [1]> pos_scores_3_axes_0 = const()[name = tensor<string, []>("pos_scores_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_3_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_3_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 1024, 1024]> pos_scores_3_cast_fp16 = reduce_sum(axes = pos_scores_3_axes_0, keep_dims = pos_scores_3_keep_dims_0, x = var_569_cast_fp16)[name = tensor<string, []>("pos_scores_3_cast_fp16")];
            tensor<int32, [4]> var_572 = const()[name = tensor<string, []>("op_572"), val = tensor<int32, [4]>([4, 1024, 1, 1024])];
            tensor<fp16, [4, 1024, 1, 1024]> var_573_cast_fp16 = reshape(shape = var_572, x = pos_scores_3_cast_fp16)[name = tensor<string, []>("op_573_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_7_cast_fp16 = transpose(perm = attn_scores_7_perm_0, x = reshape_14_cast_fp16)[name = tensor<string, []>("transpose_306")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_9_cast_fp16 = add(x = attn_scores_7_cast_fp16, y = var_573_cast_fp16)[name = tensor<string, []>("attn_scores_9_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_11_cast_fp16 = add(x = attn_scores_9_cast_fp16, y = attn_bias_1_cast_fp16)[name = tensor<string, []>("attn_scores_11_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_weights_3_cast_fp16 = softmax(axis = var_41, x = attn_scores_11_cast_fp16)[name = tensor<string, []>("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_43_cast_fp16 = add(x = input_41_cast_fp16, y = time_emb_1_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<string, []> x_55_pad_type_0 = const()[name = tensor<string, []>("x_55_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_55_strides_0 = const()[name = tensor<string, []>("x_55_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_55_pad_0 = const()[name = tensor<string, []>("x_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_55_dilations_0 = const()[name = tensor<string, []>("x_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_55_groups_0 = const()[name = tensor<string, []>("x_55_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_0_layers_1_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83221504)))];
            tensor<fp16, [1152]> core_stacks_0_layers_1_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84401216)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_55_cast_fp16 = conv(bias = core_stacks_0_layers_1_feed_forward1_in_proj_bias_to_fp16, dilations = x_55_dilations_0, groups = x_55_groups_0, pad = x_55_pad_0, pad_type = x_55_pad_type_0, strides = x_55_strides_0, weight = core_stacks_0_layers_1_feed_forward1_in_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("x_55_cast_fp16")];
            tensor<fp16, []> var_587_to_fp16 = const()[name = tensor<string, []>("op_587_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 1024]> x_57_cast_fp16 = sub(x = x_55_cast_fp16, y = var_587_to_fp16)[name = tensor<string, []>("x_57_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_589_cast_fp16 = relu(x = x_57_cast_fp16)[name = tensor<string, []>("op_589_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_590_cast_fp16 = abs(x = x_57_cast_fp16)[name = tensor<string, []>("op_590_cast_fp16")];
            tensor<fp16, []> const_11_promoted_to_fp16 = const()[name = tensor<string, []>("const_11_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_591_cast_fp16 = mul(x = var_590_cast_fp16, y = const_11_promoted_to_fp16)[name = tensor<string, []>("op_591_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_592_cast_fp16 = exp(x = var_591_cast_fp16)[name = tensor<string, []>("op_592_cast_fp16")];
            tensor<fp32, []> var_593_epsilon_0 = const()[name = tensor<string, []>("op_593_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_593_cast_fp16 = log(epsilon = var_593_epsilon_0, x = var_592_cast_fp16)[name = tensor<string, []>("op_593_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_594_cast_fp16 = add(x = var_589_cast_fp16, y = var_593_cast_fp16)[name = tensor<string, []>("op_594_cast_fp16")];
            tensor<fp16, []> var_595_to_fp16 = const()[name = tensor<string, []>("op_595_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 1024]> var_596_cast_fp16 = mul(x = x_55_cast_fp16, y = var_595_to_fp16)[name = tensor<string, []>("op_596_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_597_cast_fp16 = sub(x = var_594_cast_fp16, y = var_596_cast_fp16)[name = tensor<string, []>("op_597_cast_fp16")];
            tensor<fp16, []> var_598_to_fp16 = const()[name = tensor<string, []>("op_598_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 1024]> input_45_cast_fp16 = sub(x = var_597_cast_fp16, y = var_598_to_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<string, []> var_606_pad_type_0 = const()[name = tensor<string, []>("op_606_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_606_strides_0 = const()[name = tensor<string, []>("op_606_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_606_pad_0 = const()[name = tensor<string, []>("op_606_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_606_dilations_0 = const()[name = tensor<string, []>("op_606_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_606_groups_0 = const()[name = tensor<string, []>("op_606_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_0_layers_1_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84403584)))];
            tensor<fp16, [512]> core_stacks_0_layers_1_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85583296)))];
            tensor<fp16, [1, 512, 1, 1024]> var_606_cast_fp16 = conv(bias = core_stacks_0_layers_1_feed_forward1_out_proj_bias_to_fp16, dilations = var_606_dilations_0, groups = var_606_groups_0, pad = var_606_pad_0, pad_type = var_606_pad_type_0, strides = var_606_strides_0, weight = core_stacks_0_layers_1_feed_forward1_out_proj_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("op_606_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_47_cast_fp16 = add(x = input_43_cast_fp16, y = var_606_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<int32, [4]> var_608_begin_0 = const()[name = tensor<string, []>("op_608_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_608_end_0 = const()[name = tensor<string, []>("op_608_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_608_end_mask_0 = const()[name = tensor<string, []>("op_608_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> var_608_cast_fp16 = slice_by_index(begin = var_608_begin_0, end = var_608_end_0, end_mask = var_608_end_mask_0, x = attn_weights_3_cast_fp16)[name = tensor<string, []>("op_608_cast_fp16")];
            tensor<string, []> x_59_pad_type_0 = const()[name = tensor<string, []>("x_59_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_59_strides_0 = const()[name = tensor<string, []>("x_59_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_59_pad_0 = const()[name = tensor<string, []>("x_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_59_dilations_0 = const()[name = tensor<string, []>("x_59_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_59_groups_0 = const()[name = tensor<string, []>("x_59_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_0_layers_1_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85584384)))];
            tensor<fp16, [1152]> core_stacks_0_layers_1_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86764096)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_59_cast_fp16 = conv(bias = core_stacks_0_layers_1_nonlin_attention_in_proj_bias_to_fp16, dilations = x_59_dilations_0, groups = x_59_groups_0, pad = x_59_pad_0, pad_type = x_59_pad_type_0, strides = x_59_strides_0, weight = core_stacks_0_layers_1_nonlin_attention_in_proj_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
            tensor<int32, [4]> var_619_begin_0 = const()[name = tensor<string, []>("op_619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_619_end_0 = const()[name = tensor<string, []>("op_619_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_619_end_mask_0 = const()[name = tensor<string, []>("op_619_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_619_cast_fp16 = slice_by_index(begin = var_619_begin_0, end = var_619_end_0, end_mask = var_619_end_mask_0, x = x_59_cast_fp16)[name = tensor<string, []>("op_619_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> s_gate_3_cast_fp16 = tanh(x = var_619_cast_fp16)[name = tensor<string, []>("s_gate_3_cast_fp16")];
            tensor<int32, [4]> var_622_begin_0 = const()[name = tensor<string, []>("op_622_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_622_end_0 = const()[name = tensor<string, []>("op_622_end_0"), val = tensor<int32, [4]>([1, 768, 1, 1024])];
            tensor<bool, [4]> var_622_end_mask_0 = const()[name = tensor<string, []>("op_622_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_622_cast_fp16 = slice_by_index(begin = var_622_begin_0, end = var_622_end_0, end_mask = var_622_end_mask_0, x = x_59_cast_fp16)[name = tensor<string, []>("op_622_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> var_623_cast_fp16 = mul(x = var_622_cast_fp16, y = s_gate_3_cast_fp16)[name = tensor<string, []>("op_623_cast_fp16")];
            tensor<int32, [4]> y_3_begin_0 = const()[name = tensor<string, []>("y_3_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_3_end_0 = const()[name = tensor<string, []>("y_3_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 1024])];
            tensor<bool, [4]> y_3_end_mask_0 = const()[name = tensor<string, []>("y_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> y_3_cast_fp16 = slice_by_index(begin = y_3_begin_0, end = y_3_end_0, end_mask = y_3_end_mask_0, x = x_59_cast_fp16)[name = tensor<string, []>("y_3_cast_fp16")];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = tensor<string, []>("transpose_10_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_54 = const()[name = tensor<string, []>("concat_54"), val = tensor<int32, [3]>([1, 1024, 1024])];
            tensor<fp16, [1, 1, 1024, 1024]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = var_608_cast_fp16)[name = tensor<string, []>("transpose_304")];
            tensor<fp16, [1, 1024, 1024]> reshape_15_cast_fp16 = reshape(shape = concat_54, x = transpose_10_cast_fp16)[name = tensor<string, []>("reshape_15_cast_fp16")];
            tensor<int32, [4]> transpose_11_perm_0 = const()[name = tensor<string, []>("transpose_11_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_55 = const()[name = tensor<string, []>("concat_55"), val = tensor<int32, [3]>([1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = var_623_cast_fp16)[name = tensor<string, []>("transpose_303")];
            tensor<fp16, [1, 1024, 384]> reshape_16_cast_fp16 = reshape(shape = concat_55, x = transpose_11_cast_fp16)[name = tensor<string, []>("reshape_16_cast_fp16")];
            tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1024, 384]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = reshape_15_cast_fp16, y = reshape_16_cast_fp16)[name = tensor<string, []>("matmul_5_cast_fp16")];
            tensor<int32, [4]> concat_59 = const()[name = tensor<string, []>("concat_59"), val = tensor<int32, [4]>([1, 1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> reshape_17_cast_fp16 = reshape(shape = concat_59, x = matmul_5_cast_fp16)[name = tensor<string, []>("reshape_17_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = tensor<string, []>("v_3_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 1024]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = reshape_17_cast_fp16)[name = tensor<string, []>("transpose_302")];
            tensor<fp16, [1, 384, 1, 1024]> input_49_cast_fp16 = mul(x = v_3_cast_fp16, y = y_3_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<string, []> var_635_pad_type_0 = const()[name = tensor<string, []>("op_635_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_635_strides_0 = const()[name = tensor<string, []>("op_635_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_635_pad_0 = const()[name = tensor<string, []>("op_635_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_635_dilations_0 = const()[name = tensor<string, []>("op_635_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_635_groups_0 = const()[name = tensor<string, []>("op_635_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_0_layers_1_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86766464)))];
            tensor<fp16, [512]> core_stacks_0_layers_1_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87159744)))];
            tensor<fp16, [1, 512, 1, 1024]> var_635_cast_fp16 = conv(bias = core_stacks_0_layers_1_nonlin_attention_out_proj_bias_to_fp16, dilations = var_635_dilations_0, groups = var_635_groups_0, pad = var_635_pad_0, pad_type = var_635_pad_type_0, strides = var_635_strides_0, weight = core_stacks_0_layers_1_nonlin_attention_out_proj_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("op_635_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_61_cast_fp16 = add(x = input_47_cast_fp16, y = var_635_cast_fp16)[name = tensor<string, []>("x_61_cast_fp16")];
            tensor<string, []> var_646_pad_type_0 = const()[name = tensor<string, []>("op_646_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_646_strides_0 = const()[name = tensor<string, []>("op_646_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_646_pad_0 = const()[name = tensor<string, []>("op_646_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_646_dilations_0 = const()[name = tensor<string, []>("op_646_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_646_groups_0 = const()[name = tensor<string, []>("op_646_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_0_layers_1_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87160832)))];
            tensor<fp16, [48]> core_stacks_0_layers_1_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87210048)))];
            tensor<fp16, [1, 48, 1, 1024]> var_646_cast_fp16 = conv(bias = core_stacks_0_layers_1_self_attn1_in_proj_bias_to_fp16, dilations = var_646_dilations_0, groups = var_646_groups_0, pad = var_646_pad_0, pad_type = var_646_pad_type_0, strides = var_646_strides_0, weight = core_stacks_0_layers_1_self_attn1_in_proj_weight_to_fp16, x = x_61_cast_fp16)[name = tensor<string, []>("op_646_cast_fp16")];
            tensor<int32, [4]> var_647 = const()[name = tensor<string, []>("op_647"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_648_cast_fp16 = reshape(shape = var_647, x = var_646_cast_fp16)[name = tensor<string, []>("op_648_cast_fp16")];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = tensor<string, []>("transpose_12_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_64 = const()[name = tensor<string, []>("concat_64"), val = tensor<int32, [3]>([4, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = attn_weights_3_cast_fp16)[name = tensor<string, []>("transpose_301")];
            tensor<fp16, [4, 1024, 1024]> reshape_18_cast_fp16 = reshape(shape = concat_64, x = transpose_12_cast_fp16)[name = tensor<string, []>("reshape_18_cast_fp16")];
            tensor<int32, [4]> transpose_13_perm_0 = const()[name = tensor<string, []>("transpose_13_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_65 = const()[name = tensor<string, []>("concat_65"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = var_648_cast_fp16)[name = tensor<string, []>("transpose_300")];
            tensor<fp16, [4, 1024, 12]> reshape_19_cast_fp16 = reshape(shape = concat_65, x = transpose_13_cast_fp16)[name = tensor<string, []>("reshape_19_cast_fp16")];
            tensor<bool, []> matmul_6_transpose_x_0 = const()[name = tensor<string, []>("matmul_6_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_6_transpose_y_0 = const()[name = tensor<string, []>("matmul_6_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = reshape_18_cast_fp16, y = reshape_19_cast_fp16)[name = tensor<string, []>("matmul_6_cast_fp16")];
            tensor<int32, [4]> concat_69 = const()[name = tensor<string, []>("concat_69"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_20_cast_fp16 = reshape(shape = concat_69, x = matmul_6_cast_fp16)[name = tensor<string, []>("reshape_20_cast_fp16")];
            tensor<int32, [4]> out_5_perm_0 = const()[name = tensor<string, []>("out_5_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_651 = const()[name = tensor<string, []>("op_651"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_5_cast_fp16 = transpose(perm = out_5_perm_0, x = reshape_20_cast_fp16)[name = tensor<string, []>("transpose_299")];
            tensor<fp16, [1, 48, 1, 1024]> input_51_cast_fp16 = reshape(shape = var_651, x = out_5_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<string, []> var_659_pad_type_0 = const()[name = tensor<string, []>("op_659_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_659_strides_0 = const()[name = tensor<string, []>("op_659_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_659_pad_0 = const()[name = tensor<string, []>("op_659_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_659_dilations_0 = const()[name = tensor<string, []>("op_659_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_659_groups_0 = const()[name = tensor<string, []>("op_659_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_0_layers_1_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87210240)))];
            tensor<fp16, [512]> core_stacks_0_layers_1_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87259456)))];
            tensor<fp16, [1, 512, 1, 1024]> var_659_cast_fp16 = conv(bias = core_stacks_0_layers_1_self_attn1_out_proj_bias_to_fp16, dilations = var_659_dilations_0, groups = var_659_groups_0, pad = var_659_pad_0, pad_type = var_659_pad_type_0, strides = var_659_strides_0, weight = core_stacks_0_layers_1_self_attn1_out_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("op_659_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = var_659_cast_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = time_emb_1_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
            tensor<string, []> x_67_pad_type_0 = const()[name = tensor<string, []>("x_67_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_67_strides_0 = const()[name = tensor<string, []>("x_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_67_pad_0 = const()[name = tensor<string, []>("x_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_67_dilations_0 = const()[name = tensor<string, []>("x_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_67_groups_0 = const()[name = tensor<string, []>("x_67_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_0_layers_1_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87260544)))];
            tensor<fp16, [1024]> core_stacks_0_layers_1_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88309184)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_67_cast_fp16 = conv(bias = core_stacks_0_layers_1_conv_module1_in_proj_bias_to_fp16, dilations = x_67_dilations_0, groups = x_67_groups_0, pad = x_67_pad_0, pad_type = x_67_pad_type_0, strides = x_67_strides_0, weight = core_stacks_0_layers_1_conv_module1_in_proj_weight_to_fp16, x = x_65_cast_fp16)[name = tensor<string, []>("x_67_cast_fp16")];
            tensor<int32, [4]> var_677_begin_0 = const()[name = tensor<string, []>("op_677_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_677_end_0 = const()[name = tensor<string, []>("op_677_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_677_end_mask_0 = const()[name = tensor<string, []>("op_677_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_677_cast_fp16 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = x_67_cast_fp16)[name = tensor<string, []>("op_677_cast_fp16")];
            tensor<int32, [4]> var_679_begin_0 = const()[name = tensor<string, []>("op_679_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_679_end_0 = const()[name = tensor<string, []>("op_679_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_679_end_mask_0 = const()[name = tensor<string, []>("op_679_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_679_cast_fp16 = slice_by_index(begin = var_679_begin_0, end = var_679_end_0, end_mask = var_679_end_mask_0, x = x_67_cast_fp16)[name = tensor<string, []>("op_679_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_680_cast_fp16 = sigmoid(x = var_679_cast_fp16)[name = tensor<string, []>("op_680_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_69_cast_fp16 = mul(x = var_677_cast_fp16, y = var_680_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_53_cast_fp16 = mul(x = x_69_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<string, []> var_689_pad_type_0 = const()[name = tensor<string, []>("op_689_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_689_pad_0 = const()[name = tensor<string, []>("op_689_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_689_groups_0 = const()[name = tensor<string, []>("op_689_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_689_strides_0 = const()[name = tensor<string, []>("op_689_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_689_dilations_0 = const()[name = tensor<string, []>("op_689_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_0_layers_1_conv_module1_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module1_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88311296)))];
            tensor<fp16, [512]> core_stacks_0_layers_1_conv_module1_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module1_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88326720)))];
            tensor<fp16, [1, 512, 1, 1040]> var_689_cast_fp16 = conv(bias = core_stacks_0_layers_1_conv_module1_dw_a_bias_to_fp16, dilations = var_689_dilations_0, groups = var_689_groups_0, pad = var_689_pad_0, pad_type = var_689_pad_type_0, strides = var_689_strides_0, weight = core_stacks_0_layers_1_conv_module1_dw_a_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("op_689_cast_fp16")];
            tensor<int32, [4]> var_690_begin_0 = const()[name = tensor<string, []>("op_690_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_690_end_0 = const()[name = tensor<string, []>("op_690_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_690_end_mask_0 = const()[name = tensor<string, []>("op_690_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_690_cast_fp16 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = var_689_cast_fp16)[name = tensor<string, []>("op_690_cast_fp16")];
            tensor<string, []> var_696_pad_type_0 = const()[name = tensor<string, []>("op_696_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_696_pad_0 = const()[name = tensor<string, []>("op_696_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_696_groups_0 = const()[name = tensor<string, []>("op_696_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_696_strides_0 = const()[name = tensor<string, []>("op_696_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_696_dilations_0 = const()[name = tensor<string, []>("op_696_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_0_layers_1_conv_module1_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module1_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88327808)))];
            tensor<fp16, [1, 512, 1, 1040]> var_696_cast_fp16 = conv(dilations = var_696_dilations_0, groups = var_696_groups_0, pad = var_696_pad_0, pad_type = var_696_pad_type_0, strides = var_696_strides_0, weight = core_stacks_0_layers_1_conv_module1_dw_b_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("op_696_cast_fp16")];
            tensor<int32, [4]> var_700_begin_0 = const()[name = tensor<string, []>("op_700_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_700_end_0 = const()[name = tensor<string, []>("op_700_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_700_end_mask_0 = const()[name = tensor<string, []>("op_700_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_700_cast_fp16 = slice_by_index(begin = var_700_begin_0, end = var_700_end_0, end_mask = var_700_end_mask_0, x = var_696_cast_fp16)[name = tensor<string, []>("op_700_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_701_cast_fp16 = add(x = var_690_cast_fp16, y = var_700_cast_fp16)[name = tensor<string, []>("op_701_cast_fp16")];
            tensor<string, []> var_707_pad_type_0 = const()[name = tensor<string, []>("op_707_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_707_pad_0 = const()[name = tensor<string, []>("op_707_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_707_groups_0 = const()[name = tensor<string, []>("op_707_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_707_strides_0 = const()[name = tensor<string, []>("op_707_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_707_dilations_0 = const()[name = tensor<string, []>("op_707_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_0_layers_1_conv_module1_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module1_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88343232)))];
            tensor<fp16, [1, 512, 1, 1054]> var_707_cast_fp16 = conv(dilations = var_707_dilations_0, groups = var_707_groups_0, pad = var_707_pad_0, pad_type = var_707_pad_type_0, strides = var_707_strides_0, weight = core_stacks_0_layers_1_conv_module1_dw_c_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("op_707_cast_fp16")];
            tensor<int32, [4]> var_711_begin_0 = const()[name = tensor<string, []>("op_711_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_711_end_0 = const()[name = tensor<string, []>("op_711_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_711_end_mask_0 = const()[name = tensor<string, []>("op_711_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_711_cast_fp16 = slice_by_index(begin = var_711_begin_0, end = var_711_end_0, end_mask = var_711_end_mask_0, x = var_707_cast_fp16)[name = tensor<string, []>("op_711_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_71_cast_fp16 = add(x = var_701_cast_fp16, y = var_711_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
            tensor<fp16, []> var_713_to_fp16 = const()[name = tensor<string, []>("op_713_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_73_cast_fp16 = sub(x = x_71_cast_fp16, y = var_713_to_fp16)[name = tensor<string, []>("x_73_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_715_cast_fp16 = relu(x = x_73_cast_fp16)[name = tensor<string, []>("op_715_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_716_cast_fp16 = abs(x = x_73_cast_fp16)[name = tensor<string, []>("op_716_cast_fp16")];
            tensor<fp16, []> const_14_promoted_to_fp16 = const()[name = tensor<string, []>("const_14_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_717_cast_fp16 = mul(x = var_716_cast_fp16, y = const_14_promoted_to_fp16)[name = tensor<string, []>("op_717_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_718_cast_fp16 = exp(x = var_717_cast_fp16)[name = tensor<string, []>("op_718_cast_fp16")];
            tensor<fp32, []> var_719_epsilon_0 = const()[name = tensor<string, []>("op_719_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_719_cast_fp16 = log(epsilon = var_719_epsilon_0, x = var_718_cast_fp16)[name = tensor<string, []>("op_719_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_720_cast_fp16 = add(x = var_715_cast_fp16, y = var_719_cast_fp16)[name = tensor<string, []>("op_720_cast_fp16")];
            tensor<fp16, []> var_721_to_fp16 = const()[name = tensor<string, []>("op_721_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_722_cast_fp16 = mul(x = x_71_cast_fp16, y = var_721_to_fp16)[name = tensor<string, []>("op_722_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_723_cast_fp16 = sub(x = var_720_cast_fp16, y = var_722_cast_fp16)[name = tensor<string, []>("op_723_cast_fp16")];
            tensor<fp16, []> var_724_to_fp16 = const()[name = tensor<string, []>("op_724_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_55_cast_fp16 = sub(x = var_723_cast_fp16, y = var_724_to_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<string, []> var_732_pad_type_0 = const()[name = tensor<string, []>("op_732_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_732_strides_0 = const()[name = tensor<string, []>("op_732_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_732_pad_0 = const()[name = tensor<string, []>("op_732_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_732_dilations_0 = const()[name = tensor<string, []>("op_732_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_732_groups_0 = const()[name = tensor<string, []>("op_732_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_0_layers_1_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88344320)))];
            tensor<fp16, [512]> core_stacks_0_layers_1_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88868672)))];
            tensor<fp16, [1, 512, 1, 1024]> var_732_cast_fp16 = conv(bias = core_stacks_0_layers_1_conv_module1_out_proj_bias_to_fp16, dilations = var_732_dilations_0, groups = var_732_groups_0, pad = var_732_pad_0, pad_type = var_732_pad_type_0, strides = var_732_strides_0, weight = core_stacks_0_layers_1_conv_module1_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("op_732_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_57_cast_fp16 = add(x = x_65_cast_fp16, y = var_732_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
            tensor<string, []> x_75_pad_type_0 = const()[name = tensor<string, []>("x_75_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_75_strides_0 = const()[name = tensor<string, []>("x_75_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_75_pad_0 = const()[name = tensor<string, []>("x_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_75_dilations_0 = const()[name = tensor<string, []>("x_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_75_groups_0 = const()[name = tensor<string, []>("x_75_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_0_layers_1_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88869760)))];
            tensor<fp16, [1536]> core_stacks_0_layers_1_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90442688)))];
            tensor<fp16, [1, 1536, 1, 1024]> x_75_cast_fp16 = conv(bias = core_stacks_0_layers_1_feed_forward2_in_proj_bias_to_fp16, dilations = x_75_dilations_0, groups = x_75_groups_0, pad = x_75_pad_0, pad_type = x_75_pad_type_0, strides = x_75_strides_0, weight = core_stacks_0_layers_1_feed_forward2_in_proj_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
            tensor<fp16, []> var_743_to_fp16 = const()[name = tensor<string, []>("op_743_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 1024]> x_77_cast_fp16 = sub(x = x_75_cast_fp16, y = var_743_to_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_745_cast_fp16 = relu(x = x_77_cast_fp16)[name = tensor<string, []>("op_745_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_746_cast_fp16 = abs(x = x_77_cast_fp16)[name = tensor<string, []>("op_746_cast_fp16")];
            tensor<fp16, []> const_15_promoted_to_fp16 = const()[name = tensor<string, []>("const_15_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_747_cast_fp16 = mul(x = var_746_cast_fp16, y = const_15_promoted_to_fp16)[name = tensor<string, []>("op_747_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_748_cast_fp16 = exp(x = var_747_cast_fp16)[name = tensor<string, []>("op_748_cast_fp16")];
            tensor<fp32, []> var_749_epsilon_0 = const()[name = tensor<string, []>("op_749_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_749_cast_fp16 = log(epsilon = var_749_epsilon_0, x = var_748_cast_fp16)[name = tensor<string, []>("op_749_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_750_cast_fp16 = add(x = var_745_cast_fp16, y = var_749_cast_fp16)[name = tensor<string, []>("op_750_cast_fp16")];
            tensor<fp16, []> var_751_to_fp16 = const()[name = tensor<string, []>("op_751_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 1024]> var_752_cast_fp16 = mul(x = x_75_cast_fp16, y = var_751_to_fp16)[name = tensor<string, []>("op_752_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_753_cast_fp16 = sub(x = var_750_cast_fp16, y = var_752_cast_fp16)[name = tensor<string, []>("op_753_cast_fp16")];
            tensor<fp16, []> var_754_to_fp16 = const()[name = tensor<string, []>("op_754_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 1024]> input_59_cast_fp16 = sub(x = var_753_cast_fp16, y = var_754_to_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<string, []> var_762_pad_type_0 = const()[name = tensor<string, []>("op_762_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_762_strides_0 = const()[name = tensor<string, []>("op_762_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_762_pad_0 = const()[name = tensor<string, []>("op_762_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_762_dilations_0 = const()[name = tensor<string, []>("op_762_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_762_groups_0 = const()[name = tensor<string, []>("op_762_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_0_layers_1_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90445824)))];
            tensor<fp16, [512]> core_stacks_0_layers_1_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92018752)))];
            tensor<fp16, [1, 512, 1, 1024]> var_762_cast_fp16 = conv(bias = core_stacks_0_layers_1_feed_forward2_out_proj_bias_to_fp16, dilations = var_762_dilations_0, groups = var_762_groups_0, pad = var_762_pad_0, pad_type = var_762_pad_type_0, strides = var_762_strides_0, weight = core_stacks_0_layers_1_feed_forward2_out_proj_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("op_762_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> src_5_cast_fp16 = add(x = input_57_cast_fp16, y = var_762_cast_fp16)[name = tensor<string, []>("src_5_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_765_cast_fp16 = sub(x = src_5_cast_fp16, y = input_41_cast_fp16)[name = tensor<string, []>("op_765_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_0_layers_1_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92019840)))];
            tensor<fp16, [1, 512, 1, 1024]> var_766_cast_fp16 = mul(x = var_765_cast_fp16, y = core_stacks_0_layers_1_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_766_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_79_cast_fp16 = add(x = input_41_cast_fp16, y = var_766_cast_fp16)[name = tensor<string, []>("x_79_cast_fp16")];
            tensor<string, []> var_777_pad_type_0 = const()[name = tensor<string, []>("op_777_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_777_strides_0 = const()[name = tensor<string, []>("op_777_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_777_pad_0 = const()[name = tensor<string, []>("op_777_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_777_dilations_0 = const()[name = tensor<string, []>("op_777_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_777_groups_0 = const()[name = tensor<string, []>("op_777_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_0_layers_1_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92020928)))];
            tensor<fp16, [48]> core_stacks_0_layers_1_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92070144)))];
            tensor<fp16, [1, 48, 1, 1024]> var_777_cast_fp16 = conv(bias = core_stacks_0_layers_1_self_attn2_in_proj_bias_to_fp16, dilations = var_777_dilations_0, groups = var_777_groups_0, pad = var_777_pad_0, pad_type = var_777_pad_type_0, strides = var_777_strides_0, weight = core_stacks_0_layers_1_self_attn2_in_proj_weight_to_fp16, x = x_79_cast_fp16)[name = tensor<string, []>("op_777_cast_fp16")];
            tensor<int32, [4]> var_778 = const()[name = tensor<string, []>("op_778"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_779_cast_fp16 = reshape(shape = var_778, x = var_777_cast_fp16)[name = tensor<string, []>("op_779_cast_fp16")];
            tensor<int32, [4]> transpose_15_perm_0 = const()[name = tensor<string, []>("transpose_15_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_75 = const()[name = tensor<string, []>("concat_75"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = var_779_cast_fp16)[name = tensor<string, []>("transpose_298")];
            tensor<fp16, [4, 1024, 12]> reshape_22_cast_fp16 = reshape(shape = concat_75, x = transpose_15_cast_fp16)[name = tensor<string, []>("reshape_22_cast_fp16")];
            tensor<bool, []> matmul_7_transpose_x_0 = const()[name = tensor<string, []>("matmul_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_7_transpose_y_0 = const()[name = tensor<string, []>("matmul_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = reshape_18_cast_fp16, y = reshape_22_cast_fp16)[name = tensor<string, []>("matmul_7_cast_fp16")];
            tensor<int32, [4]> concat_79 = const()[name = tensor<string, []>("concat_79"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_23_cast_fp16 = reshape(shape = concat_79, x = matmul_7_cast_fp16)[name = tensor<string, []>("reshape_23_cast_fp16")];
            tensor<int32, [4]> out_7_perm_0 = const()[name = tensor<string, []>("out_7_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_782 = const()[name = tensor<string, []>("op_782"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_7_cast_fp16 = transpose(perm = out_7_perm_0, x = reshape_23_cast_fp16)[name = tensor<string, []>("transpose_297")];
            tensor<fp16, [1, 48, 1, 1024]> input_61_cast_fp16 = reshape(shape = var_782, x = out_7_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
            tensor<string, []> var_790_pad_type_0 = const()[name = tensor<string, []>("op_790_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_790_strides_0 = const()[name = tensor<string, []>("op_790_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_790_pad_0 = const()[name = tensor<string, []>("op_790_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_790_dilations_0 = const()[name = tensor<string, []>("op_790_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_790_groups_0 = const()[name = tensor<string, []>("op_790_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_0_layers_1_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92070336)))];
            tensor<fp16, [512]> core_stacks_0_layers_1_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92119552)))];
            tensor<fp16, [1, 512, 1, 1024]> var_790_cast_fp16 = conv(bias = core_stacks_0_layers_1_self_attn2_out_proj_bias_to_fp16, dilations = var_790_dilations_0, groups = var_790_groups_0, pad = var_790_pad_0, pad_type = var_790_pad_type_0, strides = var_790_strides_0, weight = core_stacks_0_layers_1_self_attn2_out_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("op_790_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = var_790_cast_fp16)[name = tensor<string, []>("x_81_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = time_emb_1_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")];
            tensor<string, []> x_85_pad_type_0 = const()[name = tensor<string, []>("x_85_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_85_strides_0 = const()[name = tensor<string, []>("x_85_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_85_pad_0 = const()[name = tensor<string, []>("x_85_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_85_dilations_0 = const()[name = tensor<string, []>("x_85_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_85_groups_0 = const()[name = tensor<string, []>("x_85_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_0_layers_1_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92120640)))];
            tensor<fp16, [1024]> core_stacks_0_layers_1_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93169280)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_85_cast_fp16 = conv(bias = core_stacks_0_layers_1_conv_module2_in_proj_bias_to_fp16, dilations = x_85_dilations_0, groups = x_85_groups_0, pad = x_85_pad_0, pad_type = x_85_pad_type_0, strides = x_85_strides_0, weight = core_stacks_0_layers_1_conv_module2_in_proj_weight_to_fp16, x = x_83_cast_fp16)[name = tensor<string, []>("x_85_cast_fp16")];
            tensor<int32, [4]> var_808_begin_0 = const()[name = tensor<string, []>("op_808_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_808_end_0 = const()[name = tensor<string, []>("op_808_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_808_end_mask_0 = const()[name = tensor<string, []>("op_808_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_808_cast_fp16 = slice_by_index(begin = var_808_begin_0, end = var_808_end_0, end_mask = var_808_end_mask_0, x = x_85_cast_fp16)[name = tensor<string, []>("op_808_cast_fp16")];
            tensor<int32, [4]> var_810_begin_0 = const()[name = tensor<string, []>("op_810_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_810_end_0 = const()[name = tensor<string, []>("op_810_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_810_end_mask_0 = const()[name = tensor<string, []>("op_810_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_810_cast_fp16 = slice_by_index(begin = var_810_begin_0, end = var_810_end_0, end_mask = var_810_end_mask_0, x = x_85_cast_fp16)[name = tensor<string, []>("op_810_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_811_cast_fp16 = sigmoid(x = var_810_cast_fp16)[name = tensor<string, []>("op_811_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_87_cast_fp16 = mul(x = var_808_cast_fp16, y = var_811_cast_fp16)[name = tensor<string, []>("x_87_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_63_cast_fp16 = mul(x = x_87_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<string, []> var_820_pad_type_0 = const()[name = tensor<string, []>("op_820_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_820_pad_0 = const()[name = tensor<string, []>("op_820_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_820_groups_0 = const()[name = tensor<string, []>("op_820_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_820_strides_0 = const()[name = tensor<string, []>("op_820_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_820_dilations_0 = const()[name = tensor<string, []>("op_820_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_0_layers_1_conv_module2_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module2_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93171392)))];
            tensor<fp16, [512]> core_stacks_0_layers_1_conv_module2_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module2_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93186816)))];
            tensor<fp16, [1, 512, 1, 1040]> var_820_cast_fp16 = conv(bias = core_stacks_0_layers_1_conv_module2_dw_a_bias_to_fp16, dilations = var_820_dilations_0, groups = var_820_groups_0, pad = var_820_pad_0, pad_type = var_820_pad_type_0, strides = var_820_strides_0, weight = core_stacks_0_layers_1_conv_module2_dw_a_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("op_820_cast_fp16")];
            tensor<int32, [4]> var_821_begin_0 = const()[name = tensor<string, []>("op_821_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_821_end_0 = const()[name = tensor<string, []>("op_821_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_821_end_mask_0 = const()[name = tensor<string, []>("op_821_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_821_cast_fp16 = slice_by_index(begin = var_821_begin_0, end = var_821_end_0, end_mask = var_821_end_mask_0, x = var_820_cast_fp16)[name = tensor<string, []>("op_821_cast_fp16")];
            tensor<string, []> var_827_pad_type_0 = const()[name = tensor<string, []>("op_827_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_827_pad_0 = const()[name = tensor<string, []>("op_827_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_827_groups_0 = const()[name = tensor<string, []>("op_827_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_827_strides_0 = const()[name = tensor<string, []>("op_827_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_827_dilations_0 = const()[name = tensor<string, []>("op_827_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_0_layers_1_conv_module2_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module2_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93187904)))];
            tensor<fp16, [1, 512, 1, 1040]> var_827_cast_fp16 = conv(dilations = var_827_dilations_0, groups = var_827_groups_0, pad = var_827_pad_0, pad_type = var_827_pad_type_0, strides = var_827_strides_0, weight = core_stacks_0_layers_1_conv_module2_dw_b_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("op_827_cast_fp16")];
            tensor<int32, [4]> var_831_begin_0 = const()[name = tensor<string, []>("op_831_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_831_end_0 = const()[name = tensor<string, []>("op_831_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_831_end_mask_0 = const()[name = tensor<string, []>("op_831_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_831_cast_fp16 = slice_by_index(begin = var_831_begin_0, end = var_831_end_0, end_mask = var_831_end_mask_0, x = var_827_cast_fp16)[name = tensor<string, []>("op_831_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_832_cast_fp16 = add(x = var_821_cast_fp16, y = var_831_cast_fp16)[name = tensor<string, []>("op_832_cast_fp16")];
            tensor<string, []> var_838_pad_type_0 = const()[name = tensor<string, []>("op_838_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_838_pad_0 = const()[name = tensor<string, []>("op_838_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_838_groups_0 = const()[name = tensor<string, []>("op_838_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_838_strides_0 = const()[name = tensor<string, []>("op_838_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_838_dilations_0 = const()[name = tensor<string, []>("op_838_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_0_layers_1_conv_module2_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module2_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93203328)))];
            tensor<fp16, [1, 512, 1, 1054]> var_838_cast_fp16 = conv(dilations = var_838_dilations_0, groups = var_838_groups_0, pad = var_838_pad_0, pad_type = var_838_pad_type_0, strides = var_838_strides_0, weight = core_stacks_0_layers_1_conv_module2_dw_c_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("op_838_cast_fp16")];
            tensor<int32, [4]> var_842_begin_0 = const()[name = tensor<string, []>("op_842_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_842_end_0 = const()[name = tensor<string, []>("op_842_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_842_end_mask_0 = const()[name = tensor<string, []>("op_842_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_842_cast_fp16 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = var_838_cast_fp16)[name = tensor<string, []>("op_842_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_89_cast_fp16 = add(x = var_832_cast_fp16, y = var_842_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
            tensor<fp16, []> var_844_to_fp16 = const()[name = tensor<string, []>("op_844_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_91_cast_fp16 = sub(x = x_89_cast_fp16, y = var_844_to_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_846_cast_fp16 = relu(x = x_91_cast_fp16)[name = tensor<string, []>("op_846_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_847_cast_fp16 = abs(x = x_91_cast_fp16)[name = tensor<string, []>("op_847_cast_fp16")];
            tensor<fp16, []> const_18_promoted_to_fp16 = const()[name = tensor<string, []>("const_18_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_848_cast_fp16 = mul(x = var_847_cast_fp16, y = const_18_promoted_to_fp16)[name = tensor<string, []>("op_848_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_849_cast_fp16 = exp(x = var_848_cast_fp16)[name = tensor<string, []>("op_849_cast_fp16")];
            tensor<fp32, []> var_850_epsilon_0 = const()[name = tensor<string, []>("op_850_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_850_cast_fp16 = log(epsilon = var_850_epsilon_0, x = var_849_cast_fp16)[name = tensor<string, []>("op_850_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_851_cast_fp16 = add(x = var_846_cast_fp16, y = var_850_cast_fp16)[name = tensor<string, []>("op_851_cast_fp16")];
            tensor<fp16, []> var_852_to_fp16 = const()[name = tensor<string, []>("op_852_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_853_cast_fp16 = mul(x = x_89_cast_fp16, y = var_852_to_fp16)[name = tensor<string, []>("op_853_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_854_cast_fp16 = sub(x = var_851_cast_fp16, y = var_853_cast_fp16)[name = tensor<string, []>("op_854_cast_fp16")];
            tensor<fp16, []> var_855_to_fp16 = const()[name = tensor<string, []>("op_855_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_65_cast_fp16 = sub(x = var_854_cast_fp16, y = var_855_to_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<string, []> var_863_pad_type_0 = const()[name = tensor<string, []>("op_863_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_863_strides_0 = const()[name = tensor<string, []>("op_863_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_863_pad_0 = const()[name = tensor<string, []>("op_863_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_863_dilations_0 = const()[name = tensor<string, []>("op_863_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_863_groups_0 = const()[name = tensor<string, []>("op_863_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_0_layers_1_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93204416)))];
            tensor<fp16, [512]> core_stacks_0_layers_1_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93728768)))];
            tensor<fp16, [1, 512, 1, 1024]> var_863_cast_fp16 = conv(bias = core_stacks_0_layers_1_conv_module2_out_proj_bias_to_fp16, dilations = var_863_dilations_0, groups = var_863_groups_0, pad = var_863_pad_0, pad_type = var_863_pad_type_0, strides = var_863_strides_0, weight = core_stacks_0_layers_1_conv_module2_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("op_863_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_67_cast_fp16 = add(x = x_83_cast_fp16, y = var_863_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<string, []> x_93_pad_type_0 = const()[name = tensor<string, []>("x_93_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_93_strides_0 = const()[name = tensor<string, []>("x_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_93_pad_0 = const()[name = tensor<string, []>("x_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_93_dilations_0 = const()[name = tensor<string, []>("x_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_93_groups_0 = const()[name = tensor<string, []>("x_93_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_0_layers_1_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93729856)))];
            tensor<fp16, [1920]> core_stacks_0_layers_1_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95696000)))];
            tensor<fp16, [1, 1920, 1, 1024]> x_93_cast_fp16 = conv(bias = core_stacks_0_layers_1_feed_forward3_in_proj_bias_to_fp16, dilations = x_93_dilations_0, groups = x_93_groups_0, pad = x_93_pad_0, pad_type = x_93_pad_type_0, strides = x_93_strides_0, weight = core_stacks_0_layers_1_feed_forward3_in_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("x_93_cast_fp16")];
            tensor<fp16, []> var_874_to_fp16 = const()[name = tensor<string, []>("op_874_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 1024]> x_95_cast_fp16 = sub(x = x_93_cast_fp16, y = var_874_to_fp16)[name = tensor<string, []>("x_95_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_876_cast_fp16 = relu(x = x_95_cast_fp16)[name = tensor<string, []>("op_876_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_877_cast_fp16 = abs(x = x_95_cast_fp16)[name = tensor<string, []>("op_877_cast_fp16")];
            tensor<fp16, []> const_19_promoted_to_fp16 = const()[name = tensor<string, []>("const_19_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_878_cast_fp16 = mul(x = var_877_cast_fp16, y = const_19_promoted_to_fp16)[name = tensor<string, []>("op_878_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_879_cast_fp16 = exp(x = var_878_cast_fp16)[name = tensor<string, []>("op_879_cast_fp16")];
            tensor<fp32, []> var_880_epsilon_0 = const()[name = tensor<string, []>("op_880_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_880_cast_fp16 = log(epsilon = var_880_epsilon_0, x = var_879_cast_fp16)[name = tensor<string, []>("op_880_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_881_cast_fp16 = add(x = var_876_cast_fp16, y = var_880_cast_fp16)[name = tensor<string, []>("op_881_cast_fp16")];
            tensor<fp16, []> var_882_to_fp16 = const()[name = tensor<string, []>("op_882_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 1024]> var_883_cast_fp16 = mul(x = x_93_cast_fp16, y = var_882_to_fp16)[name = tensor<string, []>("op_883_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_884_cast_fp16 = sub(x = var_881_cast_fp16, y = var_883_cast_fp16)[name = tensor<string, []>("op_884_cast_fp16")];
            tensor<fp16, []> var_885_to_fp16 = const()[name = tensor<string, []>("op_885_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 1024]> input_69_cast_fp16 = sub(x = var_884_cast_fp16, y = var_885_to_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
            tensor<string, []> var_893_pad_type_0 = const()[name = tensor<string, []>("op_893_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_893_strides_0 = const()[name = tensor<string, []>("op_893_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_893_pad_0 = const()[name = tensor<string, []>("op_893_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_893_dilations_0 = const()[name = tensor<string, []>("op_893_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_893_groups_0 = const()[name = tensor<string, []>("op_893_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_0_layers_1_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(95699904)))];
            tensor<fp16, [512]> core_stacks_0_layers_1_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97666048)))];
            tensor<fp16, [1, 512, 1, 1024]> var_893_cast_fp16 = conv(bias = core_stacks_0_layers_1_feed_forward3_out_proj_bias_to_fp16, dilations = var_893_dilations_0, groups = var_893_groups_0, pad = var_893_pad_0, pad_type = var_893_pad_type_0, strides = var_893_strides_0, weight = core_stacks_0_layers_1_feed_forward3_out_proj_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("op_893_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_97_cast_fp16 = add(x = input_67_cast_fp16, y = var_893_cast_fp16)[name = tensor<string, []>("x_97_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_0_layers_1_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97667136)))];
            tensor<fp16, [1, 512, 1, 1024]> d_3_cast_fp16 = sub(x = x_97_cast_fp16, y = core_stacks_0_layers_1_norm_bias_to_fp16)[name = tensor<string, []>("d_3_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_898_cast_fp16 = mul(x = d_3_cast_fp16, y = d_3_cast_fp16)[name = tensor<string, []>("op_898_cast_fp16")];
            tensor<int32, [1]> var_900_axes_0 = const()[name = tensor<string, []>("op_900_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_900_keep_dims_0 = const()[name = tensor<string, []>("op_900_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 1024]> var_900_cast_fp16 = reduce_mean(axes = var_900_axes_0, keep_dims = var_900_keep_dims_0, x = var_898_cast_fp16)[name = tensor<string, []>("op_900_cast_fp16")];
            tensor<fp32, []> scales_3_epsilon_0 = const()[name = tensor<string, []>("scales_3_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1024]> scales_3_cast_fp16 = rsqrt(epsilon = scales_3_epsilon_0, x = var_900_cast_fp16)[name = tensor<string, []>("scales_3_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_902_cast_fp16 = mul(x = x_97_cast_fp16, y = scales_3_cast_fp16)[name = tensor<string, []>("op_902_cast_fp16")];
            tensor<fp16, []> core_stacks_0_layers_1_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.a34p+0)];
            tensor<fp16, [1, 512, 1, 1024]> src_7_cast_fp16 = mul(x = var_902_cast_fp16, y = core_stacks_0_layers_1_norm_scale_to_fp16)[name = tensor<string, []>("src_7_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_905_cast_fp16 = sub(x = src_7_cast_fp16, y = input_41_cast_fp16)[name = tensor<string, []>("op_905_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_0_layers_1_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_0_layers_1_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97668224)))];
            tensor<fp16, [1, 512, 1, 1024]> var_906_cast_fp16 = mul(x = var_905_cast_fp16, y = core_stacks_0_layers_1_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_906_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_71_cast_fp16 = add(x = input_41_cast_fp16, y = var_906_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<string, []> input_75_pad_type_0 = const()[name = tensor<string, []>("input_75_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_75_strides_0 = const()[name = tensor<string, []>("input_75_strides_0"), val = tensor<int32, [2]>([1, 2])];
            tensor<int32, []> input_75_groups_0 = const()[name = tensor<string, []>("input_75_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [4]> input_75_pad_0 = const()[name = tensor<string, []>("input_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_75_dilations_0 = const()[name = tensor<string, []>("input_75_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 2]> core_stacks_1_down_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_down_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 2]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97669312)))];
            tensor<fp16, [1, 512, 1, 512]> input_75_cast_fp16 = conv(dilations = input_75_dilations_0, groups = input_75_groups_0, pad = input_75_pad_0, pad_type = input_75_pad_type_0, strides = input_75_strides_0, weight = core_stacks_1_down_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<string, []> time_emb_3_pad_type_0 = const()[name = tensor<string, []>("time_emb_3_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> time_emb_3_strides_0 = const()[name = tensor<string, []>("time_emb_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> time_emb_3_pad_0 = const()[name = tensor<string, []>("time_emb_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> time_emb_3_dilations_0 = const()[name = tensor<string, []>("time_emb_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> time_emb_3_groups_0 = const()[name = tensor<string, []>("time_emb_3_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 192, 1, 1]> core_stacks_1_stack_time_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_time_proj_weight_to_fp16"), val = tensor<fp16, [512, 192, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97671424)))];
            tensor<fp16, [512]> core_stacks_1_stack_time_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_time_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97868096)))];
            tensor<fp16, [1, 512, 1, 1]> time_emb_3_cast_fp16 = conv(bias = core_stacks_1_stack_time_proj_bias_to_fp16, dilations = time_emb_3_dilations_0, groups = time_emb_3_groups_0, pad = time_emb_3_pad_0, pad_type = time_emb_3_pad_type_0, strides = time_emb_3_strides_0, weight = core_stacks_1_stack_time_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("time_emb_3_cast_fp16")];
            tensor<string, []> proj_5_pad_type_0 = const()[name = tensor<string, []>("proj_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_5_strides_0 = const()[name = tensor<string, []>("proj_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_5_pad_0 = const()[name = tensor<string, []>("proj_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_5_dilations_0 = const()[name = tensor<string, []>("proj_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_5_groups_0 = const()[name = tensor<string, []>("proj_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_1_stack_layers_0_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97869184)))];
            tensor<fp16, [384]> core_stacks_1_stack_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98262464)))];
            tensor<fp16, [1, 384, 1, 512]> proj_5_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_5_dilations_0, groups = proj_5_groups_0, pad = proj_5_pad_0, pad_type = proj_5_pad_type_0, strides = proj_5_strides_0, weight = core_stacks_1_stack_layers_0_self_attn_weights_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("proj_5_cast_fp16")];
            tensor<int32, [4]> var_963_begin_0 = const()[name = tensor<string, []>("op_963_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_963_end_0 = const()[name = tensor<string, []>("op_963_end_0"), val = tensor<int32, [4]>([1, 128, 1, 512])];
            tensor<bool, [4]> var_963_end_mask_0 = const()[name = tensor<string, []>("op_963_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_963_cast_fp16 = slice_by_index(begin = var_963_begin_0, end = var_963_end_0, end_mask = var_963_end_mask_0, x = proj_5_cast_fp16)[name = tensor<string, []>("op_963_cast_fp16")];
            tensor<int32, [4]> var_964 = const()[name = tensor<string, []>("op_964"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_965_cast_fp16 = reshape(shape = var_964, x = var_963_cast_fp16)[name = tensor<string, []>("op_965_cast_fp16")];
            tensor<int32, [4]> var_967_begin_0 = const()[name = tensor<string, []>("op_967_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_967_end_0 = const()[name = tensor<string, []>("op_967_end_0"), val = tensor<int32, [4]>([1, 256, 1, 512])];
            tensor<bool, [4]> var_967_end_mask_0 = const()[name = tensor<string, []>("op_967_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_967_cast_fp16 = slice_by_index(begin = var_967_begin_0, end = var_967_end_0, end_mask = var_967_end_mask_0, x = proj_5_cast_fp16)[name = tensor<string, []>("op_967_cast_fp16")];
            tensor<int32, [4]> var_968 = const()[name = tensor<string, []>("op_968"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_969_cast_fp16 = reshape(shape = var_968, x = var_967_cast_fp16)[name = tensor<string, []>("op_969_cast_fp16")];
            tensor<int32, [4]> var_971_begin_0 = const()[name = tensor<string, []>("op_971_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_971_end_0 = const()[name = tensor<string, []>("op_971_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_971_end_mask_0 = const()[name = tensor<string, []>("op_971_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_971_cast_fp16 = slice_by_index(begin = var_971_begin_0, end = var_971_end_0, end_mask = var_971_end_mask_0, x = proj_5_cast_fp16)[name = tensor<string, []>("op_971_cast_fp16")];
            tensor<int32, [4]> var_972 = const()[name = tensor<string, []>("op_972"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> pt_5_cast_fp16 = reshape(shape = var_972, x = var_971_cast_fp16)[name = tensor<string, []>("pt_5_cast_fp16")];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = tensor<string, []>("transpose_16_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_84 = const()[name = tensor<string, []>("concat_84"), val = tensor<int32, [3]>([4, 512, 32])];
            tensor<fp16, [4, 1, 512, 32]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = var_965_cast_fp16)[name = tensor<string, []>("transpose_296")];
            tensor<fp16, [4, 512, 32]> reshape_24_cast_fp16 = reshape(shape = concat_84, x = transpose_16_cast_fp16)[name = tensor<string, []>("reshape_24_cast_fp16")];
            tensor<int32, [4]> transpose_17_perm_0 = const()[name = tensor<string, []>("transpose_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_85 = const()[name = tensor<string, []>("concat_85"), val = tensor<int32, [3]>([4, 32, 512])];
            tensor<fp16, [4, 1, 32, 512]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = var_969_cast_fp16)[name = tensor<string, []>("transpose_295")];
            tensor<fp16, [4, 32, 512]> reshape_25_cast_fp16 = reshape(shape = concat_85, x = transpose_17_cast_fp16)[name = tensor<string, []>("reshape_25_cast_fp16")];
            tensor<bool, []> matmul_8_transpose_x_0 = const()[name = tensor<string, []>("matmul_8_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_8_transpose_y_0 = const()[name = tensor<string, []>("matmul_8_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 512]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = reshape_24_cast_fp16, y = reshape_25_cast_fp16)[name = tensor<string, []>("matmul_8_cast_fp16")];
            tensor<int32, [4]> concat_89 = const()[name = tensor<string, []>("concat_89"), val = tensor<int32, [4]>([4, 1, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> reshape_26_cast_fp16 = reshape(shape = concat_89, x = matmul_8_cast_fp16)[name = tensor<string, []>("reshape_26_cast_fp16")];
            tensor<int32, [4]> attn_scores_13_perm_0 = const()[name = tensor<string, []>("attn_scores_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_976 = const()[name = tensor<string, []>("op_976"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 32, 512, 512]> core_pos_basis_512_to_fp16 = const()[name = tensor<string, []>("core_pos_basis_512_to_fp16"), val = tensor<fp16, [1, 32, 512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98263296)))];
            tensor<fp16, [4, 32, 512, 1]> var_977_cast_fp16 = transpose(perm = var_976, x = pt_5_cast_fp16)[name = tensor<string, []>("transpose_293")];
            tensor<fp16, [4, 32, 512, 512]> var_978_cast_fp16 = mul(x = var_977_cast_fp16, y = core_pos_basis_512_to_fp16)[name = tensor<string, []>("op_978_cast_fp16")];
            tensor<int32, [1]> pos_scores_5_axes_0 = const()[name = tensor<string, []>("pos_scores_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_5_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_5_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 512, 512]> pos_scores_5_cast_fp16 = reduce_sum(axes = pos_scores_5_axes_0, keep_dims = pos_scores_5_keep_dims_0, x = var_978_cast_fp16)[name = tensor<string, []>("pos_scores_5_cast_fp16")];
            tensor<int32, [4]> var_981 = const()[name = tensor<string, []>("op_981"), val = tensor<int32, [4]>([4, 512, 1, 512])];
            tensor<fp16, [4, 512, 1, 512]> var_982_cast_fp16 = reshape(shape = var_981, x = pos_scores_5_cast_fp16)[name = tensor<string, []>("op_982_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_13_cast_fp16 = transpose(perm = attn_scores_13_perm_0, x = reshape_26_cast_fp16)[name = tensor<string, []>("transpose_294")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_15_cast_fp16 = add(x = attn_scores_13_cast_fp16, y = var_982_cast_fp16)[name = tensor<string, []>("attn_scores_15_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_17_cast_fp16 = add(x = attn_scores_15_cast_fp16, y = attn_bias_3_cast_fp16)[name = tensor<string, []>("attn_scores_17_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_weights_5_cast_fp16 = softmax(axis = var_41, x = attn_scores_17_cast_fp16)[name = tensor<string, []>("attn_weights_5_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_77_cast_fp16 = add(x = input_75_cast_fp16, y = time_emb_3_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<string, []> x_101_pad_type_0 = const()[name = tensor<string, []>("x_101_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_101_strides_0 = const()[name = tensor<string, []>("x_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_101_pad_0 = const()[name = tensor<string, []>("x_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_101_dilations_0 = const()[name = tensor<string, []>("x_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_101_groups_0 = const()[name = tensor<string, []>("x_101_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_1_stack_layers_0_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115040576)))];
            tensor<fp16, [1152]> core_stacks_1_stack_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116220288)))];
            tensor<fp16, [1, 1152, 1, 512]> x_101_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_feed_forward1_in_proj_bias_to_fp16, dilations = x_101_dilations_0, groups = x_101_groups_0, pad = x_101_pad_0, pad_type = x_101_pad_type_0, strides = x_101_strides_0, weight = core_stacks_1_stack_layers_0_feed_forward1_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("x_101_cast_fp16")];
            tensor<fp16, []> var_996_to_fp16 = const()[name = tensor<string, []>("op_996_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 512]> x_103_cast_fp16 = sub(x = x_101_cast_fp16, y = var_996_to_fp16)[name = tensor<string, []>("x_103_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_998_cast_fp16 = relu(x = x_103_cast_fp16)[name = tensor<string, []>("op_998_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_999_cast_fp16 = abs(x = x_103_cast_fp16)[name = tensor<string, []>("op_999_cast_fp16")];
            tensor<fp16, []> const_21_promoted_to_fp16 = const()[name = tensor<string, []>("const_21_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_1000_cast_fp16 = mul(x = var_999_cast_fp16, y = const_21_promoted_to_fp16)[name = tensor<string, []>("op_1000_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_1001_cast_fp16 = exp(x = var_1000_cast_fp16)[name = tensor<string, []>("op_1001_cast_fp16")];
            tensor<fp32, []> var_1002_epsilon_0 = const()[name = tensor<string, []>("op_1002_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_1002_cast_fp16 = log(epsilon = var_1002_epsilon_0, x = var_1001_cast_fp16)[name = tensor<string, []>("op_1002_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_1003_cast_fp16 = add(x = var_998_cast_fp16, y = var_1002_cast_fp16)[name = tensor<string, []>("op_1003_cast_fp16")];
            tensor<fp16, []> var_1004_to_fp16 = const()[name = tensor<string, []>("op_1004_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 512]> var_1005_cast_fp16 = mul(x = x_101_cast_fp16, y = var_1004_to_fp16)[name = tensor<string, []>("op_1005_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_1006_cast_fp16 = sub(x = var_1003_cast_fp16, y = var_1005_cast_fp16)[name = tensor<string, []>("op_1006_cast_fp16")];
            tensor<fp16, []> var_1007_to_fp16 = const()[name = tensor<string, []>("op_1007_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 512]> input_79_cast_fp16 = sub(x = var_1006_cast_fp16, y = var_1007_to_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<string, []> var_1015_pad_type_0 = const()[name = tensor<string, []>("op_1015_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1015_strides_0 = const()[name = tensor<string, []>("op_1015_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1015_pad_0 = const()[name = tensor<string, []>("op_1015_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1015_dilations_0 = const()[name = tensor<string, []>("op_1015_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1015_groups_0 = const()[name = tensor<string, []>("op_1015_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_1_stack_layers_0_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116222656)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117402368)))];
            tensor<fp16, [1, 512, 1, 512]> var_1015_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_feed_forward1_out_proj_bias_to_fp16, dilations = var_1015_dilations_0, groups = var_1015_groups_0, pad = var_1015_pad_0, pad_type = var_1015_pad_type_0, strides = var_1015_strides_0, weight = core_stacks_1_stack_layers_0_feed_forward1_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("op_1015_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_81_cast_fp16 = add(x = input_77_cast_fp16, y = var_1015_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<int32, [4]> var_1017_begin_0 = const()[name = tensor<string, []>("op_1017_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1017_end_0 = const()[name = tensor<string, []>("op_1017_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_1017_end_mask_0 = const()[name = tensor<string, []>("op_1017_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_1017_cast_fp16 = slice_by_index(begin = var_1017_begin_0, end = var_1017_end_0, end_mask = var_1017_end_mask_0, x = attn_weights_5_cast_fp16)[name = tensor<string, []>("op_1017_cast_fp16")];
            tensor<string, []> x_105_pad_type_0 = const()[name = tensor<string, []>("x_105_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_105_strides_0 = const()[name = tensor<string, []>("x_105_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_105_pad_0 = const()[name = tensor<string, []>("x_105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_105_dilations_0 = const()[name = tensor<string, []>("x_105_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_105_groups_0 = const()[name = tensor<string, []>("x_105_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_1_stack_layers_0_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117403456)))];
            tensor<fp16, [1152]> core_stacks_1_stack_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118583168)))];
            tensor<fp16, [1, 1152, 1, 512]> x_105_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_nonlin_attention_in_proj_bias_to_fp16, dilations = x_105_dilations_0, groups = x_105_groups_0, pad = x_105_pad_0, pad_type = x_105_pad_type_0, strides = x_105_strides_0, weight = core_stacks_1_stack_layers_0_nonlin_attention_in_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("x_105_cast_fp16")];
            tensor<int32, [4]> var_1028_begin_0 = const()[name = tensor<string, []>("op_1028_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1028_end_0 = const()[name = tensor<string, []>("op_1028_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_1028_end_mask_0 = const()[name = tensor<string, []>("op_1028_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_1028_cast_fp16 = slice_by_index(begin = var_1028_begin_0, end = var_1028_end_0, end_mask = var_1028_end_mask_0, x = x_105_cast_fp16)[name = tensor<string, []>("op_1028_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> s_gate_5_cast_fp16 = tanh(x = var_1028_cast_fp16)[name = tensor<string, []>("s_gate_5_cast_fp16")];
            tensor<int32, [4]> var_1031_begin_0 = const()[name = tensor<string, []>("op_1031_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_1031_end_0 = const()[name = tensor<string, []>("op_1031_end_0"), val = tensor<int32, [4]>([1, 768, 1, 512])];
            tensor<bool, [4]> var_1031_end_mask_0 = const()[name = tensor<string, []>("op_1031_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_1031_cast_fp16 = slice_by_index(begin = var_1031_begin_0, end = var_1031_end_0, end_mask = var_1031_end_mask_0, x = x_105_cast_fp16)[name = tensor<string, []>("op_1031_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> var_1032_cast_fp16 = mul(x = var_1031_cast_fp16, y = s_gate_5_cast_fp16)[name = tensor<string, []>("op_1032_cast_fp16")];
            tensor<int32, [4]> y_5_begin_0 = const()[name = tensor<string, []>("y_5_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_5_end_0 = const()[name = tensor<string, []>("y_5_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 512])];
            tensor<bool, [4]> y_5_end_mask_0 = const()[name = tensor<string, []>("y_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 512]> y_5_cast_fp16 = slice_by_index(begin = y_5_begin_0, end = y_5_end_0, end_mask = y_5_end_mask_0, x = x_105_cast_fp16)[name = tensor<string, []>("y_5_cast_fp16")];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = tensor<string, []>("transpose_18_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_94 = const()[name = tensor<string, []>("concat_94"), val = tensor<int32, [3]>([1, 512, 512])];
            tensor<fp16, [1, 1, 512, 512]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = var_1017_cast_fp16)[name = tensor<string, []>("transpose_292")];
            tensor<fp16, [1, 512, 512]> reshape_27_cast_fp16 = reshape(shape = concat_94, x = transpose_18_cast_fp16)[name = tensor<string, []>("reshape_27_cast_fp16")];
            tensor<int32, [4]> transpose_19_perm_0 = const()[name = tensor<string, []>("transpose_19_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_95 = const()[name = tensor<string, []>("concat_95"), val = tensor<int32, [3]>([1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = var_1032_cast_fp16)[name = tensor<string, []>("transpose_291")];
            tensor<fp16, [1, 512, 384]> reshape_28_cast_fp16 = reshape(shape = concat_95, x = transpose_19_cast_fp16)[name = tensor<string, []>("reshape_28_cast_fp16")];
            tensor<bool, []> matmul_9_transpose_x_0 = const()[name = tensor<string, []>("matmul_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_9_transpose_y_0 = const()[name = tensor<string, []>("matmul_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 384]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = reshape_27_cast_fp16, y = reshape_28_cast_fp16)[name = tensor<string, []>("matmul_9_cast_fp16")];
            tensor<int32, [4]> concat_99 = const()[name = tensor<string, []>("concat_99"), val = tensor<int32, [4]>([1, 1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> reshape_29_cast_fp16 = reshape(shape = concat_99, x = matmul_9_cast_fp16)[name = tensor<string, []>("reshape_29_cast_fp16")];
            tensor<int32, [4]> v_5_perm_0 = const()[name = tensor<string, []>("v_5_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 512]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = reshape_29_cast_fp16)[name = tensor<string, []>("transpose_290")];
            tensor<fp16, [1, 384, 1, 512]> input_83_cast_fp16 = mul(x = v_5_cast_fp16, y = y_5_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<string, []> var_1044_pad_type_0 = const()[name = tensor<string, []>("op_1044_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1044_strides_0 = const()[name = tensor<string, []>("op_1044_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1044_pad_0 = const()[name = tensor<string, []>("op_1044_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1044_dilations_0 = const()[name = tensor<string, []>("op_1044_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1044_groups_0 = const()[name = tensor<string, []>("op_1044_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_1_stack_layers_0_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118585536)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118978816)))];
            tensor<fp16, [1, 512, 1, 512]> var_1044_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_nonlin_attention_out_proj_bias_to_fp16, dilations = var_1044_dilations_0, groups = var_1044_groups_0, pad = var_1044_pad_0, pad_type = var_1044_pad_type_0, strides = var_1044_strides_0, weight = core_stacks_1_stack_layers_0_nonlin_attention_out_proj_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("op_1044_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_107_cast_fp16 = add(x = input_81_cast_fp16, y = var_1044_cast_fp16)[name = tensor<string, []>("x_107_cast_fp16")];
            tensor<string, []> var_1055_pad_type_0 = const()[name = tensor<string, []>("op_1055_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1055_strides_0 = const()[name = tensor<string, []>("op_1055_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1055_pad_0 = const()[name = tensor<string, []>("op_1055_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1055_dilations_0 = const()[name = tensor<string, []>("op_1055_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1055_groups_0 = const()[name = tensor<string, []>("op_1055_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_1_stack_layers_0_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118979904)))];
            tensor<fp16, [48]> core_stacks_1_stack_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119029120)))];
            tensor<fp16, [1, 48, 1, 512]> var_1055_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_self_attn1_in_proj_bias_to_fp16, dilations = var_1055_dilations_0, groups = var_1055_groups_0, pad = var_1055_pad_0, pad_type = var_1055_pad_type_0, strides = var_1055_strides_0, weight = core_stacks_1_stack_layers_0_self_attn1_in_proj_weight_to_fp16, x = x_107_cast_fp16)[name = tensor<string, []>("op_1055_cast_fp16")];
            tensor<int32, [4]> var_1056 = const()[name = tensor<string, []>("op_1056"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_1057_cast_fp16 = reshape(shape = var_1056, x = var_1055_cast_fp16)[name = tensor<string, []>("op_1057_cast_fp16")];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = tensor<string, []>("transpose_20_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_104 = const()[name = tensor<string, []>("concat_104"), val = tensor<int32, [3]>([4, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = attn_weights_5_cast_fp16)[name = tensor<string, []>("transpose_289")];
            tensor<fp16, [4, 512, 512]> reshape_30_cast_fp16 = reshape(shape = concat_104, x = transpose_20_cast_fp16)[name = tensor<string, []>("reshape_30_cast_fp16")];
            tensor<int32, [4]> transpose_21_perm_0 = const()[name = tensor<string, []>("transpose_21_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_105 = const()[name = tensor<string, []>("concat_105"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = var_1057_cast_fp16)[name = tensor<string, []>("transpose_288")];
            tensor<fp16, [4, 512, 12]> reshape_31_cast_fp16 = reshape(shape = concat_105, x = transpose_21_cast_fp16)[name = tensor<string, []>("reshape_31_cast_fp16")];
            tensor<bool, []> matmul_10_transpose_x_0 = const()[name = tensor<string, []>("matmul_10_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_10_transpose_y_0 = const()[name = tensor<string, []>("matmul_10_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = reshape_30_cast_fp16, y = reshape_31_cast_fp16)[name = tensor<string, []>("matmul_10_cast_fp16")];
            tensor<int32, [4]> concat_109 = const()[name = tensor<string, []>("concat_109"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_32_cast_fp16 = reshape(shape = concat_109, x = matmul_10_cast_fp16)[name = tensor<string, []>("reshape_32_cast_fp16")];
            tensor<int32, [4]> out_9_perm_0 = const()[name = tensor<string, []>("out_9_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_1060 = const()[name = tensor<string, []>("op_1060"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_9_cast_fp16 = transpose(perm = out_9_perm_0, x = reshape_32_cast_fp16)[name = tensor<string, []>("transpose_287")];
            tensor<fp16, [1, 48, 1, 512]> input_85_cast_fp16 = reshape(shape = var_1060, x = out_9_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<string, []> var_1068_pad_type_0 = const()[name = tensor<string, []>("op_1068_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1068_strides_0 = const()[name = tensor<string, []>("op_1068_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1068_pad_0 = const()[name = tensor<string, []>("op_1068_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1068_dilations_0 = const()[name = tensor<string, []>("op_1068_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1068_groups_0 = const()[name = tensor<string, []>("op_1068_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_1_stack_layers_0_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119029312)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119078528)))];
            tensor<fp16, [1, 512, 1, 512]> var_1068_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_self_attn1_out_proj_bias_to_fp16, dilations = var_1068_dilations_0, groups = var_1068_groups_0, pad = var_1068_pad_0, pad_type = var_1068_pad_type_0, strides = var_1068_strides_0, weight = core_stacks_1_stack_layers_0_self_attn1_out_proj_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("op_1068_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_109_cast_fp16 = add(x = x_107_cast_fp16, y = var_1068_cast_fp16)[name = tensor<string, []>("x_109_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_111_cast_fp16 = add(x = x_109_cast_fp16, y = time_emb_3_cast_fp16)[name = tensor<string, []>("x_111_cast_fp16")];
            tensor<string, []> x_113_pad_type_0 = const()[name = tensor<string, []>("x_113_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_113_strides_0 = const()[name = tensor<string, []>("x_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_113_pad_0 = const()[name = tensor<string, []>("x_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_113_dilations_0 = const()[name = tensor<string, []>("x_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_113_groups_0 = const()[name = tensor<string, []>("x_113_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_1_stack_layers_0_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(119079616)))];
            tensor<fp16, [1024]> core_stacks_1_stack_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120128256)))];
            tensor<fp16, [1, 1024, 1, 512]> x_113_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_conv_module1_in_proj_bias_to_fp16, dilations = x_113_dilations_0, groups = x_113_groups_0, pad = x_113_pad_0, pad_type = x_113_pad_type_0, strides = x_113_strides_0, weight = core_stacks_1_stack_layers_0_conv_module1_in_proj_weight_to_fp16, x = x_111_cast_fp16)[name = tensor<string, []>("x_113_cast_fp16")];
            tensor<int32, [4]> var_1082_begin_0 = const()[name = tensor<string, []>("op_1082_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1082_end_0 = const()[name = tensor<string, []>("op_1082_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_1082_end_mask_0 = const()[name = tensor<string, []>("op_1082_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_1082_cast_fp16 = slice_by_index(begin = var_1082_begin_0, end = var_1082_end_0, end_mask = var_1082_end_mask_0, x = x_113_cast_fp16)[name = tensor<string, []>("op_1082_cast_fp16")];
            tensor<int32, [4]> var_1084_begin_0 = const()[name = tensor<string, []>("op_1084_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_1084_end_0 = const()[name = tensor<string, []>("op_1084_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_1084_end_mask_0 = const()[name = tensor<string, []>("op_1084_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_1084_cast_fp16 = slice_by_index(begin = var_1084_begin_0, end = var_1084_end_0, end_mask = var_1084_end_mask_0, x = x_113_cast_fp16)[name = tensor<string, []>("op_1084_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1085_cast_fp16 = sigmoid(x = var_1084_cast_fp16)[name = tensor<string, []>("op_1085_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_115_cast_fp16 = mul(x = var_1082_cast_fp16, y = var_1085_cast_fp16)[name = tensor<string, []>("x_115_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_87_cast_fp16 = mul(x = x_115_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<string, []> x_117_pad_type_0 = const()[name = tensor<string, []>("x_117_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_117_pad_0 = const()[name = tensor<string, []>("x_117_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_117_groups_0 = const()[name = tensor<string, []>("x_117_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_117_strides_0 = const()[name = tensor<string, []>("x_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_117_dilations_0 = const()[name = tensor<string, []>("x_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_1_stack_layers_0_conv_module1_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module1_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120130368)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_0_conv_module1_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module1_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120145792)))];
            tensor<fp16, [1, 512, 1, 512]> x_117_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_conv_module1_dw_bias_to_fp16, dilations = x_117_dilations_0, groups = x_117_groups_0, pad = x_117_pad_0, pad_type = x_117_pad_type_0, strides = x_117_strides_0, weight = core_stacks_1_stack_layers_0_conv_module1_dw_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("x_117_cast_fp16")];
            tensor<fp16, []> var_1095_to_fp16 = const()[name = tensor<string, []>("op_1095_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_119_cast_fp16 = sub(x = x_117_cast_fp16, y = var_1095_to_fp16)[name = tensor<string, []>("x_119_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1097_cast_fp16 = relu(x = x_119_cast_fp16)[name = tensor<string, []>("op_1097_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1098_cast_fp16 = abs(x = x_119_cast_fp16)[name = tensor<string, []>("op_1098_cast_fp16")];
            tensor<fp16, []> const_23_promoted_to_fp16 = const()[name = tensor<string, []>("const_23_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_1099_cast_fp16 = mul(x = var_1098_cast_fp16, y = const_23_promoted_to_fp16)[name = tensor<string, []>("op_1099_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1100_cast_fp16 = exp(x = var_1099_cast_fp16)[name = tensor<string, []>("op_1100_cast_fp16")];
            tensor<fp32, []> var_1101_epsilon_0 = const()[name = tensor<string, []>("op_1101_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_1101_cast_fp16 = log(epsilon = var_1101_epsilon_0, x = var_1100_cast_fp16)[name = tensor<string, []>("op_1101_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1102_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1101_cast_fp16)[name = tensor<string, []>("op_1102_cast_fp16")];
            tensor<fp16, []> var_1103_to_fp16 = const()[name = tensor<string, []>("op_1103_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_1104_cast_fp16 = mul(x = x_117_cast_fp16, y = var_1103_to_fp16)[name = tensor<string, []>("op_1104_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1105_cast_fp16 = sub(x = var_1102_cast_fp16, y = var_1104_cast_fp16)[name = tensor<string, []>("op_1105_cast_fp16")];
            tensor<fp16, []> var_1106_to_fp16 = const()[name = tensor<string, []>("op_1106_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_89_cast_fp16 = sub(x = var_1105_cast_fp16, y = var_1106_to_fp16)[name = tensor<string, []>("input_89_cast_fp16")];
            tensor<string, []> var_1114_pad_type_0 = const()[name = tensor<string, []>("op_1114_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1114_strides_0 = const()[name = tensor<string, []>("op_1114_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1114_pad_0 = const()[name = tensor<string, []>("op_1114_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1114_dilations_0 = const()[name = tensor<string, []>("op_1114_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1114_groups_0 = const()[name = tensor<string, []>("op_1114_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_1_stack_layers_0_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120146880)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120671232)))];
            tensor<fp16, [1, 512, 1, 512]> var_1114_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_conv_module1_out_proj_bias_to_fp16, dilations = var_1114_dilations_0, groups = var_1114_groups_0, pad = var_1114_pad_0, pad_type = var_1114_pad_type_0, strides = var_1114_strides_0, weight = core_stacks_1_stack_layers_0_conv_module1_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("op_1114_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_91_cast_fp16 = add(x = x_111_cast_fp16, y = var_1114_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<string, []> x_121_pad_type_0 = const()[name = tensor<string, []>("x_121_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_121_strides_0 = const()[name = tensor<string, []>("x_121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_121_pad_0 = const()[name = tensor<string, []>("x_121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_121_dilations_0 = const()[name = tensor<string, []>("x_121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_121_groups_0 = const()[name = tensor<string, []>("x_121_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_1_stack_layers_0_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120672320)))];
            tensor<fp16, [1536]> core_stacks_1_stack_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122245248)))];
            tensor<fp16, [1, 1536, 1, 512]> x_121_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_feed_forward2_in_proj_bias_to_fp16, dilations = x_121_dilations_0, groups = x_121_groups_0, pad = x_121_pad_0, pad_type = x_121_pad_type_0, strides = x_121_strides_0, weight = core_stacks_1_stack_layers_0_feed_forward2_in_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("x_121_cast_fp16")];
            tensor<fp16, []> var_1125_to_fp16 = const()[name = tensor<string, []>("op_1125_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 512]> x_123_cast_fp16 = sub(x = x_121_cast_fp16, y = var_1125_to_fp16)[name = tensor<string, []>("x_123_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_1127_cast_fp16 = relu(x = x_123_cast_fp16)[name = tensor<string, []>("op_1127_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_1128_cast_fp16 = abs(x = x_123_cast_fp16)[name = tensor<string, []>("op_1128_cast_fp16")];
            tensor<fp16, []> const_24_promoted_to_fp16 = const()[name = tensor<string, []>("const_24_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_1129_cast_fp16 = mul(x = var_1128_cast_fp16, y = const_24_promoted_to_fp16)[name = tensor<string, []>("op_1129_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_1130_cast_fp16 = exp(x = var_1129_cast_fp16)[name = tensor<string, []>("op_1130_cast_fp16")];
            tensor<fp32, []> var_1131_epsilon_0 = const()[name = tensor<string, []>("op_1131_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_1131_cast_fp16 = log(epsilon = var_1131_epsilon_0, x = var_1130_cast_fp16)[name = tensor<string, []>("op_1131_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_1132_cast_fp16 = add(x = var_1127_cast_fp16, y = var_1131_cast_fp16)[name = tensor<string, []>("op_1132_cast_fp16")];
            tensor<fp16, []> var_1133_to_fp16 = const()[name = tensor<string, []>("op_1133_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 512]> var_1134_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1133_to_fp16)[name = tensor<string, []>("op_1134_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_1135_cast_fp16 = sub(x = var_1132_cast_fp16, y = var_1134_cast_fp16)[name = tensor<string, []>("op_1135_cast_fp16")];
            tensor<fp16, []> var_1136_to_fp16 = const()[name = tensor<string, []>("op_1136_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 512]> input_93_cast_fp16 = sub(x = var_1135_cast_fp16, y = var_1136_to_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<string, []> var_1144_pad_type_0 = const()[name = tensor<string, []>("op_1144_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1144_strides_0 = const()[name = tensor<string, []>("op_1144_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1144_pad_0 = const()[name = tensor<string, []>("op_1144_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1144_dilations_0 = const()[name = tensor<string, []>("op_1144_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1144_groups_0 = const()[name = tensor<string, []>("op_1144_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_1_stack_layers_0_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122248384)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123821312)))];
            tensor<fp16, [1, 512, 1, 512]> var_1144_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_feed_forward2_out_proj_bias_to_fp16, dilations = var_1144_dilations_0, groups = var_1144_groups_0, pad = var_1144_pad_0, pad_type = var_1144_pad_type_0, strides = var_1144_strides_0, weight = core_stacks_1_stack_layers_0_feed_forward2_out_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("op_1144_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> src_9_cast_fp16 = add(x = input_91_cast_fp16, y = var_1144_cast_fp16)[name = tensor<string, []>("src_9_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1147_cast_fp16 = sub(x = src_9_cast_fp16, y = input_75_cast_fp16)[name = tensor<string, []>("op_1147_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_1_stack_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123822400)))];
            tensor<fp16, [1, 512, 1, 512]> var_1148_cast_fp16 = mul(x = var_1147_cast_fp16, y = core_stacks_1_stack_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_1148_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_125_cast_fp16 = add(x = input_75_cast_fp16, y = var_1148_cast_fp16)[name = tensor<string, []>("x_125_cast_fp16")];
            tensor<string, []> var_1159_pad_type_0 = const()[name = tensor<string, []>("op_1159_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1159_strides_0 = const()[name = tensor<string, []>("op_1159_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1159_pad_0 = const()[name = tensor<string, []>("op_1159_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1159_dilations_0 = const()[name = tensor<string, []>("op_1159_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1159_groups_0 = const()[name = tensor<string, []>("op_1159_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_1_stack_layers_0_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123823488)))];
            tensor<fp16, [48]> core_stacks_1_stack_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123872704)))];
            tensor<fp16, [1, 48, 1, 512]> var_1159_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_self_attn2_in_proj_bias_to_fp16, dilations = var_1159_dilations_0, groups = var_1159_groups_0, pad = var_1159_pad_0, pad_type = var_1159_pad_type_0, strides = var_1159_strides_0, weight = core_stacks_1_stack_layers_0_self_attn2_in_proj_weight_to_fp16, x = x_125_cast_fp16)[name = tensor<string, []>("op_1159_cast_fp16")];
            tensor<int32, [4]> var_1160 = const()[name = tensor<string, []>("op_1160"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_1161_cast_fp16 = reshape(shape = var_1160, x = var_1159_cast_fp16)[name = tensor<string, []>("op_1161_cast_fp16")];
            tensor<int32, [4]> transpose_23_perm_0 = const()[name = tensor<string, []>("transpose_23_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_115 = const()[name = tensor<string, []>("concat_115"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = var_1161_cast_fp16)[name = tensor<string, []>("transpose_286")];
            tensor<fp16, [4, 512, 12]> reshape_34_cast_fp16 = reshape(shape = concat_115, x = transpose_23_cast_fp16)[name = tensor<string, []>("reshape_34_cast_fp16")];
            tensor<bool, []> matmul_11_transpose_x_0 = const()[name = tensor<string, []>("matmul_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_11_transpose_y_0 = const()[name = tensor<string, []>("matmul_11_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = reshape_30_cast_fp16, y = reshape_34_cast_fp16)[name = tensor<string, []>("matmul_11_cast_fp16")];
            tensor<int32, [4]> concat_119 = const()[name = tensor<string, []>("concat_119"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_35_cast_fp16 = reshape(shape = concat_119, x = matmul_11_cast_fp16)[name = tensor<string, []>("reshape_35_cast_fp16")];
            tensor<int32, [4]> out_11_perm_0 = const()[name = tensor<string, []>("out_11_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_1164 = const()[name = tensor<string, []>("op_1164"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_11_cast_fp16 = transpose(perm = out_11_perm_0, x = reshape_35_cast_fp16)[name = tensor<string, []>("transpose_285")];
            tensor<fp16, [1, 48, 1, 512]> input_95_cast_fp16 = reshape(shape = var_1164, x = out_11_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<string, []> var_1172_pad_type_0 = const()[name = tensor<string, []>("op_1172_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1172_strides_0 = const()[name = tensor<string, []>("op_1172_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1172_pad_0 = const()[name = tensor<string, []>("op_1172_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1172_dilations_0 = const()[name = tensor<string, []>("op_1172_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1172_groups_0 = const()[name = tensor<string, []>("op_1172_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_1_stack_layers_0_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123872896)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123922112)))];
            tensor<fp16, [1, 512, 1, 512]> var_1172_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_self_attn2_out_proj_bias_to_fp16, dilations = var_1172_dilations_0, groups = var_1172_groups_0, pad = var_1172_pad_0, pad_type = var_1172_pad_type_0, strides = var_1172_strides_0, weight = core_stacks_1_stack_layers_0_self_attn2_out_proj_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("op_1172_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_127_cast_fp16 = add(x = x_125_cast_fp16, y = var_1172_cast_fp16)[name = tensor<string, []>("x_127_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_129_cast_fp16 = add(x = x_127_cast_fp16, y = time_emb_3_cast_fp16)[name = tensor<string, []>("x_129_cast_fp16")];
            tensor<string, []> x_131_pad_type_0 = const()[name = tensor<string, []>("x_131_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_131_strides_0 = const()[name = tensor<string, []>("x_131_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_131_pad_0 = const()[name = tensor<string, []>("x_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_131_dilations_0 = const()[name = tensor<string, []>("x_131_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_131_groups_0 = const()[name = tensor<string, []>("x_131_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_1_stack_layers_0_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123923200)))];
            tensor<fp16, [1024]> core_stacks_1_stack_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124971840)))];
            tensor<fp16, [1, 1024, 1, 512]> x_131_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_conv_module2_in_proj_bias_to_fp16, dilations = x_131_dilations_0, groups = x_131_groups_0, pad = x_131_pad_0, pad_type = x_131_pad_type_0, strides = x_131_strides_0, weight = core_stacks_1_stack_layers_0_conv_module2_in_proj_weight_to_fp16, x = x_129_cast_fp16)[name = tensor<string, []>("x_131_cast_fp16")];
            tensor<int32, [4]> var_1186_begin_0 = const()[name = tensor<string, []>("op_1186_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1186_end_0 = const()[name = tensor<string, []>("op_1186_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_1186_end_mask_0 = const()[name = tensor<string, []>("op_1186_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_1186_cast_fp16 = slice_by_index(begin = var_1186_begin_0, end = var_1186_end_0, end_mask = var_1186_end_mask_0, x = x_131_cast_fp16)[name = tensor<string, []>("op_1186_cast_fp16")];
            tensor<int32, [4]> var_1188_begin_0 = const()[name = tensor<string, []>("op_1188_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_1188_end_0 = const()[name = tensor<string, []>("op_1188_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_1188_end_mask_0 = const()[name = tensor<string, []>("op_1188_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_1188_cast_fp16 = slice_by_index(begin = var_1188_begin_0, end = var_1188_end_0, end_mask = var_1188_end_mask_0, x = x_131_cast_fp16)[name = tensor<string, []>("op_1188_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1189_cast_fp16 = sigmoid(x = var_1188_cast_fp16)[name = tensor<string, []>("op_1189_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_133_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1189_cast_fp16)[name = tensor<string, []>("x_133_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_97_cast_fp16 = mul(x = x_133_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")];
            tensor<string, []> x_135_pad_type_0 = const()[name = tensor<string, []>("x_135_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_135_pad_0 = const()[name = tensor<string, []>("x_135_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_135_groups_0 = const()[name = tensor<string, []>("x_135_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_135_strides_0 = const()[name = tensor<string, []>("x_135_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_135_dilations_0 = const()[name = tensor<string, []>("x_135_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_1_stack_layers_0_conv_module2_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module2_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124973952)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_0_conv_module2_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module2_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124989376)))];
            tensor<fp16, [1, 512, 1, 512]> x_135_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_conv_module2_dw_bias_to_fp16, dilations = x_135_dilations_0, groups = x_135_groups_0, pad = x_135_pad_0, pad_type = x_135_pad_type_0, strides = x_135_strides_0, weight = core_stacks_1_stack_layers_0_conv_module2_dw_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("x_135_cast_fp16")];
            tensor<fp16, []> var_1199_to_fp16 = const()[name = tensor<string, []>("op_1199_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_137_cast_fp16 = sub(x = x_135_cast_fp16, y = var_1199_to_fp16)[name = tensor<string, []>("x_137_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1201_cast_fp16 = relu(x = x_137_cast_fp16)[name = tensor<string, []>("op_1201_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1202_cast_fp16 = abs(x = x_137_cast_fp16)[name = tensor<string, []>("op_1202_cast_fp16")];
            tensor<fp16, []> const_26_promoted_to_fp16 = const()[name = tensor<string, []>("const_26_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_1203_cast_fp16 = mul(x = var_1202_cast_fp16, y = const_26_promoted_to_fp16)[name = tensor<string, []>("op_1203_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1204_cast_fp16 = exp(x = var_1203_cast_fp16)[name = tensor<string, []>("op_1204_cast_fp16")];
            tensor<fp32, []> var_1205_epsilon_0 = const()[name = tensor<string, []>("op_1205_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_1205_cast_fp16 = log(epsilon = var_1205_epsilon_0, x = var_1204_cast_fp16)[name = tensor<string, []>("op_1205_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1206_cast_fp16 = add(x = var_1201_cast_fp16, y = var_1205_cast_fp16)[name = tensor<string, []>("op_1206_cast_fp16")];
            tensor<fp16, []> var_1207_to_fp16 = const()[name = tensor<string, []>("op_1207_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_1208_cast_fp16 = mul(x = x_135_cast_fp16, y = var_1207_to_fp16)[name = tensor<string, []>("op_1208_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1209_cast_fp16 = sub(x = var_1206_cast_fp16, y = var_1208_cast_fp16)[name = tensor<string, []>("op_1209_cast_fp16")];
            tensor<fp16, []> var_1210_to_fp16 = const()[name = tensor<string, []>("op_1210_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_99_cast_fp16 = sub(x = var_1209_cast_fp16, y = var_1210_to_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<string, []> var_1218_pad_type_0 = const()[name = tensor<string, []>("op_1218_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1218_strides_0 = const()[name = tensor<string, []>("op_1218_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1218_pad_0 = const()[name = tensor<string, []>("op_1218_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1218_dilations_0 = const()[name = tensor<string, []>("op_1218_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1218_groups_0 = const()[name = tensor<string, []>("op_1218_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_1_stack_layers_0_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124990464)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125514816)))];
            tensor<fp16, [1, 512, 1, 512]> var_1218_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_conv_module2_out_proj_bias_to_fp16, dilations = var_1218_dilations_0, groups = var_1218_groups_0, pad = var_1218_pad_0, pad_type = var_1218_pad_type_0, strides = var_1218_strides_0, weight = core_stacks_1_stack_layers_0_conv_module2_out_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("op_1218_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_101_cast_fp16 = add(x = x_129_cast_fp16, y = var_1218_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
            tensor<string, []> x_139_pad_type_0 = const()[name = tensor<string, []>("x_139_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_139_strides_0 = const()[name = tensor<string, []>("x_139_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_139_pad_0 = const()[name = tensor<string, []>("x_139_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_139_dilations_0 = const()[name = tensor<string, []>("x_139_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_139_groups_0 = const()[name = tensor<string, []>("x_139_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_1_stack_layers_0_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125515904)))];
            tensor<fp16, [1920]> core_stacks_1_stack_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127482048)))];
            tensor<fp16, [1, 1920, 1, 512]> x_139_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_feed_forward3_in_proj_bias_to_fp16, dilations = x_139_dilations_0, groups = x_139_groups_0, pad = x_139_pad_0, pad_type = x_139_pad_type_0, strides = x_139_strides_0, weight = core_stacks_1_stack_layers_0_feed_forward3_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("x_139_cast_fp16")];
            tensor<fp16, []> var_1229_to_fp16 = const()[name = tensor<string, []>("op_1229_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 512]> x_141_cast_fp16 = sub(x = x_139_cast_fp16, y = var_1229_to_fp16)[name = tensor<string, []>("x_141_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_1231_cast_fp16 = relu(x = x_141_cast_fp16)[name = tensor<string, []>("op_1231_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_1232_cast_fp16 = abs(x = x_141_cast_fp16)[name = tensor<string, []>("op_1232_cast_fp16")];
            tensor<fp16, []> const_27_promoted_to_fp16 = const()[name = tensor<string, []>("const_27_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_1233_cast_fp16 = mul(x = var_1232_cast_fp16, y = const_27_promoted_to_fp16)[name = tensor<string, []>("op_1233_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_1234_cast_fp16 = exp(x = var_1233_cast_fp16)[name = tensor<string, []>("op_1234_cast_fp16")];
            tensor<fp32, []> var_1235_epsilon_0 = const()[name = tensor<string, []>("op_1235_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_1235_cast_fp16 = log(epsilon = var_1235_epsilon_0, x = var_1234_cast_fp16)[name = tensor<string, []>("op_1235_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_1236_cast_fp16 = add(x = var_1231_cast_fp16, y = var_1235_cast_fp16)[name = tensor<string, []>("op_1236_cast_fp16")];
            tensor<fp16, []> var_1237_to_fp16 = const()[name = tensor<string, []>("op_1237_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 512]> var_1238_cast_fp16 = mul(x = x_139_cast_fp16, y = var_1237_to_fp16)[name = tensor<string, []>("op_1238_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_1239_cast_fp16 = sub(x = var_1236_cast_fp16, y = var_1238_cast_fp16)[name = tensor<string, []>("op_1239_cast_fp16")];
            tensor<fp16, []> var_1240_to_fp16 = const()[name = tensor<string, []>("op_1240_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 512]> input_103_cast_fp16 = sub(x = var_1239_cast_fp16, y = var_1240_to_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<string, []> var_1248_pad_type_0 = const()[name = tensor<string, []>("op_1248_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1248_strides_0 = const()[name = tensor<string, []>("op_1248_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1248_pad_0 = const()[name = tensor<string, []>("op_1248_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1248_dilations_0 = const()[name = tensor<string, []>("op_1248_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1248_groups_0 = const()[name = tensor<string, []>("op_1248_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_1_stack_layers_0_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(127485952)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129452096)))];
            tensor<fp16, [1, 512, 1, 512]> var_1248_cast_fp16 = conv(bias = core_stacks_1_stack_layers_0_feed_forward3_out_proj_bias_to_fp16, dilations = var_1248_dilations_0, groups = var_1248_groups_0, pad = var_1248_pad_0, pad_type = var_1248_pad_type_0, strides = var_1248_strides_0, weight = core_stacks_1_stack_layers_0_feed_forward3_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("op_1248_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_143_cast_fp16 = add(x = input_101_cast_fp16, y = var_1248_cast_fp16)[name = tensor<string, []>("x_143_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_1_stack_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129453184)))];
            tensor<fp16, [1, 512, 1, 512]> d_5_cast_fp16 = sub(x = x_143_cast_fp16, y = core_stacks_1_stack_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("d_5_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1253_cast_fp16 = mul(x = d_5_cast_fp16, y = d_5_cast_fp16)[name = tensor<string, []>("op_1253_cast_fp16")];
            tensor<int32, [1]> var_1255_axes_0 = const()[name = tensor<string, []>("op_1255_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_1255_keep_dims_0 = const()[name = tensor<string, []>("op_1255_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 512]> var_1255_cast_fp16 = reduce_mean(axes = var_1255_axes_0, keep_dims = var_1255_keep_dims_0, x = var_1253_cast_fp16)[name = tensor<string, []>("op_1255_cast_fp16")];
            tensor<fp32, []> scales_5_epsilon_0 = const()[name = tensor<string, []>("scales_5_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 512]> scales_5_cast_fp16 = rsqrt(epsilon = scales_5_epsilon_0, x = var_1255_cast_fp16)[name = tensor<string, []>("scales_5_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1257_cast_fp16 = mul(x = x_143_cast_fp16, y = scales_5_cast_fp16)[name = tensor<string, []>("op_1257_cast_fp16")];
            tensor<fp16, []> core_stacks_1_stack_layers_0_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.af8p-2)];
            tensor<fp16, [1, 512, 1, 512]> src_11_cast_fp16 = mul(x = var_1257_cast_fp16, y = core_stacks_1_stack_layers_0_norm_scale_to_fp16)[name = tensor<string, []>("src_11_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1260_cast_fp16 = sub(x = src_11_cast_fp16, y = input_75_cast_fp16)[name = tensor<string, []>("op_1260_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_1_stack_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129454272)))];
            tensor<fp16, [1, 512, 1, 512]> var_1261_cast_fp16 = mul(x = var_1260_cast_fp16, y = core_stacks_1_stack_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_1261_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_105_cast_fp16 = add(x = input_75_cast_fp16, y = var_1261_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<string, []> proj_7_pad_type_0 = const()[name = tensor<string, []>("proj_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_7_strides_0 = const()[name = tensor<string, []>("proj_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_7_pad_0 = const()[name = tensor<string, []>("proj_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_7_dilations_0 = const()[name = tensor<string, []>("proj_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_7_groups_0 = const()[name = tensor<string, []>("proj_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_1_stack_layers_1_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129455360)))];
            tensor<fp16, [384]> core_stacks_1_stack_layers_1_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129848640)))];
            tensor<fp16, [1, 384, 1, 512]> proj_7_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_7_dilations_0, groups = proj_7_groups_0, pad = proj_7_pad_0, pad_type = proj_7_pad_type_0, strides = proj_7_strides_0, weight = core_stacks_1_stack_layers_1_self_attn_weights_in_proj_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("proj_7_cast_fp16")];
            tensor<int32, [4]> var_1284_begin_0 = const()[name = tensor<string, []>("op_1284_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1284_end_0 = const()[name = tensor<string, []>("op_1284_end_0"), val = tensor<int32, [4]>([1, 128, 1, 512])];
            tensor<bool, [4]> var_1284_end_mask_0 = const()[name = tensor<string, []>("op_1284_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_1284_cast_fp16 = slice_by_index(begin = var_1284_begin_0, end = var_1284_end_0, end_mask = var_1284_end_mask_0, x = proj_7_cast_fp16)[name = tensor<string, []>("op_1284_cast_fp16")];
            tensor<int32, [4]> var_1285 = const()[name = tensor<string, []>("op_1285"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_1286_cast_fp16 = reshape(shape = var_1285, x = var_1284_cast_fp16)[name = tensor<string, []>("op_1286_cast_fp16")];
            tensor<int32, [4]> var_1288_begin_0 = const()[name = tensor<string, []>("op_1288_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_1288_end_0 = const()[name = tensor<string, []>("op_1288_end_0"), val = tensor<int32, [4]>([1, 256, 1, 512])];
            tensor<bool, [4]> var_1288_end_mask_0 = const()[name = tensor<string, []>("op_1288_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_1288_cast_fp16 = slice_by_index(begin = var_1288_begin_0, end = var_1288_end_0, end_mask = var_1288_end_mask_0, x = proj_7_cast_fp16)[name = tensor<string, []>("op_1288_cast_fp16")];
            tensor<int32, [4]> var_1289 = const()[name = tensor<string, []>("op_1289"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_1290_cast_fp16 = reshape(shape = var_1289, x = var_1288_cast_fp16)[name = tensor<string, []>("op_1290_cast_fp16")];
            tensor<int32, [4]> var_1292_begin_0 = const()[name = tensor<string, []>("op_1292_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_1292_end_0 = const()[name = tensor<string, []>("op_1292_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_1292_end_mask_0 = const()[name = tensor<string, []>("op_1292_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_1292_cast_fp16 = slice_by_index(begin = var_1292_begin_0, end = var_1292_end_0, end_mask = var_1292_end_mask_0, x = proj_7_cast_fp16)[name = tensor<string, []>("op_1292_cast_fp16")];
            tensor<int32, [4]> var_1293 = const()[name = tensor<string, []>("op_1293"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> pt_7_cast_fp16 = reshape(shape = var_1293, x = var_1292_cast_fp16)[name = tensor<string, []>("pt_7_cast_fp16")];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = tensor<string, []>("transpose_24_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_124 = const()[name = tensor<string, []>("concat_124"), val = tensor<int32, [3]>([4, 512, 32])];
            tensor<fp16, [4, 1, 512, 32]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = var_1286_cast_fp16)[name = tensor<string, []>("transpose_284")];
            tensor<fp16, [4, 512, 32]> reshape_36_cast_fp16 = reshape(shape = concat_124, x = transpose_24_cast_fp16)[name = tensor<string, []>("reshape_36_cast_fp16")];
            tensor<int32, [4]> transpose_25_perm_0 = const()[name = tensor<string, []>("transpose_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_125 = const()[name = tensor<string, []>("concat_125"), val = tensor<int32, [3]>([4, 32, 512])];
            tensor<fp16, [4, 1, 32, 512]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = var_1290_cast_fp16)[name = tensor<string, []>("transpose_283")];
            tensor<fp16, [4, 32, 512]> reshape_37_cast_fp16 = reshape(shape = concat_125, x = transpose_25_cast_fp16)[name = tensor<string, []>("reshape_37_cast_fp16")];
            tensor<bool, []> matmul_12_transpose_x_0 = const()[name = tensor<string, []>("matmul_12_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_12_transpose_y_0 = const()[name = tensor<string, []>("matmul_12_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 512]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = reshape_36_cast_fp16, y = reshape_37_cast_fp16)[name = tensor<string, []>("matmul_12_cast_fp16")];
            tensor<int32, [4]> concat_129 = const()[name = tensor<string, []>("concat_129"), val = tensor<int32, [4]>([4, 1, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> reshape_38_cast_fp16 = reshape(shape = concat_129, x = matmul_12_cast_fp16)[name = tensor<string, []>("reshape_38_cast_fp16")];
            tensor<int32, [4]> attn_scores_19_perm_0 = const()[name = tensor<string, []>("attn_scores_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1297 = const()[name = tensor<string, []>("op_1297"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 512, 1]> var_1298_cast_fp16 = transpose(perm = var_1297, x = pt_7_cast_fp16)[name = tensor<string, []>("transpose_281")];
            tensor<fp16, [4, 32, 512, 512]> var_1299_cast_fp16 = mul(x = var_1298_cast_fp16, y = core_pos_basis_512_to_fp16)[name = tensor<string, []>("op_1299_cast_fp16")];
            tensor<int32, [1]> pos_scores_7_axes_0 = const()[name = tensor<string, []>("pos_scores_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_7_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_7_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 512, 512]> pos_scores_7_cast_fp16 = reduce_sum(axes = pos_scores_7_axes_0, keep_dims = pos_scores_7_keep_dims_0, x = var_1299_cast_fp16)[name = tensor<string, []>("pos_scores_7_cast_fp16")];
            tensor<int32, [4]> var_1302 = const()[name = tensor<string, []>("op_1302"), val = tensor<int32, [4]>([4, 512, 1, 512])];
            tensor<fp16, [4, 512, 1, 512]> var_1303_cast_fp16 = reshape(shape = var_1302, x = pos_scores_7_cast_fp16)[name = tensor<string, []>("op_1303_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_19_cast_fp16 = transpose(perm = attn_scores_19_perm_0, x = reshape_38_cast_fp16)[name = tensor<string, []>("transpose_282")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_21_cast_fp16 = add(x = attn_scores_19_cast_fp16, y = var_1303_cast_fp16)[name = tensor<string, []>("attn_scores_21_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_23_cast_fp16 = add(x = attn_scores_21_cast_fp16, y = attn_bias_3_cast_fp16)[name = tensor<string, []>("attn_scores_23_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_weights_7_cast_fp16 = softmax(axis = var_41, x = attn_scores_23_cast_fp16)[name = tensor<string, []>("attn_weights_7_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_107_cast_fp16 = add(x = input_105_cast_fp16, y = time_emb_3_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<string, []> x_145_pad_type_0 = const()[name = tensor<string, []>("x_145_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_145_strides_0 = const()[name = tensor<string, []>("x_145_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_145_pad_0 = const()[name = tensor<string, []>("x_145_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_145_dilations_0 = const()[name = tensor<string, []>("x_145_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_145_groups_0 = const()[name = tensor<string, []>("x_145_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_1_stack_layers_1_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(129849472)))];
            tensor<fp16, [1152]> core_stacks_1_stack_layers_1_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131029184)))];
            tensor<fp16, [1, 1152, 1, 512]> x_145_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_feed_forward1_in_proj_bias_to_fp16, dilations = x_145_dilations_0, groups = x_145_groups_0, pad = x_145_pad_0, pad_type = x_145_pad_type_0, strides = x_145_strides_0, weight = core_stacks_1_stack_layers_1_feed_forward1_in_proj_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("x_145_cast_fp16")];
            tensor<fp16, []> var_1317_to_fp16 = const()[name = tensor<string, []>("op_1317_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 512]> x_147_cast_fp16 = sub(x = x_145_cast_fp16, y = var_1317_to_fp16)[name = tensor<string, []>("x_147_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_1319_cast_fp16 = relu(x = x_147_cast_fp16)[name = tensor<string, []>("op_1319_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_1320_cast_fp16 = abs(x = x_147_cast_fp16)[name = tensor<string, []>("op_1320_cast_fp16")];
            tensor<fp16, []> const_28_promoted_to_fp16 = const()[name = tensor<string, []>("const_28_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_1321_cast_fp16 = mul(x = var_1320_cast_fp16, y = const_28_promoted_to_fp16)[name = tensor<string, []>("op_1321_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_1322_cast_fp16 = exp(x = var_1321_cast_fp16)[name = tensor<string, []>("op_1322_cast_fp16")];
            tensor<fp32, []> var_1323_epsilon_0 = const()[name = tensor<string, []>("op_1323_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_1323_cast_fp16 = log(epsilon = var_1323_epsilon_0, x = var_1322_cast_fp16)[name = tensor<string, []>("op_1323_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_1324_cast_fp16 = add(x = var_1319_cast_fp16, y = var_1323_cast_fp16)[name = tensor<string, []>("op_1324_cast_fp16")];
            tensor<fp16, []> var_1325_to_fp16 = const()[name = tensor<string, []>("op_1325_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 512]> var_1326_cast_fp16 = mul(x = x_145_cast_fp16, y = var_1325_to_fp16)[name = tensor<string, []>("op_1326_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_1327_cast_fp16 = sub(x = var_1324_cast_fp16, y = var_1326_cast_fp16)[name = tensor<string, []>("op_1327_cast_fp16")];
            tensor<fp16, []> var_1328_to_fp16 = const()[name = tensor<string, []>("op_1328_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 512]> input_109_cast_fp16 = sub(x = var_1327_cast_fp16, y = var_1328_to_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<string, []> var_1336_pad_type_0 = const()[name = tensor<string, []>("op_1336_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1336_strides_0 = const()[name = tensor<string, []>("op_1336_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1336_pad_0 = const()[name = tensor<string, []>("op_1336_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1336_dilations_0 = const()[name = tensor<string, []>("op_1336_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1336_groups_0 = const()[name = tensor<string, []>("op_1336_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_1_stack_layers_1_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131031552)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_1_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132211264)))];
            tensor<fp16, [1, 512, 1, 512]> var_1336_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_feed_forward1_out_proj_bias_to_fp16, dilations = var_1336_dilations_0, groups = var_1336_groups_0, pad = var_1336_pad_0, pad_type = var_1336_pad_type_0, strides = var_1336_strides_0, weight = core_stacks_1_stack_layers_1_feed_forward1_out_proj_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("op_1336_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_111_cast_fp16 = add(x = input_107_cast_fp16, y = var_1336_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
            tensor<int32, [4]> var_1338_begin_0 = const()[name = tensor<string, []>("op_1338_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1338_end_0 = const()[name = tensor<string, []>("op_1338_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_1338_end_mask_0 = const()[name = tensor<string, []>("op_1338_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_1338_cast_fp16 = slice_by_index(begin = var_1338_begin_0, end = var_1338_end_0, end_mask = var_1338_end_mask_0, x = attn_weights_7_cast_fp16)[name = tensor<string, []>("op_1338_cast_fp16")];
            tensor<string, []> x_149_pad_type_0 = const()[name = tensor<string, []>("x_149_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_149_strides_0 = const()[name = tensor<string, []>("x_149_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_149_pad_0 = const()[name = tensor<string, []>("x_149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_149_dilations_0 = const()[name = tensor<string, []>("x_149_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_149_groups_0 = const()[name = tensor<string, []>("x_149_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_1_stack_layers_1_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132212352)))];
            tensor<fp16, [1152]> core_stacks_1_stack_layers_1_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133392064)))];
            tensor<fp16, [1, 1152, 1, 512]> x_149_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_nonlin_attention_in_proj_bias_to_fp16, dilations = x_149_dilations_0, groups = x_149_groups_0, pad = x_149_pad_0, pad_type = x_149_pad_type_0, strides = x_149_strides_0, weight = core_stacks_1_stack_layers_1_nonlin_attention_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("x_149_cast_fp16")];
            tensor<int32, [4]> var_1349_begin_0 = const()[name = tensor<string, []>("op_1349_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1349_end_0 = const()[name = tensor<string, []>("op_1349_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_1349_end_mask_0 = const()[name = tensor<string, []>("op_1349_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_1349_cast_fp16 = slice_by_index(begin = var_1349_begin_0, end = var_1349_end_0, end_mask = var_1349_end_mask_0, x = x_149_cast_fp16)[name = tensor<string, []>("op_1349_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> s_gate_7_cast_fp16 = tanh(x = var_1349_cast_fp16)[name = tensor<string, []>("s_gate_7_cast_fp16")];
            tensor<int32, [4]> var_1352_begin_0 = const()[name = tensor<string, []>("op_1352_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_1352_end_0 = const()[name = tensor<string, []>("op_1352_end_0"), val = tensor<int32, [4]>([1, 768, 1, 512])];
            tensor<bool, [4]> var_1352_end_mask_0 = const()[name = tensor<string, []>("op_1352_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_1352_cast_fp16 = slice_by_index(begin = var_1352_begin_0, end = var_1352_end_0, end_mask = var_1352_end_mask_0, x = x_149_cast_fp16)[name = tensor<string, []>("op_1352_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> var_1353_cast_fp16 = mul(x = var_1352_cast_fp16, y = s_gate_7_cast_fp16)[name = tensor<string, []>("op_1353_cast_fp16")];
            tensor<int32, [4]> y_7_begin_0 = const()[name = tensor<string, []>("y_7_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_7_end_0 = const()[name = tensor<string, []>("y_7_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 512])];
            tensor<bool, [4]> y_7_end_mask_0 = const()[name = tensor<string, []>("y_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 512]> y_7_cast_fp16 = slice_by_index(begin = y_7_begin_0, end = y_7_end_0, end_mask = y_7_end_mask_0, x = x_149_cast_fp16)[name = tensor<string, []>("y_7_cast_fp16")];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = tensor<string, []>("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_134 = const()[name = tensor<string, []>("concat_134"), val = tensor<int32, [3]>([1, 512, 512])];
            tensor<fp16, [1, 1, 512, 512]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = var_1338_cast_fp16)[name = tensor<string, []>("transpose_280")];
            tensor<fp16, [1, 512, 512]> reshape_39_cast_fp16 = reshape(shape = concat_134, x = transpose_26_cast_fp16)[name = tensor<string, []>("reshape_39_cast_fp16")];
            tensor<int32, [4]> transpose_27_perm_0 = const()[name = tensor<string, []>("transpose_27_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_135 = const()[name = tensor<string, []>("concat_135"), val = tensor<int32, [3]>([1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = var_1353_cast_fp16)[name = tensor<string, []>("transpose_279")];
            tensor<fp16, [1, 512, 384]> reshape_40_cast_fp16 = reshape(shape = concat_135, x = transpose_27_cast_fp16)[name = tensor<string, []>("reshape_40_cast_fp16")];
            tensor<bool, []> matmul_13_transpose_x_0 = const()[name = tensor<string, []>("matmul_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_13_transpose_y_0 = const()[name = tensor<string, []>("matmul_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 384]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = reshape_39_cast_fp16, y = reshape_40_cast_fp16)[name = tensor<string, []>("matmul_13_cast_fp16")];
            tensor<int32, [4]> concat_139 = const()[name = tensor<string, []>("concat_139"), val = tensor<int32, [4]>([1, 1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> reshape_41_cast_fp16 = reshape(shape = concat_139, x = matmul_13_cast_fp16)[name = tensor<string, []>("reshape_41_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = tensor<string, []>("v_7_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 512]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = reshape_41_cast_fp16)[name = tensor<string, []>("transpose_278")];
            tensor<fp16, [1, 384, 1, 512]> input_113_cast_fp16 = mul(x = v_7_cast_fp16, y = y_7_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
            tensor<string, []> var_1365_pad_type_0 = const()[name = tensor<string, []>("op_1365_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1365_strides_0 = const()[name = tensor<string, []>("op_1365_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1365_pad_0 = const()[name = tensor<string, []>("op_1365_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1365_dilations_0 = const()[name = tensor<string, []>("op_1365_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1365_groups_0 = const()[name = tensor<string, []>("op_1365_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_1_stack_layers_1_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133394432)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_1_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133787712)))];
            tensor<fp16, [1, 512, 1, 512]> var_1365_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_nonlin_attention_out_proj_bias_to_fp16, dilations = var_1365_dilations_0, groups = var_1365_groups_0, pad = var_1365_pad_0, pad_type = var_1365_pad_type_0, strides = var_1365_strides_0, weight = core_stacks_1_stack_layers_1_nonlin_attention_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("op_1365_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_151_cast_fp16 = add(x = input_111_cast_fp16, y = var_1365_cast_fp16)[name = tensor<string, []>("x_151_cast_fp16")];
            tensor<string, []> var_1376_pad_type_0 = const()[name = tensor<string, []>("op_1376_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1376_strides_0 = const()[name = tensor<string, []>("op_1376_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1376_pad_0 = const()[name = tensor<string, []>("op_1376_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1376_dilations_0 = const()[name = tensor<string, []>("op_1376_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1376_groups_0 = const()[name = tensor<string, []>("op_1376_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_1_stack_layers_1_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133788800)))];
            tensor<fp16, [48]> core_stacks_1_stack_layers_1_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133838016)))];
            tensor<fp16, [1, 48, 1, 512]> var_1376_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_self_attn1_in_proj_bias_to_fp16, dilations = var_1376_dilations_0, groups = var_1376_groups_0, pad = var_1376_pad_0, pad_type = var_1376_pad_type_0, strides = var_1376_strides_0, weight = core_stacks_1_stack_layers_1_self_attn1_in_proj_weight_to_fp16, x = x_151_cast_fp16)[name = tensor<string, []>("op_1376_cast_fp16")];
            tensor<int32, [4]> var_1377 = const()[name = tensor<string, []>("op_1377"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_1378_cast_fp16 = reshape(shape = var_1377, x = var_1376_cast_fp16)[name = tensor<string, []>("op_1378_cast_fp16")];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = tensor<string, []>("transpose_28_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_144 = const()[name = tensor<string, []>("concat_144"), val = tensor<int32, [3]>([4, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = attn_weights_7_cast_fp16)[name = tensor<string, []>("transpose_277")];
            tensor<fp16, [4, 512, 512]> reshape_42_cast_fp16 = reshape(shape = concat_144, x = transpose_28_cast_fp16)[name = tensor<string, []>("reshape_42_cast_fp16")];
            tensor<int32, [4]> transpose_29_perm_0 = const()[name = tensor<string, []>("transpose_29_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_145 = const()[name = tensor<string, []>("concat_145"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = var_1378_cast_fp16)[name = tensor<string, []>("transpose_276")];
            tensor<fp16, [4, 512, 12]> reshape_43_cast_fp16 = reshape(shape = concat_145, x = transpose_29_cast_fp16)[name = tensor<string, []>("reshape_43_cast_fp16")];
            tensor<bool, []> matmul_14_transpose_x_0 = const()[name = tensor<string, []>("matmul_14_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_14_transpose_y_0 = const()[name = tensor<string, []>("matmul_14_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = reshape_42_cast_fp16, y = reshape_43_cast_fp16)[name = tensor<string, []>("matmul_14_cast_fp16")];
            tensor<int32, [4]> concat_149 = const()[name = tensor<string, []>("concat_149"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_44_cast_fp16 = reshape(shape = concat_149, x = matmul_14_cast_fp16)[name = tensor<string, []>("reshape_44_cast_fp16")];
            tensor<int32, [4]> out_13_perm_0 = const()[name = tensor<string, []>("out_13_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_1381 = const()[name = tensor<string, []>("op_1381"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_13_cast_fp16 = transpose(perm = out_13_perm_0, x = reshape_44_cast_fp16)[name = tensor<string, []>("transpose_275")];
            tensor<fp16, [1, 48, 1, 512]> input_115_cast_fp16 = reshape(shape = var_1381, x = out_13_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<string, []> var_1389_pad_type_0 = const()[name = tensor<string, []>("op_1389_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1389_strides_0 = const()[name = tensor<string, []>("op_1389_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1389_pad_0 = const()[name = tensor<string, []>("op_1389_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1389_dilations_0 = const()[name = tensor<string, []>("op_1389_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1389_groups_0 = const()[name = tensor<string, []>("op_1389_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_1_stack_layers_1_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133838208)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_1_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133887424)))];
            tensor<fp16, [1, 512, 1, 512]> var_1389_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_self_attn1_out_proj_bias_to_fp16, dilations = var_1389_dilations_0, groups = var_1389_groups_0, pad = var_1389_pad_0, pad_type = var_1389_pad_type_0, strides = var_1389_strides_0, weight = core_stacks_1_stack_layers_1_self_attn1_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("op_1389_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_153_cast_fp16 = add(x = x_151_cast_fp16, y = var_1389_cast_fp16)[name = tensor<string, []>("x_153_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_155_cast_fp16 = add(x = x_153_cast_fp16, y = time_emb_3_cast_fp16)[name = tensor<string, []>("x_155_cast_fp16")];
            tensor<string, []> x_157_pad_type_0 = const()[name = tensor<string, []>("x_157_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_157_strides_0 = const()[name = tensor<string, []>("x_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_157_pad_0 = const()[name = tensor<string, []>("x_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_157_dilations_0 = const()[name = tensor<string, []>("x_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_157_groups_0 = const()[name = tensor<string, []>("x_157_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_1_stack_layers_1_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133888512)))];
            tensor<fp16, [1024]> core_stacks_1_stack_layers_1_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134937152)))];
            tensor<fp16, [1, 1024, 1, 512]> x_157_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_conv_module1_in_proj_bias_to_fp16, dilations = x_157_dilations_0, groups = x_157_groups_0, pad = x_157_pad_0, pad_type = x_157_pad_type_0, strides = x_157_strides_0, weight = core_stacks_1_stack_layers_1_conv_module1_in_proj_weight_to_fp16, x = x_155_cast_fp16)[name = tensor<string, []>("x_157_cast_fp16")];
            tensor<int32, [4]> var_1403_begin_0 = const()[name = tensor<string, []>("op_1403_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1403_end_0 = const()[name = tensor<string, []>("op_1403_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_1403_end_mask_0 = const()[name = tensor<string, []>("op_1403_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_1403_cast_fp16 = slice_by_index(begin = var_1403_begin_0, end = var_1403_end_0, end_mask = var_1403_end_mask_0, x = x_157_cast_fp16)[name = tensor<string, []>("op_1403_cast_fp16")];
            tensor<int32, [4]> var_1405_begin_0 = const()[name = tensor<string, []>("op_1405_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_1405_end_0 = const()[name = tensor<string, []>("op_1405_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_1405_end_mask_0 = const()[name = tensor<string, []>("op_1405_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_1405_cast_fp16 = slice_by_index(begin = var_1405_begin_0, end = var_1405_end_0, end_mask = var_1405_end_mask_0, x = x_157_cast_fp16)[name = tensor<string, []>("op_1405_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1406_cast_fp16 = sigmoid(x = var_1405_cast_fp16)[name = tensor<string, []>("op_1406_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_159_cast_fp16 = mul(x = var_1403_cast_fp16, y = var_1406_cast_fp16)[name = tensor<string, []>("x_159_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_117_cast_fp16 = mul(x = x_159_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")];
            tensor<string, []> x_161_pad_type_0 = const()[name = tensor<string, []>("x_161_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_161_pad_0 = const()[name = tensor<string, []>("x_161_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_161_groups_0 = const()[name = tensor<string, []>("x_161_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_161_strides_0 = const()[name = tensor<string, []>("x_161_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_161_dilations_0 = const()[name = tensor<string, []>("x_161_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_1_stack_layers_1_conv_module1_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module1_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134939264)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_1_conv_module1_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module1_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134954688)))];
            tensor<fp16, [1, 512, 1, 512]> x_161_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_conv_module1_dw_bias_to_fp16, dilations = x_161_dilations_0, groups = x_161_groups_0, pad = x_161_pad_0, pad_type = x_161_pad_type_0, strides = x_161_strides_0, weight = core_stacks_1_stack_layers_1_conv_module1_dw_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("x_161_cast_fp16")];
            tensor<fp16, []> var_1416_to_fp16 = const()[name = tensor<string, []>("op_1416_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_163_cast_fp16 = sub(x = x_161_cast_fp16, y = var_1416_to_fp16)[name = tensor<string, []>("x_163_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1418_cast_fp16 = relu(x = x_163_cast_fp16)[name = tensor<string, []>("op_1418_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1419_cast_fp16 = abs(x = x_163_cast_fp16)[name = tensor<string, []>("op_1419_cast_fp16")];
            tensor<fp16, []> const_30_promoted_to_fp16 = const()[name = tensor<string, []>("const_30_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_1420_cast_fp16 = mul(x = var_1419_cast_fp16, y = const_30_promoted_to_fp16)[name = tensor<string, []>("op_1420_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1421_cast_fp16 = exp(x = var_1420_cast_fp16)[name = tensor<string, []>("op_1421_cast_fp16")];
            tensor<fp32, []> var_1422_epsilon_0 = const()[name = tensor<string, []>("op_1422_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_1422_cast_fp16 = log(epsilon = var_1422_epsilon_0, x = var_1421_cast_fp16)[name = tensor<string, []>("op_1422_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1423_cast_fp16 = add(x = var_1418_cast_fp16, y = var_1422_cast_fp16)[name = tensor<string, []>("op_1423_cast_fp16")];
            tensor<fp16, []> var_1424_to_fp16 = const()[name = tensor<string, []>("op_1424_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_1425_cast_fp16 = mul(x = x_161_cast_fp16, y = var_1424_to_fp16)[name = tensor<string, []>("op_1425_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1426_cast_fp16 = sub(x = var_1423_cast_fp16, y = var_1425_cast_fp16)[name = tensor<string, []>("op_1426_cast_fp16")];
            tensor<fp16, []> var_1427_to_fp16 = const()[name = tensor<string, []>("op_1427_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_119_cast_fp16 = sub(x = var_1426_cast_fp16, y = var_1427_to_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
            tensor<string, []> var_1435_pad_type_0 = const()[name = tensor<string, []>("op_1435_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1435_strides_0 = const()[name = tensor<string, []>("op_1435_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1435_pad_0 = const()[name = tensor<string, []>("op_1435_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1435_dilations_0 = const()[name = tensor<string, []>("op_1435_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1435_groups_0 = const()[name = tensor<string, []>("op_1435_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_1_stack_layers_1_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134955776)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_1_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135480128)))];
            tensor<fp16, [1, 512, 1, 512]> var_1435_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_conv_module1_out_proj_bias_to_fp16, dilations = var_1435_dilations_0, groups = var_1435_groups_0, pad = var_1435_pad_0, pad_type = var_1435_pad_type_0, strides = var_1435_strides_0, weight = core_stacks_1_stack_layers_1_conv_module1_out_proj_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("op_1435_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_121_cast_fp16 = add(x = x_155_cast_fp16, y = var_1435_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<string, []> x_165_pad_type_0 = const()[name = tensor<string, []>("x_165_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_165_strides_0 = const()[name = tensor<string, []>("x_165_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_165_pad_0 = const()[name = tensor<string, []>("x_165_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_165_dilations_0 = const()[name = tensor<string, []>("x_165_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_165_groups_0 = const()[name = tensor<string, []>("x_165_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_1_stack_layers_1_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135481216)))];
            tensor<fp16, [1536]> core_stacks_1_stack_layers_1_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137054144)))];
            tensor<fp16, [1, 1536, 1, 512]> x_165_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_feed_forward2_in_proj_bias_to_fp16, dilations = x_165_dilations_0, groups = x_165_groups_0, pad = x_165_pad_0, pad_type = x_165_pad_type_0, strides = x_165_strides_0, weight = core_stacks_1_stack_layers_1_feed_forward2_in_proj_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("x_165_cast_fp16")];
            tensor<fp16, []> var_1446_to_fp16 = const()[name = tensor<string, []>("op_1446_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 512]> x_167_cast_fp16 = sub(x = x_165_cast_fp16, y = var_1446_to_fp16)[name = tensor<string, []>("x_167_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_1448_cast_fp16 = relu(x = x_167_cast_fp16)[name = tensor<string, []>("op_1448_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_1449_cast_fp16 = abs(x = x_167_cast_fp16)[name = tensor<string, []>("op_1449_cast_fp16")];
            tensor<fp16, []> const_31_promoted_to_fp16 = const()[name = tensor<string, []>("const_31_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_1450_cast_fp16 = mul(x = var_1449_cast_fp16, y = const_31_promoted_to_fp16)[name = tensor<string, []>("op_1450_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_1451_cast_fp16 = exp(x = var_1450_cast_fp16)[name = tensor<string, []>("op_1451_cast_fp16")];
            tensor<fp32, []> var_1452_epsilon_0 = const()[name = tensor<string, []>("op_1452_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_1452_cast_fp16 = log(epsilon = var_1452_epsilon_0, x = var_1451_cast_fp16)[name = tensor<string, []>("op_1452_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_1453_cast_fp16 = add(x = var_1448_cast_fp16, y = var_1452_cast_fp16)[name = tensor<string, []>("op_1453_cast_fp16")];
            tensor<fp16, []> var_1454_to_fp16 = const()[name = tensor<string, []>("op_1454_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 512]> var_1455_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1454_to_fp16)[name = tensor<string, []>("op_1455_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_1456_cast_fp16 = sub(x = var_1453_cast_fp16, y = var_1455_cast_fp16)[name = tensor<string, []>("op_1456_cast_fp16")];
            tensor<fp16, []> var_1457_to_fp16 = const()[name = tensor<string, []>("op_1457_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 512]> input_123_cast_fp16 = sub(x = var_1456_cast_fp16, y = var_1457_to_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
            tensor<string, []> var_1465_pad_type_0 = const()[name = tensor<string, []>("op_1465_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1465_strides_0 = const()[name = tensor<string, []>("op_1465_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1465_pad_0 = const()[name = tensor<string, []>("op_1465_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1465_dilations_0 = const()[name = tensor<string, []>("op_1465_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1465_groups_0 = const()[name = tensor<string, []>("op_1465_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_1_stack_layers_1_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137057280)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_1_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138630208)))];
            tensor<fp16, [1, 512, 1, 512]> var_1465_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_feed_forward2_out_proj_bias_to_fp16, dilations = var_1465_dilations_0, groups = var_1465_groups_0, pad = var_1465_pad_0, pad_type = var_1465_pad_type_0, strides = var_1465_strides_0, weight = core_stacks_1_stack_layers_1_feed_forward2_out_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("op_1465_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> src_13_cast_fp16 = add(x = input_121_cast_fp16, y = var_1465_cast_fp16)[name = tensor<string, []>("src_13_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1468_cast_fp16 = sub(x = src_13_cast_fp16, y = input_105_cast_fp16)[name = tensor<string, []>("op_1468_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_1_stack_layers_1_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138631296)))];
            tensor<fp16, [1, 512, 1, 512]> var_1469_cast_fp16 = mul(x = var_1468_cast_fp16, y = core_stacks_1_stack_layers_1_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_1469_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_169_cast_fp16 = add(x = input_105_cast_fp16, y = var_1469_cast_fp16)[name = tensor<string, []>("x_169_cast_fp16")];
            tensor<string, []> var_1480_pad_type_0 = const()[name = tensor<string, []>("op_1480_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1480_strides_0 = const()[name = tensor<string, []>("op_1480_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1480_pad_0 = const()[name = tensor<string, []>("op_1480_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1480_dilations_0 = const()[name = tensor<string, []>("op_1480_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1480_groups_0 = const()[name = tensor<string, []>("op_1480_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_1_stack_layers_1_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138632384)))];
            tensor<fp16, [48]> core_stacks_1_stack_layers_1_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138681600)))];
            tensor<fp16, [1, 48, 1, 512]> var_1480_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_self_attn2_in_proj_bias_to_fp16, dilations = var_1480_dilations_0, groups = var_1480_groups_0, pad = var_1480_pad_0, pad_type = var_1480_pad_type_0, strides = var_1480_strides_0, weight = core_stacks_1_stack_layers_1_self_attn2_in_proj_weight_to_fp16, x = x_169_cast_fp16)[name = tensor<string, []>("op_1480_cast_fp16")];
            tensor<int32, [4]> var_1481 = const()[name = tensor<string, []>("op_1481"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_1482_cast_fp16 = reshape(shape = var_1481, x = var_1480_cast_fp16)[name = tensor<string, []>("op_1482_cast_fp16")];
            tensor<int32, [4]> transpose_31_perm_0 = const()[name = tensor<string, []>("transpose_31_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_155 = const()[name = tensor<string, []>("concat_155"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_31_cast_fp16 = transpose(perm = transpose_31_perm_0, x = var_1482_cast_fp16)[name = tensor<string, []>("transpose_274")];
            tensor<fp16, [4, 512, 12]> reshape_46_cast_fp16 = reshape(shape = concat_155, x = transpose_31_cast_fp16)[name = tensor<string, []>("reshape_46_cast_fp16")];
            tensor<bool, []> matmul_15_transpose_x_0 = const()[name = tensor<string, []>("matmul_15_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_15_transpose_y_0 = const()[name = tensor<string, []>("matmul_15_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = reshape_42_cast_fp16, y = reshape_46_cast_fp16)[name = tensor<string, []>("matmul_15_cast_fp16")];
            tensor<int32, [4]> concat_159 = const()[name = tensor<string, []>("concat_159"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_47_cast_fp16 = reshape(shape = concat_159, x = matmul_15_cast_fp16)[name = tensor<string, []>("reshape_47_cast_fp16")];
            tensor<int32, [4]> out_15_perm_0 = const()[name = tensor<string, []>("out_15_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_1485 = const()[name = tensor<string, []>("op_1485"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_15_cast_fp16 = transpose(perm = out_15_perm_0, x = reshape_47_cast_fp16)[name = tensor<string, []>("transpose_273")];
            tensor<fp16, [1, 48, 1, 512]> input_125_cast_fp16 = reshape(shape = var_1485, x = out_15_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
            tensor<string, []> var_1493_pad_type_0 = const()[name = tensor<string, []>("op_1493_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1493_strides_0 = const()[name = tensor<string, []>("op_1493_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1493_pad_0 = const()[name = tensor<string, []>("op_1493_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1493_dilations_0 = const()[name = tensor<string, []>("op_1493_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1493_groups_0 = const()[name = tensor<string, []>("op_1493_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_1_stack_layers_1_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138681792)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_1_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138731008)))];
            tensor<fp16, [1, 512, 1, 512]> var_1493_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_self_attn2_out_proj_bias_to_fp16, dilations = var_1493_dilations_0, groups = var_1493_groups_0, pad = var_1493_pad_0, pad_type = var_1493_pad_type_0, strides = var_1493_strides_0, weight = core_stacks_1_stack_layers_1_self_attn2_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor<string, []>("op_1493_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_171_cast_fp16 = add(x = x_169_cast_fp16, y = var_1493_cast_fp16)[name = tensor<string, []>("x_171_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_173_cast_fp16 = add(x = x_171_cast_fp16, y = time_emb_3_cast_fp16)[name = tensor<string, []>("x_173_cast_fp16")];
            tensor<string, []> x_175_pad_type_0 = const()[name = tensor<string, []>("x_175_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_175_strides_0 = const()[name = tensor<string, []>("x_175_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_175_pad_0 = const()[name = tensor<string, []>("x_175_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_175_dilations_0 = const()[name = tensor<string, []>("x_175_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_175_groups_0 = const()[name = tensor<string, []>("x_175_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_1_stack_layers_1_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138732096)))];
            tensor<fp16, [1024]> core_stacks_1_stack_layers_1_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139780736)))];
            tensor<fp16, [1, 1024, 1, 512]> x_175_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_conv_module2_in_proj_bias_to_fp16, dilations = x_175_dilations_0, groups = x_175_groups_0, pad = x_175_pad_0, pad_type = x_175_pad_type_0, strides = x_175_strides_0, weight = core_stacks_1_stack_layers_1_conv_module2_in_proj_weight_to_fp16, x = x_173_cast_fp16)[name = tensor<string, []>("x_175_cast_fp16")];
            tensor<int32, [4]> var_1507_begin_0 = const()[name = tensor<string, []>("op_1507_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1507_end_0 = const()[name = tensor<string, []>("op_1507_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_1507_end_mask_0 = const()[name = tensor<string, []>("op_1507_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_1507_cast_fp16 = slice_by_index(begin = var_1507_begin_0, end = var_1507_end_0, end_mask = var_1507_end_mask_0, x = x_175_cast_fp16)[name = tensor<string, []>("op_1507_cast_fp16")];
            tensor<int32, [4]> var_1509_begin_0 = const()[name = tensor<string, []>("op_1509_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_1509_end_0 = const()[name = tensor<string, []>("op_1509_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_1509_end_mask_0 = const()[name = tensor<string, []>("op_1509_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_1509_cast_fp16 = slice_by_index(begin = var_1509_begin_0, end = var_1509_end_0, end_mask = var_1509_end_mask_0, x = x_175_cast_fp16)[name = tensor<string, []>("op_1509_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1510_cast_fp16 = sigmoid(x = var_1509_cast_fp16)[name = tensor<string, []>("op_1510_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_177_cast_fp16 = mul(x = var_1507_cast_fp16, y = var_1510_cast_fp16)[name = tensor<string, []>("x_177_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_127_cast_fp16 = mul(x = x_177_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<string, []> x_179_pad_type_0 = const()[name = tensor<string, []>("x_179_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_179_pad_0 = const()[name = tensor<string, []>("x_179_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_179_groups_0 = const()[name = tensor<string, []>("x_179_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_179_strides_0 = const()[name = tensor<string, []>("x_179_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_179_dilations_0 = const()[name = tensor<string, []>("x_179_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_1_stack_layers_1_conv_module2_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module2_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139782848)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_1_conv_module2_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module2_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139798272)))];
            tensor<fp16, [1, 512, 1, 512]> x_179_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_conv_module2_dw_bias_to_fp16, dilations = x_179_dilations_0, groups = x_179_groups_0, pad = x_179_pad_0, pad_type = x_179_pad_type_0, strides = x_179_strides_0, weight = core_stacks_1_stack_layers_1_conv_module2_dw_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("x_179_cast_fp16")];
            tensor<fp16, []> var_1520_to_fp16 = const()[name = tensor<string, []>("op_1520_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_181_cast_fp16 = sub(x = x_179_cast_fp16, y = var_1520_to_fp16)[name = tensor<string, []>("x_181_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1522_cast_fp16 = relu(x = x_181_cast_fp16)[name = tensor<string, []>("op_1522_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1523_cast_fp16 = abs(x = x_181_cast_fp16)[name = tensor<string, []>("op_1523_cast_fp16")];
            tensor<fp16, []> const_33_promoted_to_fp16 = const()[name = tensor<string, []>("const_33_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_1524_cast_fp16 = mul(x = var_1523_cast_fp16, y = const_33_promoted_to_fp16)[name = tensor<string, []>("op_1524_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1525_cast_fp16 = exp(x = var_1524_cast_fp16)[name = tensor<string, []>("op_1525_cast_fp16")];
            tensor<fp32, []> var_1526_epsilon_0 = const()[name = tensor<string, []>("op_1526_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_1526_cast_fp16 = log(epsilon = var_1526_epsilon_0, x = var_1525_cast_fp16)[name = tensor<string, []>("op_1526_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1527_cast_fp16 = add(x = var_1522_cast_fp16, y = var_1526_cast_fp16)[name = tensor<string, []>("op_1527_cast_fp16")];
            tensor<fp16, []> var_1528_to_fp16 = const()[name = tensor<string, []>("op_1528_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_1529_cast_fp16 = mul(x = x_179_cast_fp16, y = var_1528_to_fp16)[name = tensor<string, []>("op_1529_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1530_cast_fp16 = sub(x = var_1527_cast_fp16, y = var_1529_cast_fp16)[name = tensor<string, []>("op_1530_cast_fp16")];
            tensor<fp16, []> var_1531_to_fp16 = const()[name = tensor<string, []>("op_1531_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_129_cast_fp16 = sub(x = var_1530_cast_fp16, y = var_1531_to_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<string, []> var_1539_pad_type_0 = const()[name = tensor<string, []>("op_1539_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1539_strides_0 = const()[name = tensor<string, []>("op_1539_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1539_pad_0 = const()[name = tensor<string, []>("op_1539_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1539_dilations_0 = const()[name = tensor<string, []>("op_1539_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1539_groups_0 = const()[name = tensor<string, []>("op_1539_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_1_stack_layers_1_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139799360)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_1_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140323712)))];
            tensor<fp16, [1, 512, 1, 512]> var_1539_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_conv_module2_out_proj_bias_to_fp16, dilations = var_1539_dilations_0, groups = var_1539_groups_0, pad = var_1539_pad_0, pad_type = var_1539_pad_type_0, strides = var_1539_strides_0, weight = core_stacks_1_stack_layers_1_conv_module2_out_proj_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("op_1539_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_131_cast_fp16 = add(x = x_173_cast_fp16, y = var_1539_cast_fp16)[name = tensor<string, []>("input_131_cast_fp16")];
            tensor<string, []> x_183_pad_type_0 = const()[name = tensor<string, []>("x_183_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_183_strides_0 = const()[name = tensor<string, []>("x_183_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_183_pad_0 = const()[name = tensor<string, []>("x_183_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_183_dilations_0 = const()[name = tensor<string, []>("x_183_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_183_groups_0 = const()[name = tensor<string, []>("x_183_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_1_stack_layers_1_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140324800)))];
            tensor<fp16, [1920]> core_stacks_1_stack_layers_1_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142290944)))];
            tensor<fp16, [1, 1920, 1, 512]> x_183_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_feed_forward3_in_proj_bias_to_fp16, dilations = x_183_dilations_0, groups = x_183_groups_0, pad = x_183_pad_0, pad_type = x_183_pad_type_0, strides = x_183_strides_0, weight = core_stacks_1_stack_layers_1_feed_forward3_in_proj_weight_to_fp16, x = input_131_cast_fp16)[name = tensor<string, []>("x_183_cast_fp16")];
            tensor<fp16, []> var_1550_to_fp16 = const()[name = tensor<string, []>("op_1550_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 512]> x_185_cast_fp16 = sub(x = x_183_cast_fp16, y = var_1550_to_fp16)[name = tensor<string, []>("x_185_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_1552_cast_fp16 = relu(x = x_185_cast_fp16)[name = tensor<string, []>("op_1552_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_1553_cast_fp16 = abs(x = x_185_cast_fp16)[name = tensor<string, []>("op_1553_cast_fp16")];
            tensor<fp16, []> const_34_promoted_to_fp16 = const()[name = tensor<string, []>("const_34_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_1554_cast_fp16 = mul(x = var_1553_cast_fp16, y = const_34_promoted_to_fp16)[name = tensor<string, []>("op_1554_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_1555_cast_fp16 = exp(x = var_1554_cast_fp16)[name = tensor<string, []>("op_1555_cast_fp16")];
            tensor<fp32, []> var_1556_epsilon_0 = const()[name = tensor<string, []>("op_1556_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_1556_cast_fp16 = log(epsilon = var_1556_epsilon_0, x = var_1555_cast_fp16)[name = tensor<string, []>("op_1556_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_1557_cast_fp16 = add(x = var_1552_cast_fp16, y = var_1556_cast_fp16)[name = tensor<string, []>("op_1557_cast_fp16")];
            tensor<fp16, []> var_1558_to_fp16 = const()[name = tensor<string, []>("op_1558_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 512]> var_1559_cast_fp16 = mul(x = x_183_cast_fp16, y = var_1558_to_fp16)[name = tensor<string, []>("op_1559_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_1560_cast_fp16 = sub(x = var_1557_cast_fp16, y = var_1559_cast_fp16)[name = tensor<string, []>("op_1560_cast_fp16")];
            tensor<fp16, []> var_1561_to_fp16 = const()[name = tensor<string, []>("op_1561_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 512]> input_133_cast_fp16 = sub(x = var_1560_cast_fp16, y = var_1561_to_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<string, []> var_1569_pad_type_0 = const()[name = tensor<string, []>("op_1569_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1569_strides_0 = const()[name = tensor<string, []>("op_1569_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1569_pad_0 = const()[name = tensor<string, []>("op_1569_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1569_dilations_0 = const()[name = tensor<string, []>("op_1569_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1569_groups_0 = const()[name = tensor<string, []>("op_1569_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_1_stack_layers_1_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142294848)))];
            tensor<fp16, [512]> core_stacks_1_stack_layers_1_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144260992)))];
            tensor<fp16, [1, 512, 1, 512]> var_1569_cast_fp16 = conv(bias = core_stacks_1_stack_layers_1_feed_forward3_out_proj_bias_to_fp16, dilations = var_1569_dilations_0, groups = var_1569_groups_0, pad = var_1569_pad_0, pad_type = var_1569_pad_type_0, strides = var_1569_strides_0, weight = core_stacks_1_stack_layers_1_feed_forward3_out_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("op_1569_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_187_cast_fp16 = add(x = input_131_cast_fp16, y = var_1569_cast_fp16)[name = tensor<string, []>("x_187_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_1_stack_layers_1_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144262080)))];
            tensor<fp16, [1, 512, 1, 512]> d_7_cast_fp16 = sub(x = x_187_cast_fp16, y = core_stacks_1_stack_layers_1_norm_bias_to_fp16)[name = tensor<string, []>("d_7_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1574_cast_fp16 = mul(x = d_7_cast_fp16, y = d_7_cast_fp16)[name = tensor<string, []>("op_1574_cast_fp16")];
            tensor<int32, [1]> var_1576_axes_0 = const()[name = tensor<string, []>("op_1576_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_1576_keep_dims_0 = const()[name = tensor<string, []>("op_1576_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 512]> var_1576_cast_fp16 = reduce_mean(axes = var_1576_axes_0, keep_dims = var_1576_keep_dims_0, x = var_1574_cast_fp16)[name = tensor<string, []>("op_1576_cast_fp16")];
            tensor<fp32, []> scales_7_epsilon_0 = const()[name = tensor<string, []>("scales_7_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 512]> scales_7_cast_fp16 = rsqrt(epsilon = scales_7_epsilon_0, x = var_1576_cast_fp16)[name = tensor<string, []>("scales_7_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1578_cast_fp16 = mul(x = x_187_cast_fp16, y = scales_7_cast_fp16)[name = tensor<string, []>("op_1578_cast_fp16")];
            tensor<fp16, []> core_stacks_1_stack_layers_1_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.504p-1)];
            tensor<fp16, [1, 512, 1, 512]> src_15_cast_fp16 = mul(x = var_1578_cast_fp16, y = core_stacks_1_stack_layers_1_norm_scale_to_fp16)[name = tensor<string, []>("src_15_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_1581_cast_fp16 = sub(x = src_15_cast_fp16, y = input_105_cast_fp16)[name = tensor<string, []>("op_1581_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_1_stack_layers_1_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_stack_layers_1_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144263168)))];
            tensor<fp16, [1, 512, 1, 512]> var_1582_cast_fp16 = mul(x = var_1581_cast_fp16, y = core_stacks_1_stack_layers_1_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_1582_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_135_cast_fp16 = add(x = input_105_cast_fp16, y = var_1582_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<fp32, []> src_17_scale_factor_height_0 = const()[name = tensor<string, []>("src_17_scale_factor_height_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, []> src_17_scale_factor_width_0 = const()[name = tensor<string, []>("src_17_scale_factor_width_0"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp16, [1, 512, 1, 1024]> src_17_cast_fp16 = upsample_nearest_neighbor(scale_factor_height = src_17_scale_factor_height_0, scale_factor_width = src_17_scale_factor_width_0, x = input_135_cast_fp16)[name = tensor<string, []>("src_17_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_1587_cast_fp16 = sub(x = src_17_cast_fp16, y = input_71_cast_fp16)[name = tensor<string, []>("op_1587_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_1_out_combiner_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_1_out_combiner_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144264256)))];
            tensor<fp16, [1, 512, 1, 1024]> var_1588_cast_fp16 = mul(x = var_1587_cast_fp16, y = core_stacks_1_out_combiner_bypass_scale_to_fp16)[name = tensor<string, []>("op_1588_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_137_cast_fp16 = add(x = input_71_cast_fp16, y = var_1588_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
            tensor<string, []> input_141_pad_type_0 = const()[name = tensor<string, []>("input_141_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_141_strides_0 = const()[name = tensor<string, []>("input_141_strides_0"), val = tensor<int32, [2]>([1, 4])];
            tensor<int32, []> input_141_groups_0 = const()[name = tensor<string, []>("input_141_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [4]> input_141_pad_0 = const()[name = tensor<string, []>("input_141_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_141_dilations_0 = const()[name = tensor<string, []>("input_141_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 4]> core_stacks_2_down_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_down_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 4]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144265344)))];
            tensor<fp16, [1, 512, 1, 256]> input_141_cast_fp16 = conv(dilations = input_141_dilations_0, groups = input_141_groups_0, pad = input_141_pad_0, pad_type = input_141_pad_type_0, strides = input_141_strides_0, weight = core_stacks_2_down_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
            tensor<string, []> time_emb_5_pad_type_0 = const()[name = tensor<string, []>("time_emb_5_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> time_emb_5_strides_0 = const()[name = tensor<string, []>("time_emb_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> time_emb_5_pad_0 = const()[name = tensor<string, []>("time_emb_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> time_emb_5_dilations_0 = const()[name = tensor<string, []>("time_emb_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> time_emb_5_groups_0 = const()[name = tensor<string, []>("time_emb_5_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 192, 1, 1]> core_stacks_2_stack_time_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_time_proj_weight_to_fp16"), val = tensor<fp16, [512, 192, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144269504)))];
            tensor<fp16, [512]> core_stacks_2_stack_time_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_time_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144466176)))];
            tensor<fp16, [1, 512, 1, 1]> time_emb_5_cast_fp16 = conv(bias = core_stacks_2_stack_time_proj_bias_to_fp16, dilations = time_emb_5_dilations_0, groups = time_emb_5_groups_0, pad = time_emb_5_pad_0, pad_type = time_emb_5_pad_type_0, strides = time_emb_5_strides_0, weight = core_stacks_2_stack_time_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("time_emb_5_cast_fp16")];
            tensor<string, []> proj_9_pad_type_0 = const()[name = tensor<string, []>("proj_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_9_strides_0 = const()[name = tensor<string, []>("proj_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_9_pad_0 = const()[name = tensor<string, []>("proj_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_9_dilations_0 = const()[name = tensor<string, []>("proj_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_9_groups_0 = const()[name = tensor<string, []>("proj_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_2_stack_layers_0_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144467264)))];
            tensor<fp16, [384]> core_stacks_2_stack_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144860544)))];
            tensor<fp16, [1, 384, 1, 256]> proj_9_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_9_dilations_0, groups = proj_9_groups_0, pad = proj_9_pad_0, pad_type = proj_9_pad_type_0, strides = proj_9_strides_0, weight = core_stacks_2_stack_layers_0_self_attn_weights_in_proj_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("proj_9_cast_fp16")];
            tensor<int32, [4]> var_1649_begin_0 = const()[name = tensor<string, []>("op_1649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1649_end_0 = const()[name = tensor<string, []>("op_1649_end_0"), val = tensor<int32, [4]>([1, 128, 1, 256])];
            tensor<bool, [4]> var_1649_end_mask_0 = const()[name = tensor<string, []>("op_1649_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_1649_cast_fp16 = slice_by_index(begin = var_1649_begin_0, end = var_1649_end_0, end_mask = var_1649_end_mask_0, x = proj_9_cast_fp16)[name = tensor<string, []>("op_1649_cast_fp16")];
            tensor<int32, [4]> var_1650 = const()[name = tensor<string, []>("op_1650"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> var_1651_cast_fp16 = reshape(shape = var_1650, x = var_1649_cast_fp16)[name = tensor<string, []>("op_1651_cast_fp16")];
            tensor<int32, [4]> var_1653_begin_0 = const()[name = tensor<string, []>("op_1653_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_1653_end_0 = const()[name = tensor<string, []>("op_1653_end_0"), val = tensor<int32, [4]>([1, 256, 1, 256])];
            tensor<bool, [4]> var_1653_end_mask_0 = const()[name = tensor<string, []>("op_1653_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, x = proj_9_cast_fp16)[name = tensor<string, []>("op_1653_cast_fp16")];
            tensor<int32, [4]> var_1654 = const()[name = tensor<string, []>("op_1654"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> var_1655_cast_fp16 = reshape(shape = var_1654, x = var_1653_cast_fp16)[name = tensor<string, []>("op_1655_cast_fp16")];
            tensor<int32, [4]> var_1657_begin_0 = const()[name = tensor<string, []>("op_1657_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_1657_end_0 = const()[name = tensor<string, []>("op_1657_end_0"), val = tensor<int32, [4]>([1, 384, 1, 256])];
            tensor<bool, [4]> var_1657_end_mask_0 = const()[name = tensor<string, []>("op_1657_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_1657_cast_fp16 = slice_by_index(begin = var_1657_begin_0, end = var_1657_end_0, end_mask = var_1657_end_mask_0, x = proj_9_cast_fp16)[name = tensor<string, []>("op_1657_cast_fp16")];
            tensor<int32, [4]> var_1658 = const()[name = tensor<string, []>("op_1658"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> pt_9_cast_fp16 = reshape(shape = var_1658, x = var_1657_cast_fp16)[name = tensor<string, []>("pt_9_cast_fp16")];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = tensor<string, []>("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_164 = const()[name = tensor<string, []>("concat_164"), val = tensor<int32, [3]>([4, 256, 32])];
            tensor<fp16, [4, 1, 256, 32]> transpose_32_cast_fp16 = transpose(perm = transpose_32_perm_0, x = var_1651_cast_fp16)[name = tensor<string, []>("transpose_272")];
            tensor<fp16, [4, 256, 32]> reshape_48_cast_fp16 = reshape(shape = concat_164, x = transpose_32_cast_fp16)[name = tensor<string, []>("reshape_48_cast_fp16")];
            tensor<int32, [4]> transpose_33_perm_0 = const()[name = tensor<string, []>("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_165 = const()[name = tensor<string, []>("concat_165"), val = tensor<int32, [3]>([4, 32, 256])];
            tensor<fp16, [4, 1, 32, 256]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = var_1655_cast_fp16)[name = tensor<string, []>("transpose_271")];
            tensor<fp16, [4, 32, 256]> reshape_49_cast_fp16 = reshape(shape = concat_165, x = transpose_33_cast_fp16)[name = tensor<string, []>("reshape_49_cast_fp16")];
            tensor<bool, []> matmul_16_transpose_x_0 = const()[name = tensor<string, []>("matmul_16_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_16_transpose_y_0 = const()[name = tensor<string, []>("matmul_16_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 256]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = reshape_48_cast_fp16, y = reshape_49_cast_fp16)[name = tensor<string, []>("matmul_16_cast_fp16")];
            tensor<int32, [4]> concat_169 = const()[name = tensor<string, []>("concat_169"), val = tensor<int32, [4]>([4, 1, 256, 256])];
            tensor<fp16, [4, 1, 256, 256]> reshape_50_cast_fp16 = reshape(shape = concat_169, x = matmul_16_cast_fp16)[name = tensor<string, []>("reshape_50_cast_fp16")];
            tensor<int32, [4]> attn_scores_25_perm_0 = const()[name = tensor<string, []>("attn_scores_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1662 = const()[name = tensor<string, []>("op_1662"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [1, 32, 256, 256]> core_pos_basis_256_to_fp16 = const()[name = tensor<string, []>("core_pos_basis_256_to_fp16"), val = tensor<fp16, [1, 32, 256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144861376)))];
            tensor<fp16, [4, 32, 256, 1]> var_1663_cast_fp16 = transpose(perm = var_1662, x = pt_9_cast_fp16)[name = tensor<string, []>("transpose_269")];
            tensor<fp16, [4, 32, 256, 256]> var_1664_cast_fp16 = mul(x = var_1663_cast_fp16, y = core_pos_basis_256_to_fp16)[name = tensor<string, []>("op_1664_cast_fp16")];
            tensor<int32, [1]> pos_scores_9_axes_0 = const()[name = tensor<string, []>("pos_scores_9_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_9_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_9_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 256, 256]> pos_scores_9_cast_fp16 = reduce_sum(axes = pos_scores_9_axes_0, keep_dims = pos_scores_9_keep_dims_0, x = var_1664_cast_fp16)[name = tensor<string, []>("pos_scores_9_cast_fp16")];
            tensor<int32, [4]> var_1667 = const()[name = tensor<string, []>("op_1667"), val = tensor<int32, [4]>([4, 256, 1, 256])];
            tensor<fp16, [4, 256, 1, 256]> var_1668_cast_fp16 = reshape(shape = var_1667, x = pos_scores_9_cast_fp16)[name = tensor<string, []>("op_1668_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_25_cast_fp16 = transpose(perm = attn_scores_25_perm_0, x = reshape_50_cast_fp16)[name = tensor<string, []>("transpose_270")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_27_cast_fp16 = add(x = attn_scores_25_cast_fp16, y = var_1668_cast_fp16)[name = tensor<string, []>("attn_scores_27_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_29_cast_fp16 = add(x = attn_scores_27_cast_fp16, y = attn_bias_cast_fp16)[name = tensor<string, []>("attn_scores_29_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_weights_9_cast_fp16 = softmax(axis = var_41, x = attn_scores_29_cast_fp16)[name = tensor<string, []>("attn_weights_9_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_143_cast_fp16 = add(x = input_141_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
            tensor<string, []> x_191_pad_type_0 = const()[name = tensor<string, []>("x_191_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_191_strides_0 = const()[name = tensor<string, []>("x_191_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_191_pad_0 = const()[name = tensor<string, []>("x_191_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_191_dilations_0 = const()[name = tensor<string, []>("x_191_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_191_groups_0 = const()[name = tensor<string, []>("x_191_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_2_stack_layers_0_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149055744)))];
            tensor<fp16, [1152]> core_stacks_2_stack_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150235456)))];
            tensor<fp16, [1, 1152, 1, 256]> x_191_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_feed_forward1_in_proj_bias_to_fp16, dilations = x_191_dilations_0, groups = x_191_groups_0, pad = x_191_pad_0, pad_type = x_191_pad_type_0, strides = x_191_strides_0, weight = core_stacks_2_stack_layers_0_feed_forward1_in_proj_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("x_191_cast_fp16")];
            tensor<fp16, []> var_1682_to_fp16 = const()[name = tensor<string, []>("op_1682_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 256]> x_193_cast_fp16 = sub(x = x_191_cast_fp16, y = var_1682_to_fp16)[name = tensor<string, []>("x_193_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_1684_cast_fp16 = relu(x = x_193_cast_fp16)[name = tensor<string, []>("op_1684_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_1685_cast_fp16 = abs(x = x_193_cast_fp16)[name = tensor<string, []>("op_1685_cast_fp16")];
            tensor<fp16, []> const_36_promoted_to_fp16 = const()[name = tensor<string, []>("const_36_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 256]> var_1686_cast_fp16 = mul(x = var_1685_cast_fp16, y = const_36_promoted_to_fp16)[name = tensor<string, []>("op_1686_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_1687_cast_fp16 = exp(x = var_1686_cast_fp16)[name = tensor<string, []>("op_1687_cast_fp16")];
            tensor<fp32, []> var_1688_epsilon_0 = const()[name = tensor<string, []>("op_1688_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 256]> var_1688_cast_fp16 = log(epsilon = var_1688_epsilon_0, x = var_1687_cast_fp16)[name = tensor<string, []>("op_1688_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_1689_cast_fp16 = add(x = var_1684_cast_fp16, y = var_1688_cast_fp16)[name = tensor<string, []>("op_1689_cast_fp16")];
            tensor<fp16, []> var_1690_to_fp16 = const()[name = tensor<string, []>("op_1690_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 256]> var_1691_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1690_to_fp16)[name = tensor<string, []>("op_1691_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_1692_cast_fp16 = sub(x = var_1689_cast_fp16, y = var_1691_cast_fp16)[name = tensor<string, []>("op_1692_cast_fp16")];
            tensor<fp16, []> var_1693_to_fp16 = const()[name = tensor<string, []>("op_1693_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 256]> input_145_cast_fp16 = sub(x = var_1692_cast_fp16, y = var_1693_to_fp16)[name = tensor<string, []>("input_145_cast_fp16")];
            tensor<string, []> var_1701_pad_type_0 = const()[name = tensor<string, []>("op_1701_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1701_strides_0 = const()[name = tensor<string, []>("op_1701_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1701_pad_0 = const()[name = tensor<string, []>("op_1701_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1701_dilations_0 = const()[name = tensor<string, []>("op_1701_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1701_groups_0 = const()[name = tensor<string, []>("op_1701_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_2_stack_layers_0_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150237824)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151417536)))];
            tensor<fp16, [1, 512, 1, 256]> var_1701_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_feed_forward1_out_proj_bias_to_fp16, dilations = var_1701_dilations_0, groups = var_1701_groups_0, pad = var_1701_pad_0, pad_type = var_1701_pad_type_0, strides = var_1701_strides_0, weight = core_stacks_2_stack_layers_0_feed_forward1_out_proj_weight_to_fp16, x = input_145_cast_fp16)[name = tensor<string, []>("op_1701_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_147_cast_fp16 = add(x = input_143_cast_fp16, y = var_1701_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<int32, [4]> var_1703_begin_0 = const()[name = tensor<string, []>("op_1703_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1703_end_0 = const()[name = tensor<string, []>("op_1703_end_0"), val = tensor<int32, [4]>([1, 256, 1, 256])];
            tensor<bool, [4]> var_1703_end_mask_0 = const()[name = tensor<string, []>("op_1703_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 256, 1, 256]> var_1703_cast_fp16 = slice_by_index(begin = var_1703_begin_0, end = var_1703_end_0, end_mask = var_1703_end_mask_0, x = attn_weights_9_cast_fp16)[name = tensor<string, []>("op_1703_cast_fp16")];
            tensor<string, []> x_195_pad_type_0 = const()[name = tensor<string, []>("x_195_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_195_strides_0 = const()[name = tensor<string, []>("x_195_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_195_pad_0 = const()[name = tensor<string, []>("x_195_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_195_dilations_0 = const()[name = tensor<string, []>("x_195_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_195_groups_0 = const()[name = tensor<string, []>("x_195_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_2_stack_layers_0_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151418624)))];
            tensor<fp16, [1152]> core_stacks_2_stack_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152598336)))];
            tensor<fp16, [1, 1152, 1, 256]> x_195_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_nonlin_attention_in_proj_bias_to_fp16, dilations = x_195_dilations_0, groups = x_195_groups_0, pad = x_195_pad_0, pad_type = x_195_pad_type_0, strides = x_195_strides_0, weight = core_stacks_2_stack_layers_0_nonlin_attention_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("x_195_cast_fp16")];
            tensor<int32, [4]> var_1714_begin_0 = const()[name = tensor<string, []>("op_1714_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1714_end_0 = const()[name = tensor<string, []>("op_1714_end_0"), val = tensor<int32, [4]>([1, 384, 1, 256])];
            tensor<bool, [4]> var_1714_end_mask_0 = const()[name = tensor<string, []>("op_1714_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 256]> var_1714_cast_fp16 = slice_by_index(begin = var_1714_begin_0, end = var_1714_end_0, end_mask = var_1714_end_mask_0, x = x_195_cast_fp16)[name = tensor<string, []>("op_1714_cast_fp16")];
            tensor<fp16, [1, 384, 1, 256]> s_gate_9_cast_fp16 = tanh(x = var_1714_cast_fp16)[name = tensor<string, []>("s_gate_9_cast_fp16")];
            tensor<int32, [4]> var_1717_begin_0 = const()[name = tensor<string, []>("op_1717_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_1717_end_0 = const()[name = tensor<string, []>("op_1717_end_0"), val = tensor<int32, [4]>([1, 768, 1, 256])];
            tensor<bool, [4]> var_1717_end_mask_0 = const()[name = tensor<string, []>("op_1717_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 256]> var_1717_cast_fp16 = slice_by_index(begin = var_1717_begin_0, end = var_1717_end_0, end_mask = var_1717_end_mask_0, x = x_195_cast_fp16)[name = tensor<string, []>("op_1717_cast_fp16")];
            tensor<fp16, [1, 384, 1, 256]> var_1718_cast_fp16 = mul(x = var_1717_cast_fp16, y = s_gate_9_cast_fp16)[name = tensor<string, []>("op_1718_cast_fp16")];
            tensor<int32, [4]> y_9_begin_0 = const()[name = tensor<string, []>("y_9_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_9_end_0 = const()[name = tensor<string, []>("y_9_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 256])];
            tensor<bool, [4]> y_9_end_mask_0 = const()[name = tensor<string, []>("y_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 256]> y_9_cast_fp16 = slice_by_index(begin = y_9_begin_0, end = y_9_end_0, end_mask = y_9_end_mask_0, x = x_195_cast_fp16)[name = tensor<string, []>("y_9_cast_fp16")];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = tensor<string, []>("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_174 = const()[name = tensor<string, []>("concat_174"), val = tensor<int32, [3]>([1, 256, 256])];
            tensor<fp16, [1, 1, 256, 256]> transpose_34_cast_fp16 = transpose(perm = transpose_34_perm_0, x = var_1703_cast_fp16)[name = tensor<string, []>("transpose_268")];
            tensor<fp16, [1, 256, 256]> reshape_51_cast_fp16 = reshape(shape = concat_174, x = transpose_34_cast_fp16)[name = tensor<string, []>("reshape_51_cast_fp16")];
            tensor<int32, [4]> transpose_35_perm_0 = const()[name = tensor<string, []>("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_175 = const()[name = tensor<string, []>("concat_175"), val = tensor<int32, [3]>([1, 256, 384])];
            tensor<fp16, [1, 1, 256, 384]> transpose_35_cast_fp16 = transpose(perm = transpose_35_perm_0, x = var_1718_cast_fp16)[name = tensor<string, []>("transpose_267")];
            tensor<fp16, [1, 256, 384]> reshape_52_cast_fp16 = reshape(shape = concat_175, x = transpose_35_cast_fp16)[name = tensor<string, []>("reshape_52_cast_fp16")];
            tensor<bool, []> matmul_17_transpose_x_0 = const()[name = tensor<string, []>("matmul_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_17_transpose_y_0 = const()[name = tensor<string, []>("matmul_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 256, 384]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = reshape_51_cast_fp16, y = reshape_52_cast_fp16)[name = tensor<string, []>("matmul_17_cast_fp16")];
            tensor<int32, [4]> concat_179 = const()[name = tensor<string, []>("concat_179"), val = tensor<int32, [4]>([1, 1, 256, 384])];
            tensor<fp16, [1, 1, 256, 384]> reshape_53_cast_fp16 = reshape(shape = concat_179, x = matmul_17_cast_fp16)[name = tensor<string, []>("reshape_53_cast_fp16")];
            tensor<int32, [4]> v_9_perm_0 = const()[name = tensor<string, []>("v_9_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 256]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = reshape_53_cast_fp16)[name = tensor<string, []>("transpose_266")];
            tensor<fp16, [1, 384, 1, 256]> input_149_cast_fp16 = mul(x = v_9_cast_fp16, y = y_9_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
            tensor<string, []> var_1730_pad_type_0 = const()[name = tensor<string, []>("op_1730_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1730_strides_0 = const()[name = tensor<string, []>("op_1730_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1730_pad_0 = const()[name = tensor<string, []>("op_1730_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1730_dilations_0 = const()[name = tensor<string, []>("op_1730_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1730_groups_0 = const()[name = tensor<string, []>("op_1730_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_2_stack_layers_0_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152600704)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152993984)))];
            tensor<fp16, [1, 512, 1, 256]> var_1730_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_nonlin_attention_out_proj_bias_to_fp16, dilations = var_1730_dilations_0, groups = var_1730_groups_0, pad = var_1730_pad_0, pad_type = var_1730_pad_type_0, strides = var_1730_strides_0, weight = core_stacks_2_stack_layers_0_nonlin_attention_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor<string, []>("op_1730_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_197_cast_fp16 = add(x = input_147_cast_fp16, y = var_1730_cast_fp16)[name = tensor<string, []>("x_197_cast_fp16")];
            tensor<string, []> var_1741_pad_type_0 = const()[name = tensor<string, []>("op_1741_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1741_strides_0 = const()[name = tensor<string, []>("op_1741_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1741_pad_0 = const()[name = tensor<string, []>("op_1741_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1741_dilations_0 = const()[name = tensor<string, []>("op_1741_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1741_groups_0 = const()[name = tensor<string, []>("op_1741_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_2_stack_layers_0_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152995072)))];
            tensor<fp16, [48]> core_stacks_2_stack_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153044288)))];
            tensor<fp16, [1, 48, 1, 256]> var_1741_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_self_attn1_in_proj_bias_to_fp16, dilations = var_1741_dilations_0, groups = var_1741_groups_0, pad = var_1741_pad_0, pad_type = var_1741_pad_type_0, strides = var_1741_strides_0, weight = core_stacks_2_stack_layers_0_self_attn1_in_proj_weight_to_fp16, x = x_197_cast_fp16)[name = tensor<string, []>("op_1741_cast_fp16")];
            tensor<int32, [4]> var_1742 = const()[name = tensor<string, []>("op_1742"), val = tensor<int32, [4]>([4, 12, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> var_1743_cast_fp16 = reshape(shape = var_1742, x = var_1741_cast_fp16)[name = tensor<string, []>("op_1743_cast_fp16")];
            tensor<int32, [4]> transpose_36_perm_0 = const()[name = tensor<string, []>("transpose_36_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_184 = const()[name = tensor<string, []>("concat_184"), val = tensor<int32, [3]>([4, 256, 256])];
            tensor<fp16, [4, 1, 256, 256]> transpose_36_cast_fp16 = transpose(perm = transpose_36_perm_0, x = attn_weights_9_cast_fp16)[name = tensor<string, []>("transpose_265")];
            tensor<fp16, [4, 256, 256]> reshape_54_cast_fp16 = reshape(shape = concat_184, x = transpose_36_cast_fp16)[name = tensor<string, []>("reshape_54_cast_fp16")];
            tensor<int32, [4]> transpose_37_perm_0 = const()[name = tensor<string, []>("transpose_37_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_185 = const()[name = tensor<string, []>("concat_185"), val = tensor<int32, [3]>([4, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> transpose_37_cast_fp16 = transpose(perm = transpose_37_perm_0, x = var_1743_cast_fp16)[name = tensor<string, []>("transpose_264")];
            tensor<fp16, [4, 256, 12]> reshape_55_cast_fp16 = reshape(shape = concat_185, x = transpose_37_cast_fp16)[name = tensor<string, []>("reshape_55_cast_fp16")];
            tensor<bool, []> matmul_18_transpose_x_0 = const()[name = tensor<string, []>("matmul_18_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_18_transpose_y_0 = const()[name = tensor<string, []>("matmul_18_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 12]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = reshape_54_cast_fp16, y = reshape_55_cast_fp16)[name = tensor<string, []>("matmul_18_cast_fp16")];
            tensor<int32, [4]> concat_189 = const()[name = tensor<string, []>("concat_189"), val = tensor<int32, [4]>([4, 1, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> reshape_56_cast_fp16 = reshape(shape = concat_189, x = matmul_18_cast_fp16)[name = tensor<string, []>("reshape_56_cast_fp16")];
            tensor<int32, [4]> out_17_perm_0 = const()[name = tensor<string, []>("out_17_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_1746 = const()[name = tensor<string, []>("op_1746"), val = tensor<int32, [4]>([1, 48, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> out_17_cast_fp16 = transpose(perm = out_17_perm_0, x = reshape_56_cast_fp16)[name = tensor<string, []>("transpose_263")];
            tensor<fp16, [1, 48, 1, 256]> input_151_cast_fp16 = reshape(shape = var_1746, x = out_17_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
            tensor<string, []> var_1754_pad_type_0 = const()[name = tensor<string, []>("op_1754_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1754_strides_0 = const()[name = tensor<string, []>("op_1754_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1754_pad_0 = const()[name = tensor<string, []>("op_1754_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1754_dilations_0 = const()[name = tensor<string, []>("op_1754_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1754_groups_0 = const()[name = tensor<string, []>("op_1754_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_2_stack_layers_0_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153044480)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153093696)))];
            tensor<fp16, [1, 512, 1, 256]> var_1754_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_self_attn1_out_proj_bias_to_fp16, dilations = var_1754_dilations_0, groups = var_1754_groups_0, pad = var_1754_pad_0, pad_type = var_1754_pad_type_0, strides = var_1754_strides_0, weight = core_stacks_2_stack_layers_0_self_attn1_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("op_1754_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_199_cast_fp16 = add(x = x_197_cast_fp16, y = var_1754_cast_fp16)[name = tensor<string, []>("x_199_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_201_cast_fp16 = add(x = x_199_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("x_201_cast_fp16")];
            tensor<string, []> x_203_pad_type_0 = const()[name = tensor<string, []>("x_203_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_203_strides_0 = const()[name = tensor<string, []>("x_203_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_203_pad_0 = const()[name = tensor<string, []>("x_203_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_203_dilations_0 = const()[name = tensor<string, []>("x_203_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_203_groups_0 = const()[name = tensor<string, []>("x_203_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_2_stack_layers_0_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153094784)))];
            tensor<fp16, [1024]> core_stacks_2_stack_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154143424)))];
            tensor<fp16, [1, 1024, 1, 256]> x_203_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_conv_module1_in_proj_bias_to_fp16, dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = core_stacks_2_stack_layers_0_conv_module1_in_proj_weight_to_fp16, x = x_201_cast_fp16)[name = tensor<string, []>("x_203_cast_fp16")];
            tensor<int32, [4]> var_1768_begin_0 = const()[name = tensor<string, []>("op_1768_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1768_end_0 = const()[name = tensor<string, []>("op_1768_end_0"), val = tensor<int32, [4]>([1, 512, 1, 256])];
            tensor<bool, [4]> var_1768_end_mask_0 = const()[name = tensor<string, []>("op_1768_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_1768_cast_fp16 = slice_by_index(begin = var_1768_begin_0, end = var_1768_end_0, end_mask = var_1768_end_mask_0, x = x_203_cast_fp16)[name = tensor<string, []>("op_1768_cast_fp16")];
            tensor<int32, [4]> var_1770_begin_0 = const()[name = tensor<string, []>("op_1770_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_1770_end_0 = const()[name = tensor<string, []>("op_1770_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 256])];
            tensor<bool, [4]> var_1770_end_mask_0 = const()[name = tensor<string, []>("op_1770_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_1770_cast_fp16 = slice_by_index(begin = var_1770_begin_0, end = var_1770_end_0, end_mask = var_1770_end_mask_0, x = x_203_cast_fp16)[name = tensor<string, []>("op_1770_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1771_cast_fp16 = sigmoid(x = var_1770_cast_fp16)[name = tensor<string, []>("op_1771_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_205_cast_fp16 = mul(x = var_1768_cast_fp16, y = var_1771_cast_fp16)[name = tensor<string, []>("x_205_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_153_cast_fp16 = mul(x = x_205_cast_fp16, y = conv_keep_cast_fp16)[name = tensor<string, []>("input_153_cast_fp16")];
            tensor<string, []> x_207_pad_type_0 = const()[name = tensor<string, []>("x_207_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_207_pad_0 = const()[name = tensor<string, []>("x_207_pad_0"), val = tensor<int32, [4]>([0, 0, 3, 3])];
            tensor<int32, []> x_207_groups_0 = const()[name = tensor<string, []>("x_207_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_207_strides_0 = const()[name = tensor<string, []>("x_207_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_207_dilations_0 = const()[name = tensor<string, []>("x_207_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 7]> core_stacks_2_stack_layers_0_conv_module1_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module1_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 7]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154145536)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_0_conv_module1_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module1_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154152768)))];
            tensor<fp16, [1, 512, 1, 256]> x_207_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_conv_module1_dw_bias_to_fp16, dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = core_stacks_2_stack_layers_0_conv_module1_dw_weight_to_fp16, x = input_153_cast_fp16)[name = tensor<string, []>("x_207_cast_fp16")];
            tensor<fp16, []> var_1781_to_fp16 = const()[name = tensor<string, []>("op_1781_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> x_209_cast_fp16 = sub(x = x_207_cast_fp16, y = var_1781_to_fp16)[name = tensor<string, []>("x_209_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1783_cast_fp16 = relu(x = x_209_cast_fp16)[name = tensor<string, []>("op_1783_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1784_cast_fp16 = abs(x = x_209_cast_fp16)[name = tensor<string, []>("op_1784_cast_fp16")];
            tensor<fp16, []> const_38_promoted_to_fp16 = const()[name = tensor<string, []>("const_38_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_1785_cast_fp16 = mul(x = var_1784_cast_fp16, y = const_38_promoted_to_fp16)[name = tensor<string, []>("op_1785_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1786_cast_fp16 = exp(x = var_1785_cast_fp16)[name = tensor<string, []>("op_1786_cast_fp16")];
            tensor<fp32, []> var_1787_epsilon_0 = const()[name = tensor<string, []>("op_1787_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_1787_cast_fp16 = log(epsilon = var_1787_epsilon_0, x = var_1786_cast_fp16)[name = tensor<string, []>("op_1787_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1788_cast_fp16 = add(x = var_1783_cast_fp16, y = var_1787_cast_fp16)[name = tensor<string, []>("op_1788_cast_fp16")];
            tensor<fp16, []> var_1789_to_fp16 = const()[name = tensor<string, []>("op_1789_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 256]> var_1790_cast_fp16 = mul(x = x_207_cast_fp16, y = var_1789_to_fp16)[name = tensor<string, []>("op_1790_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1791_cast_fp16 = sub(x = var_1788_cast_fp16, y = var_1790_cast_fp16)[name = tensor<string, []>("op_1791_cast_fp16")];
            tensor<fp16, []> var_1792_to_fp16 = const()[name = tensor<string, []>("op_1792_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 256]> input_155_cast_fp16 = sub(x = var_1791_cast_fp16, y = var_1792_to_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
            tensor<string, []> var_1800_pad_type_0 = const()[name = tensor<string, []>("op_1800_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1800_strides_0 = const()[name = tensor<string, []>("op_1800_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1800_pad_0 = const()[name = tensor<string, []>("op_1800_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1800_dilations_0 = const()[name = tensor<string, []>("op_1800_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1800_groups_0 = const()[name = tensor<string, []>("op_1800_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_2_stack_layers_0_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154153856)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154678208)))];
            tensor<fp16, [1, 512, 1, 256]> var_1800_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_conv_module1_out_proj_bias_to_fp16, dilations = var_1800_dilations_0, groups = var_1800_groups_0, pad = var_1800_pad_0, pad_type = var_1800_pad_type_0, strides = var_1800_strides_0, weight = core_stacks_2_stack_layers_0_conv_module1_out_proj_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("op_1800_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_157_cast_fp16 = add(x = x_201_cast_fp16, y = var_1800_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
            tensor<string, []> x_211_pad_type_0 = const()[name = tensor<string, []>("x_211_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_211_strides_0 = const()[name = tensor<string, []>("x_211_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_211_pad_0 = const()[name = tensor<string, []>("x_211_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_211_dilations_0 = const()[name = tensor<string, []>("x_211_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_211_groups_0 = const()[name = tensor<string, []>("x_211_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_2_stack_layers_0_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154679296)))];
            tensor<fp16, [1536]> core_stacks_2_stack_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156252224)))];
            tensor<fp16, [1, 1536, 1, 256]> x_211_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_feed_forward2_in_proj_bias_to_fp16, dilations = x_211_dilations_0, groups = x_211_groups_0, pad = x_211_pad_0, pad_type = x_211_pad_type_0, strides = x_211_strides_0, weight = core_stacks_2_stack_layers_0_feed_forward2_in_proj_weight_to_fp16, x = input_157_cast_fp16)[name = tensor<string, []>("x_211_cast_fp16")];
            tensor<fp16, []> var_1811_to_fp16 = const()[name = tensor<string, []>("op_1811_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 256]> x_213_cast_fp16 = sub(x = x_211_cast_fp16, y = var_1811_to_fp16)[name = tensor<string, []>("x_213_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_1813_cast_fp16 = relu(x = x_213_cast_fp16)[name = tensor<string, []>("op_1813_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_1814_cast_fp16 = abs(x = x_213_cast_fp16)[name = tensor<string, []>("op_1814_cast_fp16")];
            tensor<fp16, []> const_39_promoted_to_fp16 = const()[name = tensor<string, []>("const_39_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 256]> var_1815_cast_fp16 = mul(x = var_1814_cast_fp16, y = const_39_promoted_to_fp16)[name = tensor<string, []>("op_1815_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_1816_cast_fp16 = exp(x = var_1815_cast_fp16)[name = tensor<string, []>("op_1816_cast_fp16")];
            tensor<fp32, []> var_1817_epsilon_0 = const()[name = tensor<string, []>("op_1817_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 256]> var_1817_cast_fp16 = log(epsilon = var_1817_epsilon_0, x = var_1816_cast_fp16)[name = tensor<string, []>("op_1817_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_1818_cast_fp16 = add(x = var_1813_cast_fp16, y = var_1817_cast_fp16)[name = tensor<string, []>("op_1818_cast_fp16")];
            tensor<fp16, []> var_1819_to_fp16 = const()[name = tensor<string, []>("op_1819_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 256]> var_1820_cast_fp16 = mul(x = x_211_cast_fp16, y = var_1819_to_fp16)[name = tensor<string, []>("op_1820_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_1821_cast_fp16 = sub(x = var_1818_cast_fp16, y = var_1820_cast_fp16)[name = tensor<string, []>("op_1821_cast_fp16")];
            tensor<fp16, []> var_1822_to_fp16 = const()[name = tensor<string, []>("op_1822_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 256]> input_159_cast_fp16 = sub(x = var_1821_cast_fp16, y = var_1822_to_fp16)[name = tensor<string, []>("input_159_cast_fp16")];
            tensor<string, []> var_1830_pad_type_0 = const()[name = tensor<string, []>("op_1830_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1830_strides_0 = const()[name = tensor<string, []>("op_1830_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1830_pad_0 = const()[name = tensor<string, []>("op_1830_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1830_dilations_0 = const()[name = tensor<string, []>("op_1830_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1830_groups_0 = const()[name = tensor<string, []>("op_1830_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_2_stack_layers_0_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(156255360)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157828288)))];
            tensor<fp16, [1, 512, 1, 256]> var_1830_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_feed_forward2_out_proj_bias_to_fp16, dilations = var_1830_dilations_0, groups = var_1830_groups_0, pad = var_1830_pad_0, pad_type = var_1830_pad_type_0, strides = var_1830_strides_0, weight = core_stacks_2_stack_layers_0_feed_forward2_out_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor<string, []>("op_1830_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> src_19_cast_fp16 = add(x = input_157_cast_fp16, y = var_1830_cast_fp16)[name = tensor<string, []>("src_19_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1833_cast_fp16 = sub(x = src_19_cast_fp16, y = input_141_cast_fp16)[name = tensor<string, []>("op_1833_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157829376)))];
            tensor<fp16, [1, 512, 1, 256]> var_1834_cast_fp16 = mul(x = var_1833_cast_fp16, y = core_stacks_2_stack_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_1834_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_215_cast_fp16 = add(x = input_141_cast_fp16, y = var_1834_cast_fp16)[name = tensor<string, []>("x_215_cast_fp16")];
            tensor<string, []> var_1845_pad_type_0 = const()[name = tensor<string, []>("op_1845_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1845_strides_0 = const()[name = tensor<string, []>("op_1845_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1845_pad_0 = const()[name = tensor<string, []>("op_1845_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1845_dilations_0 = const()[name = tensor<string, []>("op_1845_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1845_groups_0 = const()[name = tensor<string, []>("op_1845_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_2_stack_layers_0_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157830464)))];
            tensor<fp16, [48]> core_stacks_2_stack_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157879680)))];
            tensor<fp16, [1, 48, 1, 256]> var_1845_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_self_attn2_in_proj_bias_to_fp16, dilations = var_1845_dilations_0, groups = var_1845_groups_0, pad = var_1845_pad_0, pad_type = var_1845_pad_type_0, strides = var_1845_strides_0, weight = core_stacks_2_stack_layers_0_self_attn2_in_proj_weight_to_fp16, x = x_215_cast_fp16)[name = tensor<string, []>("op_1845_cast_fp16")];
            tensor<int32, [4]> var_1846 = const()[name = tensor<string, []>("op_1846"), val = tensor<int32, [4]>([4, 12, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> var_1847_cast_fp16 = reshape(shape = var_1846, x = var_1845_cast_fp16)[name = tensor<string, []>("op_1847_cast_fp16")];
            tensor<int32, [4]> transpose_39_perm_0 = const()[name = tensor<string, []>("transpose_39_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_195 = const()[name = tensor<string, []>("concat_195"), val = tensor<int32, [3]>([4, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> transpose_39_cast_fp16 = transpose(perm = transpose_39_perm_0, x = var_1847_cast_fp16)[name = tensor<string, []>("transpose_262")];
            tensor<fp16, [4, 256, 12]> reshape_58_cast_fp16 = reshape(shape = concat_195, x = transpose_39_cast_fp16)[name = tensor<string, []>("reshape_58_cast_fp16")];
            tensor<bool, []> matmul_19_transpose_x_0 = const()[name = tensor<string, []>("matmul_19_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_19_transpose_y_0 = const()[name = tensor<string, []>("matmul_19_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 12]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = reshape_54_cast_fp16, y = reshape_58_cast_fp16)[name = tensor<string, []>("matmul_19_cast_fp16")];
            tensor<int32, [4]> concat_199 = const()[name = tensor<string, []>("concat_199"), val = tensor<int32, [4]>([4, 1, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> reshape_59_cast_fp16 = reshape(shape = concat_199, x = matmul_19_cast_fp16)[name = tensor<string, []>("reshape_59_cast_fp16")];
            tensor<int32, [4]> out_19_perm_0 = const()[name = tensor<string, []>("out_19_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_1850 = const()[name = tensor<string, []>("op_1850"), val = tensor<int32, [4]>([1, 48, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> out_19_cast_fp16 = transpose(perm = out_19_perm_0, x = reshape_59_cast_fp16)[name = tensor<string, []>("transpose_261")];
            tensor<fp16, [1, 48, 1, 256]> input_161_cast_fp16 = reshape(shape = var_1850, x = out_19_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
            tensor<string, []> var_1858_pad_type_0 = const()[name = tensor<string, []>("op_1858_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1858_strides_0 = const()[name = tensor<string, []>("op_1858_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1858_pad_0 = const()[name = tensor<string, []>("op_1858_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1858_dilations_0 = const()[name = tensor<string, []>("op_1858_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1858_groups_0 = const()[name = tensor<string, []>("op_1858_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_2_stack_layers_0_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157879872)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157929088)))];
            tensor<fp16, [1, 512, 1, 256]> var_1858_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_self_attn2_out_proj_bias_to_fp16, dilations = var_1858_dilations_0, groups = var_1858_groups_0, pad = var_1858_pad_0, pad_type = var_1858_pad_type_0, strides = var_1858_strides_0, weight = core_stacks_2_stack_layers_0_self_attn2_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("op_1858_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_217_cast_fp16 = add(x = x_215_cast_fp16, y = var_1858_cast_fp16)[name = tensor<string, []>("x_217_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_219_cast_fp16 = add(x = x_217_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")];
            tensor<string, []> x_221_pad_type_0 = const()[name = tensor<string, []>("x_221_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_221_strides_0 = const()[name = tensor<string, []>("x_221_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_221_pad_0 = const()[name = tensor<string, []>("x_221_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_221_dilations_0 = const()[name = tensor<string, []>("x_221_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_221_groups_0 = const()[name = tensor<string, []>("x_221_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_2_stack_layers_0_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(157930176)))];
            tensor<fp16, [1024]> core_stacks_2_stack_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158978816)))];
            tensor<fp16, [1, 1024, 1, 256]> x_221_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_conv_module2_in_proj_bias_to_fp16, dilations = x_221_dilations_0, groups = x_221_groups_0, pad = x_221_pad_0, pad_type = x_221_pad_type_0, strides = x_221_strides_0, weight = core_stacks_2_stack_layers_0_conv_module2_in_proj_weight_to_fp16, x = x_219_cast_fp16)[name = tensor<string, []>("x_221_cast_fp16")];
            tensor<int32, [4]> var_1872_begin_0 = const()[name = tensor<string, []>("op_1872_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1872_end_0 = const()[name = tensor<string, []>("op_1872_end_0"), val = tensor<int32, [4]>([1, 512, 1, 256])];
            tensor<bool, [4]> var_1872_end_mask_0 = const()[name = tensor<string, []>("op_1872_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_1872_cast_fp16 = slice_by_index(begin = var_1872_begin_0, end = var_1872_end_0, end_mask = var_1872_end_mask_0, x = x_221_cast_fp16)[name = tensor<string, []>("op_1872_cast_fp16")];
            tensor<int32, [4]> var_1874_begin_0 = const()[name = tensor<string, []>("op_1874_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_1874_end_0 = const()[name = tensor<string, []>("op_1874_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 256])];
            tensor<bool, [4]> var_1874_end_mask_0 = const()[name = tensor<string, []>("op_1874_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_1874_cast_fp16 = slice_by_index(begin = var_1874_begin_0, end = var_1874_end_0, end_mask = var_1874_end_mask_0, x = x_221_cast_fp16)[name = tensor<string, []>("op_1874_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1875_cast_fp16 = sigmoid(x = var_1874_cast_fp16)[name = tensor<string, []>("op_1875_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_223_cast_fp16 = mul(x = var_1872_cast_fp16, y = var_1875_cast_fp16)[name = tensor<string, []>("x_223_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_163_cast_fp16 = mul(x = x_223_cast_fp16, y = conv_keep_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
            tensor<string, []> x_225_pad_type_0 = const()[name = tensor<string, []>("x_225_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_225_pad_0 = const()[name = tensor<string, []>("x_225_pad_0"), val = tensor<int32, [4]>([0, 0, 3, 3])];
            tensor<int32, []> x_225_groups_0 = const()[name = tensor<string, []>("x_225_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_225_strides_0 = const()[name = tensor<string, []>("x_225_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_225_dilations_0 = const()[name = tensor<string, []>("x_225_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 7]> core_stacks_2_stack_layers_0_conv_module2_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module2_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 7]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158980928)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_0_conv_module2_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module2_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158988160)))];
            tensor<fp16, [1, 512, 1, 256]> x_225_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_conv_module2_dw_bias_to_fp16, dilations = x_225_dilations_0, groups = x_225_groups_0, pad = x_225_pad_0, pad_type = x_225_pad_type_0, strides = x_225_strides_0, weight = core_stacks_2_stack_layers_0_conv_module2_dw_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("x_225_cast_fp16")];
            tensor<fp16, []> var_1885_to_fp16 = const()[name = tensor<string, []>("op_1885_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> x_227_cast_fp16 = sub(x = x_225_cast_fp16, y = var_1885_to_fp16)[name = tensor<string, []>("x_227_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1887_cast_fp16 = relu(x = x_227_cast_fp16)[name = tensor<string, []>("op_1887_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1888_cast_fp16 = abs(x = x_227_cast_fp16)[name = tensor<string, []>("op_1888_cast_fp16")];
            tensor<fp16, []> const_41_promoted_to_fp16 = const()[name = tensor<string, []>("const_41_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_1889_cast_fp16 = mul(x = var_1888_cast_fp16, y = const_41_promoted_to_fp16)[name = tensor<string, []>("op_1889_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1890_cast_fp16 = exp(x = var_1889_cast_fp16)[name = tensor<string, []>("op_1890_cast_fp16")];
            tensor<fp32, []> var_1891_epsilon_0 = const()[name = tensor<string, []>("op_1891_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_1891_cast_fp16 = log(epsilon = var_1891_epsilon_0, x = var_1890_cast_fp16)[name = tensor<string, []>("op_1891_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1892_cast_fp16 = add(x = var_1887_cast_fp16, y = var_1891_cast_fp16)[name = tensor<string, []>("op_1892_cast_fp16")];
            tensor<fp16, []> var_1893_to_fp16 = const()[name = tensor<string, []>("op_1893_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 256]> var_1894_cast_fp16 = mul(x = x_225_cast_fp16, y = var_1893_to_fp16)[name = tensor<string, []>("op_1894_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1895_cast_fp16 = sub(x = var_1892_cast_fp16, y = var_1894_cast_fp16)[name = tensor<string, []>("op_1895_cast_fp16")];
            tensor<fp16, []> var_1896_to_fp16 = const()[name = tensor<string, []>("op_1896_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 256]> input_165_cast_fp16 = sub(x = var_1895_cast_fp16, y = var_1896_to_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
            tensor<string, []> var_1904_pad_type_0 = const()[name = tensor<string, []>("op_1904_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1904_strides_0 = const()[name = tensor<string, []>("op_1904_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1904_pad_0 = const()[name = tensor<string, []>("op_1904_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1904_dilations_0 = const()[name = tensor<string, []>("op_1904_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1904_groups_0 = const()[name = tensor<string, []>("op_1904_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_2_stack_layers_0_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158989248)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159513600)))];
            tensor<fp16, [1, 512, 1, 256]> var_1904_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_conv_module2_out_proj_bias_to_fp16, dilations = var_1904_dilations_0, groups = var_1904_groups_0, pad = var_1904_pad_0, pad_type = var_1904_pad_type_0, strides = var_1904_strides_0, weight = core_stacks_2_stack_layers_0_conv_module2_out_proj_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("op_1904_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_167_cast_fp16 = add(x = x_219_cast_fp16, y = var_1904_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
            tensor<string, []> x_229_pad_type_0 = const()[name = tensor<string, []>("x_229_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_229_strides_0 = const()[name = tensor<string, []>("x_229_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_229_pad_0 = const()[name = tensor<string, []>("x_229_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_229_dilations_0 = const()[name = tensor<string, []>("x_229_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_229_groups_0 = const()[name = tensor<string, []>("x_229_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_2_stack_layers_0_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159514688)))];
            tensor<fp16, [1920]> core_stacks_2_stack_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161480832)))];
            tensor<fp16, [1, 1920, 1, 256]> x_229_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_feed_forward3_in_proj_bias_to_fp16, dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = core_stacks_2_stack_layers_0_feed_forward3_in_proj_weight_to_fp16, x = input_167_cast_fp16)[name = tensor<string, []>("x_229_cast_fp16")];
            tensor<fp16, []> var_1915_to_fp16 = const()[name = tensor<string, []>("op_1915_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 256]> x_231_cast_fp16 = sub(x = x_229_cast_fp16, y = var_1915_to_fp16)[name = tensor<string, []>("x_231_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_1917_cast_fp16 = relu(x = x_231_cast_fp16)[name = tensor<string, []>("op_1917_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_1918_cast_fp16 = abs(x = x_231_cast_fp16)[name = tensor<string, []>("op_1918_cast_fp16")];
            tensor<fp16, []> const_42_promoted_to_fp16 = const()[name = tensor<string, []>("const_42_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 256]> var_1919_cast_fp16 = mul(x = var_1918_cast_fp16, y = const_42_promoted_to_fp16)[name = tensor<string, []>("op_1919_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_1920_cast_fp16 = exp(x = var_1919_cast_fp16)[name = tensor<string, []>("op_1920_cast_fp16")];
            tensor<fp32, []> var_1921_epsilon_0 = const()[name = tensor<string, []>("op_1921_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 256]> var_1921_cast_fp16 = log(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = tensor<string, []>("op_1921_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_1922_cast_fp16 = add(x = var_1917_cast_fp16, y = var_1921_cast_fp16)[name = tensor<string, []>("op_1922_cast_fp16")];
            tensor<fp16, []> var_1923_to_fp16 = const()[name = tensor<string, []>("op_1923_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 256]> var_1924_cast_fp16 = mul(x = x_229_cast_fp16, y = var_1923_to_fp16)[name = tensor<string, []>("op_1924_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_1925_cast_fp16 = sub(x = var_1922_cast_fp16, y = var_1924_cast_fp16)[name = tensor<string, []>("op_1925_cast_fp16")];
            tensor<fp16, []> var_1926_to_fp16 = const()[name = tensor<string, []>("op_1926_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 256]> input_169_cast_fp16 = sub(x = var_1925_cast_fp16, y = var_1926_to_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
            tensor<string, []> var_1934_pad_type_0 = const()[name = tensor<string, []>("op_1934_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_1934_strides_0 = const()[name = tensor<string, []>("op_1934_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1934_pad_0 = const()[name = tensor<string, []>("op_1934_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1934_dilations_0 = const()[name = tensor<string, []>("op_1934_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_1934_groups_0 = const()[name = tensor<string, []>("op_1934_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_2_stack_layers_0_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161484736)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163450880)))];
            tensor<fp16, [1, 512, 1, 256]> var_1934_cast_fp16 = conv(bias = core_stacks_2_stack_layers_0_feed_forward3_out_proj_bias_to_fp16, dilations = var_1934_dilations_0, groups = var_1934_groups_0, pad = var_1934_pad_0, pad_type = var_1934_pad_type_0, strides = var_1934_strides_0, weight = core_stacks_2_stack_layers_0_feed_forward3_out_proj_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("op_1934_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_233_cast_fp16 = add(x = input_167_cast_fp16, y = var_1934_cast_fp16)[name = tensor<string, []>("x_233_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163451968)))];
            tensor<fp16, [1, 512, 1, 256]> d_9_cast_fp16 = sub(x = x_233_cast_fp16, y = core_stacks_2_stack_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("d_9_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1939_cast_fp16 = mul(x = d_9_cast_fp16, y = d_9_cast_fp16)[name = tensor<string, []>("op_1939_cast_fp16")];
            tensor<int32, [1]> var_1941_axes_0 = const()[name = tensor<string, []>("op_1941_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_1941_keep_dims_0 = const()[name = tensor<string, []>("op_1941_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 256]> var_1941_cast_fp16 = reduce_mean(axes = var_1941_axes_0, keep_dims = var_1941_keep_dims_0, x = var_1939_cast_fp16)[name = tensor<string, []>("op_1941_cast_fp16")];
            tensor<fp32, []> scales_9_epsilon_0 = const()[name = tensor<string, []>("scales_9_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 256]> scales_9_cast_fp16 = rsqrt(epsilon = scales_9_epsilon_0, x = var_1941_cast_fp16)[name = tensor<string, []>("scales_9_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1943_cast_fp16 = mul(x = x_233_cast_fp16, y = scales_9_cast_fp16)[name = tensor<string, []>("op_1943_cast_fp16")];
            tensor<fp16, []> core_stacks_2_stack_layers_0_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.65cp-2)];
            tensor<fp16, [1, 512, 1, 256]> src_21_cast_fp16 = mul(x = var_1943_cast_fp16, y = core_stacks_2_stack_layers_0_norm_scale_to_fp16)[name = tensor<string, []>("src_21_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1946_cast_fp16 = sub(x = src_21_cast_fp16, y = input_141_cast_fp16)[name = tensor<string, []>("op_1946_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163453056)))];
            tensor<fp16, [1, 512, 1, 256]> var_1947_cast_fp16 = mul(x = var_1946_cast_fp16, y = core_stacks_2_stack_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_1947_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_171_cast_fp16 = add(x = input_141_cast_fp16, y = var_1947_cast_fp16)[name = tensor<string, []>("input_171_cast_fp16")];
            tensor<string, []> proj_11_pad_type_0 = const()[name = tensor<string, []>("proj_11_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_11_strides_0 = const()[name = tensor<string, []>("proj_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_11_pad_0 = const()[name = tensor<string, []>("proj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_11_dilations_0 = const()[name = tensor<string, []>("proj_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_11_groups_0 = const()[name = tensor<string, []>("proj_11_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_2_stack_layers_1_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163454144)))];
            tensor<fp16, [384]> core_stacks_2_stack_layers_1_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163847424)))];
            tensor<fp16, [1, 384, 1, 256]> proj_11_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_11_dilations_0, groups = proj_11_groups_0, pad = proj_11_pad_0, pad_type = proj_11_pad_type_0, strides = proj_11_strides_0, weight = core_stacks_2_stack_layers_1_self_attn_weights_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor<string, []>("proj_11_cast_fp16")];
            tensor<int32, [4]> var_1970_begin_0 = const()[name = tensor<string, []>("op_1970_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1970_end_0 = const()[name = tensor<string, []>("op_1970_end_0"), val = tensor<int32, [4]>([1, 128, 1, 256])];
            tensor<bool, [4]> var_1970_end_mask_0 = const()[name = tensor<string, []>("op_1970_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_1970_cast_fp16 = slice_by_index(begin = var_1970_begin_0, end = var_1970_end_0, end_mask = var_1970_end_mask_0, x = proj_11_cast_fp16)[name = tensor<string, []>("op_1970_cast_fp16")];
            tensor<int32, [4]> var_1971 = const()[name = tensor<string, []>("op_1971"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> var_1972_cast_fp16 = reshape(shape = var_1971, x = var_1970_cast_fp16)[name = tensor<string, []>("op_1972_cast_fp16")];
            tensor<int32, [4]> var_1974_begin_0 = const()[name = tensor<string, []>("op_1974_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_1974_end_0 = const()[name = tensor<string, []>("op_1974_end_0"), val = tensor<int32, [4]>([1, 256, 1, 256])];
            tensor<bool, [4]> var_1974_end_mask_0 = const()[name = tensor<string, []>("op_1974_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_1974_cast_fp16 = slice_by_index(begin = var_1974_begin_0, end = var_1974_end_0, end_mask = var_1974_end_mask_0, x = proj_11_cast_fp16)[name = tensor<string, []>("op_1974_cast_fp16")];
            tensor<int32, [4]> var_1975 = const()[name = tensor<string, []>("op_1975"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> var_1976_cast_fp16 = reshape(shape = var_1975, x = var_1974_cast_fp16)[name = tensor<string, []>("op_1976_cast_fp16")];
            tensor<int32, [4]> var_1978_begin_0 = const()[name = tensor<string, []>("op_1978_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_1978_end_0 = const()[name = tensor<string, []>("op_1978_end_0"), val = tensor<int32, [4]>([1, 384, 1, 256])];
            tensor<bool, [4]> var_1978_end_mask_0 = const()[name = tensor<string, []>("op_1978_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_1978_cast_fp16 = slice_by_index(begin = var_1978_begin_0, end = var_1978_end_0, end_mask = var_1978_end_mask_0, x = proj_11_cast_fp16)[name = tensor<string, []>("op_1978_cast_fp16")];
            tensor<int32, [4]> var_1979 = const()[name = tensor<string, []>("op_1979"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> pt_11_cast_fp16 = reshape(shape = var_1979, x = var_1978_cast_fp16)[name = tensor<string, []>("pt_11_cast_fp16")];
            tensor<int32, [4]> transpose_40_perm_0 = const()[name = tensor<string, []>("transpose_40_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_204 = const()[name = tensor<string, []>("concat_204"), val = tensor<int32, [3]>([4, 256, 32])];
            tensor<fp16, [4, 1, 256, 32]> transpose_40_cast_fp16 = transpose(perm = transpose_40_perm_0, x = var_1972_cast_fp16)[name = tensor<string, []>("transpose_260")];
            tensor<fp16, [4, 256, 32]> reshape_60_cast_fp16 = reshape(shape = concat_204, x = transpose_40_cast_fp16)[name = tensor<string, []>("reshape_60_cast_fp16")];
            tensor<int32, [4]> transpose_41_perm_0 = const()[name = tensor<string, []>("transpose_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_205 = const()[name = tensor<string, []>("concat_205"), val = tensor<int32, [3]>([4, 32, 256])];
            tensor<fp16, [4, 1, 32, 256]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = var_1976_cast_fp16)[name = tensor<string, []>("transpose_259")];
            tensor<fp16, [4, 32, 256]> reshape_61_cast_fp16 = reshape(shape = concat_205, x = transpose_41_cast_fp16)[name = tensor<string, []>("reshape_61_cast_fp16")];
            tensor<bool, []> matmul_20_transpose_x_0 = const()[name = tensor<string, []>("matmul_20_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_20_transpose_y_0 = const()[name = tensor<string, []>("matmul_20_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 256]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = reshape_60_cast_fp16, y = reshape_61_cast_fp16)[name = tensor<string, []>("matmul_20_cast_fp16")];
            tensor<int32, [4]> concat_209 = const()[name = tensor<string, []>("concat_209"), val = tensor<int32, [4]>([4, 1, 256, 256])];
            tensor<fp16, [4, 1, 256, 256]> reshape_62_cast_fp16 = reshape(shape = concat_209, x = matmul_20_cast_fp16)[name = tensor<string, []>("reshape_62_cast_fp16")];
            tensor<int32, [4]> attn_scores_31_perm_0 = const()[name = tensor<string, []>("attn_scores_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_1983 = const()[name = tensor<string, []>("op_1983"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 256, 1]> var_1984_cast_fp16 = transpose(perm = var_1983, x = pt_11_cast_fp16)[name = tensor<string, []>("transpose_257")];
            tensor<fp16, [4, 32, 256, 256]> var_1985_cast_fp16 = mul(x = var_1984_cast_fp16, y = core_pos_basis_256_to_fp16)[name = tensor<string, []>("op_1985_cast_fp16")];
            tensor<int32, [1]> pos_scores_11_axes_0 = const()[name = tensor<string, []>("pos_scores_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_11_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_11_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 256, 256]> pos_scores_11_cast_fp16 = reduce_sum(axes = pos_scores_11_axes_0, keep_dims = pos_scores_11_keep_dims_0, x = var_1985_cast_fp16)[name = tensor<string, []>("pos_scores_11_cast_fp16")];
            tensor<int32, [4]> var_1988 = const()[name = tensor<string, []>("op_1988"), val = tensor<int32, [4]>([4, 256, 1, 256])];
            tensor<fp16, [4, 256, 1, 256]> var_1989_cast_fp16 = reshape(shape = var_1988, x = pos_scores_11_cast_fp16)[name = tensor<string, []>("op_1989_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_31_cast_fp16 = transpose(perm = attn_scores_31_perm_0, x = reshape_62_cast_fp16)[name = tensor<string, []>("transpose_258")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_33_cast_fp16 = add(x = attn_scores_31_cast_fp16, y = var_1989_cast_fp16)[name = tensor<string, []>("attn_scores_33_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_35_cast_fp16 = add(x = attn_scores_33_cast_fp16, y = attn_bias_cast_fp16)[name = tensor<string, []>("attn_scores_35_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_weights_11_cast_fp16 = softmax(axis = var_41, x = attn_scores_35_cast_fp16)[name = tensor<string, []>("attn_weights_11_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_173_cast_fp16 = add(x = input_171_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
            tensor<string, []> x_235_pad_type_0 = const()[name = tensor<string, []>("x_235_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_235_strides_0 = const()[name = tensor<string, []>("x_235_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_235_pad_0 = const()[name = tensor<string, []>("x_235_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_235_dilations_0 = const()[name = tensor<string, []>("x_235_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_235_groups_0 = const()[name = tensor<string, []>("x_235_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_2_stack_layers_1_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163848256)))];
            tensor<fp16, [1152]> core_stacks_2_stack_layers_1_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165027968)))];
            tensor<fp16, [1, 1152, 1, 256]> x_235_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_feed_forward1_in_proj_bias_to_fp16, dilations = x_235_dilations_0, groups = x_235_groups_0, pad = x_235_pad_0, pad_type = x_235_pad_type_0, strides = x_235_strides_0, weight = core_stacks_2_stack_layers_1_feed_forward1_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("x_235_cast_fp16")];
            tensor<fp16, []> var_2003_to_fp16 = const()[name = tensor<string, []>("op_2003_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 256]> x_237_cast_fp16 = sub(x = x_235_cast_fp16, y = var_2003_to_fp16)[name = tensor<string, []>("x_237_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2005_cast_fp16 = relu(x = x_237_cast_fp16)[name = tensor<string, []>("op_2005_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2006_cast_fp16 = abs(x = x_237_cast_fp16)[name = tensor<string, []>("op_2006_cast_fp16")];
            tensor<fp16, []> const_43_promoted_to_fp16 = const()[name = tensor<string, []>("const_43_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 256]> var_2007_cast_fp16 = mul(x = var_2006_cast_fp16, y = const_43_promoted_to_fp16)[name = tensor<string, []>("op_2007_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2008_cast_fp16 = exp(x = var_2007_cast_fp16)[name = tensor<string, []>("op_2008_cast_fp16")];
            tensor<fp32, []> var_2009_epsilon_0 = const()[name = tensor<string, []>("op_2009_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 256]> var_2009_cast_fp16 = log(epsilon = var_2009_epsilon_0, x = var_2008_cast_fp16)[name = tensor<string, []>("op_2009_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2010_cast_fp16 = add(x = var_2005_cast_fp16, y = var_2009_cast_fp16)[name = tensor<string, []>("op_2010_cast_fp16")];
            tensor<fp16, []> var_2011_to_fp16 = const()[name = tensor<string, []>("op_2011_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 256]> var_2012_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2011_to_fp16)[name = tensor<string, []>("op_2012_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2013_cast_fp16 = sub(x = var_2010_cast_fp16, y = var_2012_cast_fp16)[name = tensor<string, []>("op_2013_cast_fp16")];
            tensor<fp16, []> var_2014_to_fp16 = const()[name = tensor<string, []>("op_2014_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 256]> input_175_cast_fp16 = sub(x = var_2013_cast_fp16, y = var_2014_to_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
            tensor<string, []> var_2022_pad_type_0 = const()[name = tensor<string, []>("op_2022_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2022_strides_0 = const()[name = tensor<string, []>("op_2022_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2022_pad_0 = const()[name = tensor<string, []>("op_2022_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2022_dilations_0 = const()[name = tensor<string, []>("op_2022_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2022_groups_0 = const()[name = tensor<string, []>("op_2022_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_2_stack_layers_1_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165030336)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_1_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166210048)))];
            tensor<fp16, [1, 512, 1, 256]> var_2022_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_feed_forward1_out_proj_bias_to_fp16, dilations = var_2022_dilations_0, groups = var_2022_groups_0, pad = var_2022_pad_0, pad_type = var_2022_pad_type_0, strides = var_2022_strides_0, weight = core_stacks_2_stack_layers_1_feed_forward1_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor<string, []>("op_2022_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_177_cast_fp16 = add(x = input_173_cast_fp16, y = var_2022_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
            tensor<int32, [4]> var_2024_begin_0 = const()[name = tensor<string, []>("op_2024_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2024_end_0 = const()[name = tensor<string, []>("op_2024_end_0"), val = tensor<int32, [4]>([1, 256, 1, 256])];
            tensor<bool, [4]> var_2024_end_mask_0 = const()[name = tensor<string, []>("op_2024_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 256, 1, 256]> var_2024_cast_fp16 = slice_by_index(begin = var_2024_begin_0, end = var_2024_end_0, end_mask = var_2024_end_mask_0, x = attn_weights_11_cast_fp16)[name = tensor<string, []>("op_2024_cast_fp16")];
            tensor<string, []> x_239_pad_type_0 = const()[name = tensor<string, []>("x_239_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_239_strides_0 = const()[name = tensor<string, []>("x_239_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_239_pad_0 = const()[name = tensor<string, []>("x_239_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_239_dilations_0 = const()[name = tensor<string, []>("x_239_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_239_groups_0 = const()[name = tensor<string, []>("x_239_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_2_stack_layers_1_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166211136)))];
            tensor<fp16, [1152]> core_stacks_2_stack_layers_1_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167390848)))];
            tensor<fp16, [1, 1152, 1, 256]> x_239_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_nonlin_attention_in_proj_bias_to_fp16, dilations = x_239_dilations_0, groups = x_239_groups_0, pad = x_239_pad_0, pad_type = x_239_pad_type_0, strides = x_239_strides_0, weight = core_stacks_2_stack_layers_1_nonlin_attention_in_proj_weight_to_fp16, x = input_177_cast_fp16)[name = tensor<string, []>("x_239_cast_fp16")];
            tensor<int32, [4]> var_2035_begin_0 = const()[name = tensor<string, []>("op_2035_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2035_end_0 = const()[name = tensor<string, []>("op_2035_end_0"), val = tensor<int32, [4]>([1, 384, 1, 256])];
            tensor<bool, [4]> var_2035_end_mask_0 = const()[name = tensor<string, []>("op_2035_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 256]> var_2035_cast_fp16 = slice_by_index(begin = var_2035_begin_0, end = var_2035_end_0, end_mask = var_2035_end_mask_0, x = x_239_cast_fp16)[name = tensor<string, []>("op_2035_cast_fp16")];
            tensor<fp16, [1, 384, 1, 256]> s_gate_11_cast_fp16 = tanh(x = var_2035_cast_fp16)[name = tensor<string, []>("s_gate_11_cast_fp16")];
            tensor<int32, [4]> var_2038_begin_0 = const()[name = tensor<string, []>("op_2038_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_2038_end_0 = const()[name = tensor<string, []>("op_2038_end_0"), val = tensor<int32, [4]>([1, 768, 1, 256])];
            tensor<bool, [4]> var_2038_end_mask_0 = const()[name = tensor<string, []>("op_2038_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 256]> var_2038_cast_fp16 = slice_by_index(begin = var_2038_begin_0, end = var_2038_end_0, end_mask = var_2038_end_mask_0, x = x_239_cast_fp16)[name = tensor<string, []>("op_2038_cast_fp16")];
            tensor<fp16, [1, 384, 1, 256]> var_2039_cast_fp16 = mul(x = var_2038_cast_fp16, y = s_gate_11_cast_fp16)[name = tensor<string, []>("op_2039_cast_fp16")];
            tensor<int32, [4]> y_11_begin_0 = const()[name = tensor<string, []>("y_11_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_11_end_0 = const()[name = tensor<string, []>("y_11_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 256])];
            tensor<bool, [4]> y_11_end_mask_0 = const()[name = tensor<string, []>("y_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 256]> y_11_cast_fp16 = slice_by_index(begin = y_11_begin_0, end = y_11_end_0, end_mask = y_11_end_mask_0, x = x_239_cast_fp16)[name = tensor<string, []>("y_11_cast_fp16")];
            tensor<int32, [4]> transpose_42_perm_0 = const()[name = tensor<string, []>("transpose_42_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_214 = const()[name = tensor<string, []>("concat_214"), val = tensor<int32, [3]>([1, 256, 256])];
            tensor<fp16, [1, 1, 256, 256]> transpose_42_cast_fp16 = transpose(perm = transpose_42_perm_0, x = var_2024_cast_fp16)[name = tensor<string, []>("transpose_256")];
            tensor<fp16, [1, 256, 256]> reshape_63_cast_fp16 = reshape(shape = concat_214, x = transpose_42_cast_fp16)[name = tensor<string, []>("reshape_63_cast_fp16")];
            tensor<int32, [4]> transpose_43_perm_0 = const()[name = tensor<string, []>("transpose_43_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_215 = const()[name = tensor<string, []>("concat_215"), val = tensor<int32, [3]>([1, 256, 384])];
            tensor<fp16, [1, 1, 256, 384]> transpose_43_cast_fp16 = transpose(perm = transpose_43_perm_0, x = var_2039_cast_fp16)[name = tensor<string, []>("transpose_255")];
            tensor<fp16, [1, 256, 384]> reshape_64_cast_fp16 = reshape(shape = concat_215, x = transpose_43_cast_fp16)[name = tensor<string, []>("reshape_64_cast_fp16")];
            tensor<bool, []> matmul_21_transpose_x_0 = const()[name = tensor<string, []>("matmul_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_21_transpose_y_0 = const()[name = tensor<string, []>("matmul_21_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 256, 384]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = reshape_63_cast_fp16, y = reshape_64_cast_fp16)[name = tensor<string, []>("matmul_21_cast_fp16")];
            tensor<int32, [4]> concat_219 = const()[name = tensor<string, []>("concat_219"), val = tensor<int32, [4]>([1, 1, 256, 384])];
            tensor<fp16, [1, 1, 256, 384]> reshape_65_cast_fp16 = reshape(shape = concat_219, x = matmul_21_cast_fp16)[name = tensor<string, []>("reshape_65_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = tensor<string, []>("v_11_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 256]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = reshape_65_cast_fp16)[name = tensor<string, []>("transpose_254")];
            tensor<fp16, [1, 384, 1, 256]> input_179_cast_fp16 = mul(x = v_11_cast_fp16, y = y_11_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
            tensor<string, []> var_2051_pad_type_0 = const()[name = tensor<string, []>("op_2051_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2051_strides_0 = const()[name = tensor<string, []>("op_2051_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2051_pad_0 = const()[name = tensor<string, []>("op_2051_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2051_dilations_0 = const()[name = tensor<string, []>("op_2051_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2051_groups_0 = const()[name = tensor<string, []>("op_2051_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_2_stack_layers_1_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167393216)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_1_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167786496)))];
            tensor<fp16, [1, 512, 1, 256]> var_2051_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_nonlin_attention_out_proj_bias_to_fp16, dilations = var_2051_dilations_0, groups = var_2051_groups_0, pad = var_2051_pad_0, pad_type = var_2051_pad_type_0, strides = var_2051_strides_0, weight = core_stacks_2_stack_layers_1_nonlin_attention_out_proj_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("op_2051_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_241_cast_fp16 = add(x = input_177_cast_fp16, y = var_2051_cast_fp16)[name = tensor<string, []>("x_241_cast_fp16")];
            tensor<string, []> var_2062_pad_type_0 = const()[name = tensor<string, []>("op_2062_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2062_strides_0 = const()[name = tensor<string, []>("op_2062_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2062_pad_0 = const()[name = tensor<string, []>("op_2062_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2062_dilations_0 = const()[name = tensor<string, []>("op_2062_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2062_groups_0 = const()[name = tensor<string, []>("op_2062_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_2_stack_layers_1_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167787584)))];
            tensor<fp16, [48]> core_stacks_2_stack_layers_1_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167836800)))];
            tensor<fp16, [1, 48, 1, 256]> var_2062_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_self_attn1_in_proj_bias_to_fp16, dilations = var_2062_dilations_0, groups = var_2062_groups_0, pad = var_2062_pad_0, pad_type = var_2062_pad_type_0, strides = var_2062_strides_0, weight = core_stacks_2_stack_layers_1_self_attn1_in_proj_weight_to_fp16, x = x_241_cast_fp16)[name = tensor<string, []>("op_2062_cast_fp16")];
            tensor<int32, [4]> var_2063 = const()[name = tensor<string, []>("op_2063"), val = tensor<int32, [4]>([4, 12, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> var_2064_cast_fp16 = reshape(shape = var_2063, x = var_2062_cast_fp16)[name = tensor<string, []>("op_2064_cast_fp16")];
            tensor<int32, [4]> transpose_44_perm_0 = const()[name = tensor<string, []>("transpose_44_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_224 = const()[name = tensor<string, []>("concat_224"), val = tensor<int32, [3]>([4, 256, 256])];
            tensor<fp16, [4, 1, 256, 256]> transpose_44_cast_fp16 = transpose(perm = transpose_44_perm_0, x = attn_weights_11_cast_fp16)[name = tensor<string, []>("transpose_253")];
            tensor<fp16, [4, 256, 256]> reshape_66_cast_fp16 = reshape(shape = concat_224, x = transpose_44_cast_fp16)[name = tensor<string, []>("reshape_66_cast_fp16")];
            tensor<int32, [4]> transpose_45_perm_0 = const()[name = tensor<string, []>("transpose_45_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_225 = const()[name = tensor<string, []>("concat_225"), val = tensor<int32, [3]>([4, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = var_2064_cast_fp16)[name = tensor<string, []>("transpose_252")];
            tensor<fp16, [4, 256, 12]> reshape_67_cast_fp16 = reshape(shape = concat_225, x = transpose_45_cast_fp16)[name = tensor<string, []>("reshape_67_cast_fp16")];
            tensor<bool, []> matmul_22_transpose_x_0 = const()[name = tensor<string, []>("matmul_22_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_22_transpose_y_0 = const()[name = tensor<string, []>("matmul_22_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 12]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = reshape_66_cast_fp16, y = reshape_67_cast_fp16)[name = tensor<string, []>("matmul_22_cast_fp16")];
            tensor<int32, [4]> concat_229 = const()[name = tensor<string, []>("concat_229"), val = tensor<int32, [4]>([4, 1, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> reshape_68_cast_fp16 = reshape(shape = concat_229, x = matmul_22_cast_fp16)[name = tensor<string, []>("reshape_68_cast_fp16")];
            tensor<int32, [4]> out_21_perm_0 = const()[name = tensor<string, []>("out_21_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_2067 = const()[name = tensor<string, []>("op_2067"), val = tensor<int32, [4]>([1, 48, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> out_21_cast_fp16 = transpose(perm = out_21_perm_0, x = reshape_68_cast_fp16)[name = tensor<string, []>("transpose_251")];
            tensor<fp16, [1, 48, 1, 256]> input_181_cast_fp16 = reshape(shape = var_2067, x = out_21_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
            tensor<string, []> var_2075_pad_type_0 = const()[name = tensor<string, []>("op_2075_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2075_strides_0 = const()[name = tensor<string, []>("op_2075_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2075_pad_0 = const()[name = tensor<string, []>("op_2075_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2075_dilations_0 = const()[name = tensor<string, []>("op_2075_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2075_groups_0 = const()[name = tensor<string, []>("op_2075_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_2_stack_layers_1_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167836992)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_1_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167886208)))];
            tensor<fp16, [1, 512, 1, 256]> var_2075_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_self_attn1_out_proj_bias_to_fp16, dilations = var_2075_dilations_0, groups = var_2075_groups_0, pad = var_2075_pad_0, pad_type = var_2075_pad_type_0, strides = var_2075_strides_0, weight = core_stacks_2_stack_layers_1_self_attn1_out_proj_weight_to_fp16, x = input_181_cast_fp16)[name = tensor<string, []>("op_2075_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_243_cast_fp16 = add(x = x_241_cast_fp16, y = var_2075_cast_fp16)[name = tensor<string, []>("x_243_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_245_cast_fp16 = add(x = x_243_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("x_245_cast_fp16")];
            tensor<string, []> x_247_pad_type_0 = const()[name = tensor<string, []>("x_247_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_247_strides_0 = const()[name = tensor<string, []>("x_247_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_247_pad_0 = const()[name = tensor<string, []>("x_247_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_247_dilations_0 = const()[name = tensor<string, []>("x_247_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_247_groups_0 = const()[name = tensor<string, []>("x_247_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_2_stack_layers_1_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167887296)))];
            tensor<fp16, [1024]> core_stacks_2_stack_layers_1_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168935936)))];
            tensor<fp16, [1, 1024, 1, 256]> x_247_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_conv_module1_in_proj_bias_to_fp16, dilations = x_247_dilations_0, groups = x_247_groups_0, pad = x_247_pad_0, pad_type = x_247_pad_type_0, strides = x_247_strides_0, weight = core_stacks_2_stack_layers_1_conv_module1_in_proj_weight_to_fp16, x = x_245_cast_fp16)[name = tensor<string, []>("x_247_cast_fp16")];
            tensor<int32, [4]> var_2089_begin_0 = const()[name = tensor<string, []>("op_2089_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2089_end_0 = const()[name = tensor<string, []>("op_2089_end_0"), val = tensor<int32, [4]>([1, 512, 1, 256])];
            tensor<bool, [4]> var_2089_end_mask_0 = const()[name = tensor<string, []>("op_2089_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2089_cast_fp16 = slice_by_index(begin = var_2089_begin_0, end = var_2089_end_0, end_mask = var_2089_end_mask_0, x = x_247_cast_fp16)[name = tensor<string, []>("op_2089_cast_fp16")];
            tensor<int32, [4]> var_2091_begin_0 = const()[name = tensor<string, []>("op_2091_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_2091_end_0 = const()[name = tensor<string, []>("op_2091_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 256])];
            tensor<bool, [4]> var_2091_end_mask_0 = const()[name = tensor<string, []>("op_2091_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2091_cast_fp16 = slice_by_index(begin = var_2091_begin_0, end = var_2091_end_0, end_mask = var_2091_end_mask_0, x = x_247_cast_fp16)[name = tensor<string, []>("op_2091_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2092_cast_fp16 = sigmoid(x = var_2091_cast_fp16)[name = tensor<string, []>("op_2092_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_249_cast_fp16 = mul(x = var_2089_cast_fp16, y = var_2092_cast_fp16)[name = tensor<string, []>("x_249_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_183_cast_fp16 = mul(x = x_249_cast_fp16, y = conv_keep_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
            tensor<string, []> x_251_pad_type_0 = const()[name = tensor<string, []>("x_251_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_251_pad_0 = const()[name = tensor<string, []>("x_251_pad_0"), val = tensor<int32, [4]>([0, 0, 3, 3])];
            tensor<int32, []> x_251_groups_0 = const()[name = tensor<string, []>("x_251_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_251_strides_0 = const()[name = tensor<string, []>("x_251_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_251_dilations_0 = const()[name = tensor<string, []>("x_251_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 7]> core_stacks_2_stack_layers_1_conv_module1_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module1_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 7]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168938048)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_1_conv_module1_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module1_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168945280)))];
            tensor<fp16, [1, 512, 1, 256]> x_251_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_conv_module1_dw_bias_to_fp16, dilations = x_251_dilations_0, groups = x_251_groups_0, pad = x_251_pad_0, pad_type = x_251_pad_type_0, strides = x_251_strides_0, weight = core_stacks_2_stack_layers_1_conv_module1_dw_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("x_251_cast_fp16")];
            tensor<fp16, []> var_2102_to_fp16 = const()[name = tensor<string, []>("op_2102_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> x_253_cast_fp16 = sub(x = x_251_cast_fp16, y = var_2102_to_fp16)[name = tensor<string, []>("x_253_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2104_cast_fp16 = relu(x = x_253_cast_fp16)[name = tensor<string, []>("op_2104_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2105_cast_fp16 = abs(x = x_253_cast_fp16)[name = tensor<string, []>("op_2105_cast_fp16")];
            tensor<fp16, []> const_45_promoted_to_fp16 = const()[name = tensor<string, []>("const_45_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2106_cast_fp16 = mul(x = var_2105_cast_fp16, y = const_45_promoted_to_fp16)[name = tensor<string, []>("op_2106_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2107_cast_fp16 = exp(x = var_2106_cast_fp16)[name = tensor<string, []>("op_2107_cast_fp16")];
            tensor<fp32, []> var_2108_epsilon_0 = const()[name = tensor<string, []>("op_2108_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2108_cast_fp16 = log(epsilon = var_2108_epsilon_0, x = var_2107_cast_fp16)[name = tensor<string, []>("op_2108_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2109_cast_fp16 = add(x = var_2104_cast_fp16, y = var_2108_cast_fp16)[name = tensor<string, []>("op_2109_cast_fp16")];
            tensor<fp16, []> var_2110_to_fp16 = const()[name = tensor<string, []>("op_2110_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 256]> var_2111_cast_fp16 = mul(x = x_251_cast_fp16, y = var_2110_to_fp16)[name = tensor<string, []>("op_2111_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2112_cast_fp16 = sub(x = var_2109_cast_fp16, y = var_2111_cast_fp16)[name = tensor<string, []>("op_2112_cast_fp16")];
            tensor<fp16, []> var_2113_to_fp16 = const()[name = tensor<string, []>("op_2113_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 256]> input_185_cast_fp16 = sub(x = var_2112_cast_fp16, y = var_2113_to_fp16)[name = tensor<string, []>("input_185_cast_fp16")];
            tensor<string, []> var_2121_pad_type_0 = const()[name = tensor<string, []>("op_2121_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2121_strides_0 = const()[name = tensor<string, []>("op_2121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2121_pad_0 = const()[name = tensor<string, []>("op_2121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2121_dilations_0 = const()[name = tensor<string, []>("op_2121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2121_groups_0 = const()[name = tensor<string, []>("op_2121_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_2_stack_layers_1_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168946368)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_1_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169470720)))];
            tensor<fp16, [1, 512, 1, 256]> var_2121_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_conv_module1_out_proj_bias_to_fp16, dilations = var_2121_dilations_0, groups = var_2121_groups_0, pad = var_2121_pad_0, pad_type = var_2121_pad_type_0, strides = var_2121_strides_0, weight = core_stacks_2_stack_layers_1_conv_module1_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("op_2121_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_187_cast_fp16 = add(x = x_245_cast_fp16, y = var_2121_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
            tensor<string, []> x_255_pad_type_0 = const()[name = tensor<string, []>("x_255_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_255_strides_0 = const()[name = tensor<string, []>("x_255_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_255_pad_0 = const()[name = tensor<string, []>("x_255_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_255_dilations_0 = const()[name = tensor<string, []>("x_255_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_255_groups_0 = const()[name = tensor<string, []>("x_255_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_2_stack_layers_1_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169471808)))];
            tensor<fp16, [1536]> core_stacks_2_stack_layers_1_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171044736)))];
            tensor<fp16, [1, 1536, 1, 256]> x_255_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_feed_forward2_in_proj_bias_to_fp16, dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = core_stacks_2_stack_layers_1_feed_forward2_in_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("x_255_cast_fp16")];
            tensor<fp16, []> var_2132_to_fp16 = const()[name = tensor<string, []>("op_2132_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 256]> x_257_cast_fp16 = sub(x = x_255_cast_fp16, y = var_2132_to_fp16)[name = tensor<string, []>("x_257_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2134_cast_fp16 = relu(x = x_257_cast_fp16)[name = tensor<string, []>("op_2134_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2135_cast_fp16 = abs(x = x_257_cast_fp16)[name = tensor<string, []>("op_2135_cast_fp16")];
            tensor<fp16, []> const_46_promoted_to_fp16 = const()[name = tensor<string, []>("const_46_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 256]> var_2136_cast_fp16 = mul(x = var_2135_cast_fp16, y = const_46_promoted_to_fp16)[name = tensor<string, []>("op_2136_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2137_cast_fp16 = exp(x = var_2136_cast_fp16)[name = tensor<string, []>("op_2137_cast_fp16")];
            tensor<fp32, []> var_2138_epsilon_0 = const()[name = tensor<string, []>("op_2138_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 256]> var_2138_cast_fp16 = log(epsilon = var_2138_epsilon_0, x = var_2137_cast_fp16)[name = tensor<string, []>("op_2138_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2139_cast_fp16 = add(x = var_2134_cast_fp16, y = var_2138_cast_fp16)[name = tensor<string, []>("op_2139_cast_fp16")];
            tensor<fp16, []> var_2140_to_fp16 = const()[name = tensor<string, []>("op_2140_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 256]> var_2141_cast_fp16 = mul(x = x_255_cast_fp16, y = var_2140_to_fp16)[name = tensor<string, []>("op_2141_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2142_cast_fp16 = sub(x = var_2139_cast_fp16, y = var_2141_cast_fp16)[name = tensor<string, []>("op_2142_cast_fp16")];
            tensor<fp16, []> var_2143_to_fp16 = const()[name = tensor<string, []>("op_2143_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 256]> input_189_cast_fp16 = sub(x = var_2142_cast_fp16, y = var_2143_to_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
            tensor<string, []> var_2151_pad_type_0 = const()[name = tensor<string, []>("op_2151_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2151_strides_0 = const()[name = tensor<string, []>("op_2151_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2151_pad_0 = const()[name = tensor<string, []>("op_2151_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2151_dilations_0 = const()[name = tensor<string, []>("op_2151_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2151_groups_0 = const()[name = tensor<string, []>("op_2151_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_2_stack_layers_1_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171047872)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_1_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172620800)))];
            tensor<fp16, [1, 512, 1, 256]> var_2151_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_feed_forward2_out_proj_bias_to_fp16, dilations = var_2151_dilations_0, groups = var_2151_groups_0, pad = var_2151_pad_0, pad_type = var_2151_pad_type_0, strides = var_2151_strides_0, weight = core_stacks_2_stack_layers_1_feed_forward2_out_proj_weight_to_fp16, x = input_189_cast_fp16)[name = tensor<string, []>("op_2151_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> src_23_cast_fp16 = add(x = input_187_cast_fp16, y = var_2151_cast_fp16)[name = tensor<string, []>("src_23_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2154_cast_fp16 = sub(x = src_23_cast_fp16, y = input_171_cast_fp16)[name = tensor<string, []>("op_2154_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_1_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172621888)))];
            tensor<fp16, [1, 512, 1, 256]> var_2155_cast_fp16 = mul(x = var_2154_cast_fp16, y = core_stacks_2_stack_layers_1_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_2155_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_259_cast_fp16 = add(x = input_171_cast_fp16, y = var_2155_cast_fp16)[name = tensor<string, []>("x_259_cast_fp16")];
            tensor<string, []> var_2166_pad_type_0 = const()[name = tensor<string, []>("op_2166_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2166_strides_0 = const()[name = tensor<string, []>("op_2166_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2166_pad_0 = const()[name = tensor<string, []>("op_2166_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2166_dilations_0 = const()[name = tensor<string, []>("op_2166_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2166_groups_0 = const()[name = tensor<string, []>("op_2166_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_2_stack_layers_1_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172622976)))];
            tensor<fp16, [48]> core_stacks_2_stack_layers_1_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172672192)))];
            tensor<fp16, [1, 48, 1, 256]> var_2166_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_self_attn2_in_proj_bias_to_fp16, dilations = var_2166_dilations_0, groups = var_2166_groups_0, pad = var_2166_pad_0, pad_type = var_2166_pad_type_0, strides = var_2166_strides_0, weight = core_stacks_2_stack_layers_1_self_attn2_in_proj_weight_to_fp16, x = x_259_cast_fp16)[name = tensor<string, []>("op_2166_cast_fp16")];
            tensor<int32, [4]> var_2167 = const()[name = tensor<string, []>("op_2167"), val = tensor<int32, [4]>([4, 12, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> var_2168_cast_fp16 = reshape(shape = var_2167, x = var_2166_cast_fp16)[name = tensor<string, []>("op_2168_cast_fp16")];
            tensor<int32, [4]> transpose_47_perm_0 = const()[name = tensor<string, []>("transpose_47_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_235 = const()[name = tensor<string, []>("concat_235"), val = tensor<int32, [3]>([4, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> transpose_47_cast_fp16 = transpose(perm = transpose_47_perm_0, x = var_2168_cast_fp16)[name = tensor<string, []>("transpose_250")];
            tensor<fp16, [4, 256, 12]> reshape_70_cast_fp16 = reshape(shape = concat_235, x = transpose_47_cast_fp16)[name = tensor<string, []>("reshape_70_cast_fp16")];
            tensor<bool, []> matmul_23_transpose_x_0 = const()[name = tensor<string, []>("matmul_23_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_23_transpose_y_0 = const()[name = tensor<string, []>("matmul_23_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 12]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = reshape_66_cast_fp16, y = reshape_70_cast_fp16)[name = tensor<string, []>("matmul_23_cast_fp16")];
            tensor<int32, [4]> concat_239 = const()[name = tensor<string, []>("concat_239"), val = tensor<int32, [4]>([4, 1, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> reshape_71_cast_fp16 = reshape(shape = concat_239, x = matmul_23_cast_fp16)[name = tensor<string, []>("reshape_71_cast_fp16")];
            tensor<int32, [4]> out_23_perm_0 = const()[name = tensor<string, []>("out_23_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_2171 = const()[name = tensor<string, []>("op_2171"), val = tensor<int32, [4]>([1, 48, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> out_23_cast_fp16 = transpose(perm = out_23_perm_0, x = reshape_71_cast_fp16)[name = tensor<string, []>("transpose_249")];
            tensor<fp16, [1, 48, 1, 256]> input_191_cast_fp16 = reshape(shape = var_2171, x = out_23_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
            tensor<string, []> var_2179_pad_type_0 = const()[name = tensor<string, []>("op_2179_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2179_strides_0 = const()[name = tensor<string, []>("op_2179_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2179_pad_0 = const()[name = tensor<string, []>("op_2179_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2179_dilations_0 = const()[name = tensor<string, []>("op_2179_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2179_groups_0 = const()[name = tensor<string, []>("op_2179_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_2_stack_layers_1_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172672384)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_1_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172721600)))];
            tensor<fp16, [1, 512, 1, 256]> var_2179_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_self_attn2_out_proj_bias_to_fp16, dilations = var_2179_dilations_0, groups = var_2179_groups_0, pad = var_2179_pad_0, pad_type = var_2179_pad_type_0, strides = var_2179_strides_0, weight = core_stacks_2_stack_layers_1_self_attn2_out_proj_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("op_2179_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_261_cast_fp16 = add(x = x_259_cast_fp16, y = var_2179_cast_fp16)[name = tensor<string, []>("x_261_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_263_cast_fp16 = add(x = x_261_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("x_263_cast_fp16")];
            tensor<string, []> x_265_pad_type_0 = const()[name = tensor<string, []>("x_265_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_265_strides_0 = const()[name = tensor<string, []>("x_265_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_265_pad_0 = const()[name = tensor<string, []>("x_265_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_265_dilations_0 = const()[name = tensor<string, []>("x_265_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_265_groups_0 = const()[name = tensor<string, []>("x_265_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_2_stack_layers_1_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172722688)))];
            tensor<fp16, [1024]> core_stacks_2_stack_layers_1_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173771328)))];
            tensor<fp16, [1, 1024, 1, 256]> x_265_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_conv_module2_in_proj_bias_to_fp16, dilations = x_265_dilations_0, groups = x_265_groups_0, pad = x_265_pad_0, pad_type = x_265_pad_type_0, strides = x_265_strides_0, weight = core_stacks_2_stack_layers_1_conv_module2_in_proj_weight_to_fp16, x = x_263_cast_fp16)[name = tensor<string, []>("x_265_cast_fp16")];
            tensor<int32, [4]> var_2193_begin_0 = const()[name = tensor<string, []>("op_2193_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2193_end_0 = const()[name = tensor<string, []>("op_2193_end_0"), val = tensor<int32, [4]>([1, 512, 1, 256])];
            tensor<bool, [4]> var_2193_end_mask_0 = const()[name = tensor<string, []>("op_2193_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2193_cast_fp16 = slice_by_index(begin = var_2193_begin_0, end = var_2193_end_0, end_mask = var_2193_end_mask_0, x = x_265_cast_fp16)[name = tensor<string, []>("op_2193_cast_fp16")];
            tensor<int32, [4]> var_2195_begin_0 = const()[name = tensor<string, []>("op_2195_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_2195_end_0 = const()[name = tensor<string, []>("op_2195_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 256])];
            tensor<bool, [4]> var_2195_end_mask_0 = const()[name = tensor<string, []>("op_2195_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2195_cast_fp16 = slice_by_index(begin = var_2195_begin_0, end = var_2195_end_0, end_mask = var_2195_end_mask_0, x = x_265_cast_fp16)[name = tensor<string, []>("op_2195_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2196_cast_fp16 = sigmoid(x = var_2195_cast_fp16)[name = tensor<string, []>("op_2196_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_267_cast_fp16 = mul(x = var_2193_cast_fp16, y = var_2196_cast_fp16)[name = tensor<string, []>("x_267_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_193_cast_fp16 = mul(x = x_267_cast_fp16, y = conv_keep_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<string, []> x_269_pad_type_0 = const()[name = tensor<string, []>("x_269_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_269_pad_0 = const()[name = tensor<string, []>("x_269_pad_0"), val = tensor<int32, [4]>([0, 0, 3, 3])];
            tensor<int32, []> x_269_groups_0 = const()[name = tensor<string, []>("x_269_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_269_strides_0 = const()[name = tensor<string, []>("x_269_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_269_dilations_0 = const()[name = tensor<string, []>("x_269_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 7]> core_stacks_2_stack_layers_1_conv_module2_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module2_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 7]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173773440)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_1_conv_module2_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module2_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173780672)))];
            tensor<fp16, [1, 512, 1, 256]> x_269_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_conv_module2_dw_bias_to_fp16, dilations = x_269_dilations_0, groups = x_269_groups_0, pad = x_269_pad_0, pad_type = x_269_pad_type_0, strides = x_269_strides_0, weight = core_stacks_2_stack_layers_1_conv_module2_dw_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("x_269_cast_fp16")];
            tensor<fp16, []> var_2206_to_fp16 = const()[name = tensor<string, []>("op_2206_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> x_271_cast_fp16 = sub(x = x_269_cast_fp16, y = var_2206_to_fp16)[name = tensor<string, []>("x_271_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2208_cast_fp16 = relu(x = x_271_cast_fp16)[name = tensor<string, []>("op_2208_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2209_cast_fp16 = abs(x = x_271_cast_fp16)[name = tensor<string, []>("op_2209_cast_fp16")];
            tensor<fp16, []> const_48_promoted_to_fp16 = const()[name = tensor<string, []>("const_48_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2210_cast_fp16 = mul(x = var_2209_cast_fp16, y = const_48_promoted_to_fp16)[name = tensor<string, []>("op_2210_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2211_cast_fp16 = exp(x = var_2210_cast_fp16)[name = tensor<string, []>("op_2211_cast_fp16")];
            tensor<fp32, []> var_2212_epsilon_0 = const()[name = tensor<string, []>("op_2212_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2212_cast_fp16 = log(epsilon = var_2212_epsilon_0, x = var_2211_cast_fp16)[name = tensor<string, []>("op_2212_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2213_cast_fp16 = add(x = var_2208_cast_fp16, y = var_2212_cast_fp16)[name = tensor<string, []>("op_2213_cast_fp16")];
            tensor<fp16, []> var_2214_to_fp16 = const()[name = tensor<string, []>("op_2214_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 256]> var_2215_cast_fp16 = mul(x = x_269_cast_fp16, y = var_2214_to_fp16)[name = tensor<string, []>("op_2215_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2216_cast_fp16 = sub(x = var_2213_cast_fp16, y = var_2215_cast_fp16)[name = tensor<string, []>("op_2216_cast_fp16")];
            tensor<fp16, []> var_2217_to_fp16 = const()[name = tensor<string, []>("op_2217_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 256]> input_195_cast_fp16 = sub(x = var_2216_cast_fp16, y = var_2217_to_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
            tensor<string, []> var_2225_pad_type_0 = const()[name = tensor<string, []>("op_2225_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2225_strides_0 = const()[name = tensor<string, []>("op_2225_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2225_pad_0 = const()[name = tensor<string, []>("op_2225_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2225_dilations_0 = const()[name = tensor<string, []>("op_2225_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2225_groups_0 = const()[name = tensor<string, []>("op_2225_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_2_stack_layers_1_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173781760)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_1_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174306112)))];
            tensor<fp16, [1, 512, 1, 256]> var_2225_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_conv_module2_out_proj_bias_to_fp16, dilations = var_2225_dilations_0, groups = var_2225_groups_0, pad = var_2225_pad_0, pad_type = var_2225_pad_type_0, strides = var_2225_strides_0, weight = core_stacks_2_stack_layers_1_conv_module2_out_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("op_2225_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_197_cast_fp16 = add(x = x_263_cast_fp16, y = var_2225_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")];
            tensor<string, []> x_273_pad_type_0 = const()[name = tensor<string, []>("x_273_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_273_strides_0 = const()[name = tensor<string, []>("x_273_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_273_pad_0 = const()[name = tensor<string, []>("x_273_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_273_dilations_0 = const()[name = tensor<string, []>("x_273_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_273_groups_0 = const()[name = tensor<string, []>("x_273_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_2_stack_layers_1_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174307200)))];
            tensor<fp16, [1920]> core_stacks_2_stack_layers_1_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176273344)))];
            tensor<fp16, [1, 1920, 1, 256]> x_273_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_feed_forward3_in_proj_bias_to_fp16, dilations = x_273_dilations_0, groups = x_273_groups_0, pad = x_273_pad_0, pad_type = x_273_pad_type_0, strides = x_273_strides_0, weight = core_stacks_2_stack_layers_1_feed_forward3_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor<string, []>("x_273_cast_fp16")];
            tensor<fp16, []> var_2236_to_fp16 = const()[name = tensor<string, []>("op_2236_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 256]> x_275_cast_fp16 = sub(x = x_273_cast_fp16, y = var_2236_to_fp16)[name = tensor<string, []>("x_275_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2238_cast_fp16 = relu(x = x_275_cast_fp16)[name = tensor<string, []>("op_2238_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2239_cast_fp16 = abs(x = x_275_cast_fp16)[name = tensor<string, []>("op_2239_cast_fp16")];
            tensor<fp16, []> const_49_promoted_to_fp16 = const()[name = tensor<string, []>("const_49_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 256]> var_2240_cast_fp16 = mul(x = var_2239_cast_fp16, y = const_49_promoted_to_fp16)[name = tensor<string, []>("op_2240_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2241_cast_fp16 = exp(x = var_2240_cast_fp16)[name = tensor<string, []>("op_2241_cast_fp16")];
            tensor<fp32, []> var_2242_epsilon_0 = const()[name = tensor<string, []>("op_2242_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 256]> var_2242_cast_fp16 = log(epsilon = var_2242_epsilon_0, x = var_2241_cast_fp16)[name = tensor<string, []>("op_2242_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2243_cast_fp16 = add(x = var_2238_cast_fp16, y = var_2242_cast_fp16)[name = tensor<string, []>("op_2243_cast_fp16")];
            tensor<fp16, []> var_2244_to_fp16 = const()[name = tensor<string, []>("op_2244_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 256]> var_2245_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2244_to_fp16)[name = tensor<string, []>("op_2245_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2246_cast_fp16 = sub(x = var_2243_cast_fp16, y = var_2245_cast_fp16)[name = tensor<string, []>("op_2246_cast_fp16")];
            tensor<fp16, []> var_2247_to_fp16 = const()[name = tensor<string, []>("op_2247_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 256]> input_199_cast_fp16 = sub(x = var_2246_cast_fp16, y = var_2247_to_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
            tensor<string, []> var_2255_pad_type_0 = const()[name = tensor<string, []>("op_2255_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2255_strides_0 = const()[name = tensor<string, []>("op_2255_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2255_pad_0 = const()[name = tensor<string, []>("op_2255_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2255_dilations_0 = const()[name = tensor<string, []>("op_2255_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2255_groups_0 = const()[name = tensor<string, []>("op_2255_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_2_stack_layers_1_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176277248)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_1_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178243392)))];
            tensor<fp16, [1, 512, 1, 256]> var_2255_cast_fp16 = conv(bias = core_stacks_2_stack_layers_1_feed_forward3_out_proj_bias_to_fp16, dilations = var_2255_dilations_0, groups = var_2255_groups_0, pad = var_2255_pad_0, pad_type = var_2255_pad_type_0, strides = var_2255_strides_0, weight = core_stacks_2_stack_layers_1_feed_forward3_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("op_2255_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_277_cast_fp16 = add(x = input_197_cast_fp16, y = var_2255_cast_fp16)[name = tensor<string, []>("x_277_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_1_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178244480)))];
            tensor<fp16, [1, 512, 1, 256]> d_11_cast_fp16 = sub(x = x_277_cast_fp16, y = core_stacks_2_stack_layers_1_norm_bias_to_fp16)[name = tensor<string, []>("d_11_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2260_cast_fp16 = mul(x = d_11_cast_fp16, y = d_11_cast_fp16)[name = tensor<string, []>("op_2260_cast_fp16")];
            tensor<int32, [1]> var_2262_axes_0 = const()[name = tensor<string, []>("op_2262_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_2262_keep_dims_0 = const()[name = tensor<string, []>("op_2262_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 256]> var_2262_cast_fp16 = reduce_mean(axes = var_2262_axes_0, keep_dims = var_2262_keep_dims_0, x = var_2260_cast_fp16)[name = tensor<string, []>("op_2262_cast_fp16")];
            tensor<fp32, []> scales_11_epsilon_0 = const()[name = tensor<string, []>("scales_11_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 256]> scales_11_cast_fp16 = rsqrt(epsilon = scales_11_epsilon_0, x = var_2262_cast_fp16)[name = tensor<string, []>("scales_11_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2264_cast_fp16 = mul(x = x_277_cast_fp16, y = scales_11_cast_fp16)[name = tensor<string, []>("op_2264_cast_fp16")];
            tensor<fp16, []> core_stacks_2_stack_layers_1_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.024p-2)];
            tensor<fp16, [1, 512, 1, 256]> src_25_cast_fp16 = mul(x = var_2264_cast_fp16, y = core_stacks_2_stack_layers_1_norm_scale_to_fp16)[name = tensor<string, []>("src_25_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2267_cast_fp16 = sub(x = src_25_cast_fp16, y = input_171_cast_fp16)[name = tensor<string, []>("op_2267_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_1_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_1_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178245568)))];
            tensor<fp16, [1, 512, 1, 256]> var_2268_cast_fp16 = mul(x = var_2267_cast_fp16, y = core_stacks_2_stack_layers_1_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_2268_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_201_cast_fp16 = add(x = input_171_cast_fp16, y = var_2268_cast_fp16)[name = tensor<string, []>("input_201_cast_fp16")];
            tensor<string, []> proj_13_pad_type_0 = const()[name = tensor<string, []>("proj_13_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_13_strides_0 = const()[name = tensor<string, []>("proj_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_13_pad_0 = const()[name = tensor<string, []>("proj_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_13_dilations_0 = const()[name = tensor<string, []>("proj_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_13_groups_0 = const()[name = tensor<string, []>("proj_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_2_stack_layers_2_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178246656)))];
            tensor<fp16, [384]> core_stacks_2_stack_layers_2_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178639936)))];
            tensor<fp16, [1, 384, 1, 256]> proj_13_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_13_dilations_0, groups = proj_13_groups_0, pad = proj_13_pad_0, pad_type = proj_13_pad_type_0, strides = proj_13_strides_0, weight = core_stacks_2_stack_layers_2_self_attn_weights_in_proj_weight_to_fp16, x = input_201_cast_fp16)[name = tensor<string, []>("proj_13_cast_fp16")];
            tensor<int32, [4]> var_2291_begin_0 = const()[name = tensor<string, []>("op_2291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2291_end_0 = const()[name = tensor<string, []>("op_2291_end_0"), val = tensor<int32, [4]>([1, 128, 1, 256])];
            tensor<bool, [4]> var_2291_end_mask_0 = const()[name = tensor<string, []>("op_2291_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_2291_cast_fp16 = slice_by_index(begin = var_2291_begin_0, end = var_2291_end_0, end_mask = var_2291_end_mask_0, x = proj_13_cast_fp16)[name = tensor<string, []>("op_2291_cast_fp16")];
            tensor<int32, [4]> var_2292 = const()[name = tensor<string, []>("op_2292"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> var_2293_cast_fp16 = reshape(shape = var_2292, x = var_2291_cast_fp16)[name = tensor<string, []>("op_2293_cast_fp16")];
            tensor<int32, [4]> var_2295_begin_0 = const()[name = tensor<string, []>("op_2295_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_2295_end_0 = const()[name = tensor<string, []>("op_2295_end_0"), val = tensor<int32, [4]>([1, 256, 1, 256])];
            tensor<bool, [4]> var_2295_end_mask_0 = const()[name = tensor<string, []>("op_2295_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_2295_cast_fp16 = slice_by_index(begin = var_2295_begin_0, end = var_2295_end_0, end_mask = var_2295_end_mask_0, x = proj_13_cast_fp16)[name = tensor<string, []>("op_2295_cast_fp16")];
            tensor<int32, [4]> var_2296 = const()[name = tensor<string, []>("op_2296"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> var_2297_cast_fp16 = reshape(shape = var_2296, x = var_2295_cast_fp16)[name = tensor<string, []>("op_2297_cast_fp16")];
            tensor<int32, [4]> var_2299_begin_0 = const()[name = tensor<string, []>("op_2299_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_2299_end_0 = const()[name = tensor<string, []>("op_2299_end_0"), val = tensor<int32, [4]>([1, 384, 1, 256])];
            tensor<bool, [4]> var_2299_end_mask_0 = const()[name = tensor<string, []>("op_2299_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_2299_cast_fp16 = slice_by_index(begin = var_2299_begin_0, end = var_2299_end_0, end_mask = var_2299_end_mask_0, x = proj_13_cast_fp16)[name = tensor<string, []>("op_2299_cast_fp16")];
            tensor<int32, [4]> var_2300 = const()[name = tensor<string, []>("op_2300"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> pt_13_cast_fp16 = reshape(shape = var_2300, x = var_2299_cast_fp16)[name = tensor<string, []>("pt_13_cast_fp16")];
            tensor<int32, [4]> transpose_48_perm_0 = const()[name = tensor<string, []>("transpose_48_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_244 = const()[name = tensor<string, []>("concat_244"), val = tensor<int32, [3]>([4, 256, 32])];
            tensor<fp16, [4, 1, 256, 32]> transpose_48_cast_fp16 = transpose(perm = transpose_48_perm_0, x = var_2293_cast_fp16)[name = tensor<string, []>("transpose_248")];
            tensor<fp16, [4, 256, 32]> reshape_72_cast_fp16 = reshape(shape = concat_244, x = transpose_48_cast_fp16)[name = tensor<string, []>("reshape_72_cast_fp16")];
            tensor<int32, [4]> transpose_49_perm_0 = const()[name = tensor<string, []>("transpose_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_245 = const()[name = tensor<string, []>("concat_245"), val = tensor<int32, [3]>([4, 32, 256])];
            tensor<fp16, [4, 1, 32, 256]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = var_2297_cast_fp16)[name = tensor<string, []>("transpose_247")];
            tensor<fp16, [4, 32, 256]> reshape_73_cast_fp16 = reshape(shape = concat_245, x = transpose_49_cast_fp16)[name = tensor<string, []>("reshape_73_cast_fp16")];
            tensor<bool, []> matmul_24_transpose_x_0 = const()[name = tensor<string, []>("matmul_24_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_24_transpose_y_0 = const()[name = tensor<string, []>("matmul_24_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 256]> matmul_24_cast_fp16 = matmul(transpose_x = matmul_24_transpose_x_0, transpose_y = matmul_24_transpose_y_0, x = reshape_72_cast_fp16, y = reshape_73_cast_fp16)[name = tensor<string, []>("matmul_24_cast_fp16")];
            tensor<int32, [4]> concat_249 = const()[name = tensor<string, []>("concat_249"), val = tensor<int32, [4]>([4, 1, 256, 256])];
            tensor<fp16, [4, 1, 256, 256]> reshape_74_cast_fp16 = reshape(shape = concat_249, x = matmul_24_cast_fp16)[name = tensor<string, []>("reshape_74_cast_fp16")];
            tensor<int32, [4]> attn_scores_37_perm_0 = const()[name = tensor<string, []>("attn_scores_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2304 = const()[name = tensor<string, []>("op_2304"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 256, 1]> var_2305_cast_fp16 = transpose(perm = var_2304, x = pt_13_cast_fp16)[name = tensor<string, []>("transpose_245")];
            tensor<fp16, [4, 32, 256, 256]> var_2306_cast_fp16 = mul(x = var_2305_cast_fp16, y = core_pos_basis_256_to_fp16)[name = tensor<string, []>("op_2306_cast_fp16")];
            tensor<int32, [1]> pos_scores_13_axes_0 = const()[name = tensor<string, []>("pos_scores_13_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_13_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_13_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 256, 256]> pos_scores_13_cast_fp16 = reduce_sum(axes = pos_scores_13_axes_0, keep_dims = pos_scores_13_keep_dims_0, x = var_2306_cast_fp16)[name = tensor<string, []>("pos_scores_13_cast_fp16")];
            tensor<int32, [4]> var_2309 = const()[name = tensor<string, []>("op_2309"), val = tensor<int32, [4]>([4, 256, 1, 256])];
            tensor<fp16, [4, 256, 1, 256]> var_2310_cast_fp16 = reshape(shape = var_2309, x = pos_scores_13_cast_fp16)[name = tensor<string, []>("op_2310_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_37_cast_fp16 = transpose(perm = attn_scores_37_perm_0, x = reshape_74_cast_fp16)[name = tensor<string, []>("transpose_246")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_39_cast_fp16 = add(x = attn_scores_37_cast_fp16, y = var_2310_cast_fp16)[name = tensor<string, []>("attn_scores_39_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_41_cast_fp16 = add(x = attn_scores_39_cast_fp16, y = attn_bias_cast_fp16)[name = tensor<string, []>("attn_scores_41_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_weights_13_cast_fp16 = softmax(axis = var_41, x = attn_scores_41_cast_fp16)[name = tensor<string, []>("attn_weights_13_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_203_cast_fp16 = add(x = input_201_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
            tensor<string, []> x_279_pad_type_0 = const()[name = tensor<string, []>("x_279_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_279_strides_0 = const()[name = tensor<string, []>("x_279_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_279_pad_0 = const()[name = tensor<string, []>("x_279_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_279_dilations_0 = const()[name = tensor<string, []>("x_279_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_279_groups_0 = const()[name = tensor<string, []>("x_279_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_2_stack_layers_2_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178640768)))];
            tensor<fp16, [1152]> core_stacks_2_stack_layers_2_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179820480)))];
            tensor<fp16, [1, 1152, 1, 256]> x_279_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_feed_forward1_in_proj_bias_to_fp16, dilations = x_279_dilations_0, groups = x_279_groups_0, pad = x_279_pad_0, pad_type = x_279_pad_type_0, strides = x_279_strides_0, weight = core_stacks_2_stack_layers_2_feed_forward1_in_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("x_279_cast_fp16")];
            tensor<fp16, []> var_2324_to_fp16 = const()[name = tensor<string, []>("op_2324_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 256]> x_281_cast_fp16 = sub(x = x_279_cast_fp16, y = var_2324_to_fp16)[name = tensor<string, []>("x_281_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2326_cast_fp16 = relu(x = x_281_cast_fp16)[name = tensor<string, []>("op_2326_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2327_cast_fp16 = abs(x = x_281_cast_fp16)[name = tensor<string, []>("op_2327_cast_fp16")];
            tensor<fp16, []> const_50_promoted_to_fp16 = const()[name = tensor<string, []>("const_50_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 256]> var_2328_cast_fp16 = mul(x = var_2327_cast_fp16, y = const_50_promoted_to_fp16)[name = tensor<string, []>("op_2328_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2329_cast_fp16 = exp(x = var_2328_cast_fp16)[name = tensor<string, []>("op_2329_cast_fp16")];
            tensor<fp32, []> var_2330_epsilon_0 = const()[name = tensor<string, []>("op_2330_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 256]> var_2330_cast_fp16 = log(epsilon = var_2330_epsilon_0, x = var_2329_cast_fp16)[name = tensor<string, []>("op_2330_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2331_cast_fp16 = add(x = var_2326_cast_fp16, y = var_2330_cast_fp16)[name = tensor<string, []>("op_2331_cast_fp16")];
            tensor<fp16, []> var_2332_to_fp16 = const()[name = tensor<string, []>("op_2332_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 256]> var_2333_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2332_to_fp16)[name = tensor<string, []>("op_2333_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2334_cast_fp16 = sub(x = var_2331_cast_fp16, y = var_2333_cast_fp16)[name = tensor<string, []>("op_2334_cast_fp16")];
            tensor<fp16, []> var_2335_to_fp16 = const()[name = tensor<string, []>("op_2335_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 256]> input_205_cast_fp16 = sub(x = var_2334_cast_fp16, y = var_2335_to_fp16)[name = tensor<string, []>("input_205_cast_fp16")];
            tensor<string, []> var_2343_pad_type_0 = const()[name = tensor<string, []>("op_2343_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2343_strides_0 = const()[name = tensor<string, []>("op_2343_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2343_pad_0 = const()[name = tensor<string, []>("op_2343_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2343_dilations_0 = const()[name = tensor<string, []>("op_2343_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2343_groups_0 = const()[name = tensor<string, []>("op_2343_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_2_stack_layers_2_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179822848)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_2_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181002560)))];
            tensor<fp16, [1, 512, 1, 256]> var_2343_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_feed_forward1_out_proj_bias_to_fp16, dilations = var_2343_dilations_0, groups = var_2343_groups_0, pad = var_2343_pad_0, pad_type = var_2343_pad_type_0, strides = var_2343_strides_0, weight = core_stacks_2_stack_layers_2_feed_forward1_out_proj_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("op_2343_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_207_cast_fp16 = add(x = input_203_cast_fp16, y = var_2343_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
            tensor<int32, [4]> var_2345_begin_0 = const()[name = tensor<string, []>("op_2345_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2345_end_0 = const()[name = tensor<string, []>("op_2345_end_0"), val = tensor<int32, [4]>([1, 256, 1, 256])];
            tensor<bool, [4]> var_2345_end_mask_0 = const()[name = tensor<string, []>("op_2345_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 256, 1, 256]> var_2345_cast_fp16 = slice_by_index(begin = var_2345_begin_0, end = var_2345_end_0, end_mask = var_2345_end_mask_0, x = attn_weights_13_cast_fp16)[name = tensor<string, []>("op_2345_cast_fp16")];
            tensor<string, []> x_283_pad_type_0 = const()[name = tensor<string, []>("x_283_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_283_strides_0 = const()[name = tensor<string, []>("x_283_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_283_pad_0 = const()[name = tensor<string, []>("x_283_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_283_dilations_0 = const()[name = tensor<string, []>("x_283_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_283_groups_0 = const()[name = tensor<string, []>("x_283_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_2_stack_layers_2_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181003648)))];
            tensor<fp16, [1152]> core_stacks_2_stack_layers_2_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182183360)))];
            tensor<fp16, [1, 1152, 1, 256]> x_283_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_nonlin_attention_in_proj_bias_to_fp16, dilations = x_283_dilations_0, groups = x_283_groups_0, pad = x_283_pad_0, pad_type = x_283_pad_type_0, strides = x_283_strides_0, weight = core_stacks_2_stack_layers_2_nonlin_attention_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor<string, []>("x_283_cast_fp16")];
            tensor<int32, [4]> var_2356_begin_0 = const()[name = tensor<string, []>("op_2356_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2356_end_0 = const()[name = tensor<string, []>("op_2356_end_0"), val = tensor<int32, [4]>([1, 384, 1, 256])];
            tensor<bool, [4]> var_2356_end_mask_0 = const()[name = tensor<string, []>("op_2356_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 256]> var_2356_cast_fp16 = slice_by_index(begin = var_2356_begin_0, end = var_2356_end_0, end_mask = var_2356_end_mask_0, x = x_283_cast_fp16)[name = tensor<string, []>("op_2356_cast_fp16")];
            tensor<fp16, [1, 384, 1, 256]> s_gate_13_cast_fp16 = tanh(x = var_2356_cast_fp16)[name = tensor<string, []>("s_gate_13_cast_fp16")];
            tensor<int32, [4]> var_2359_begin_0 = const()[name = tensor<string, []>("op_2359_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_2359_end_0 = const()[name = tensor<string, []>("op_2359_end_0"), val = tensor<int32, [4]>([1, 768, 1, 256])];
            tensor<bool, [4]> var_2359_end_mask_0 = const()[name = tensor<string, []>("op_2359_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 256]> var_2359_cast_fp16 = slice_by_index(begin = var_2359_begin_0, end = var_2359_end_0, end_mask = var_2359_end_mask_0, x = x_283_cast_fp16)[name = tensor<string, []>("op_2359_cast_fp16")];
            tensor<fp16, [1, 384, 1, 256]> var_2360_cast_fp16 = mul(x = var_2359_cast_fp16, y = s_gate_13_cast_fp16)[name = tensor<string, []>("op_2360_cast_fp16")];
            tensor<int32, [4]> y_13_begin_0 = const()[name = tensor<string, []>("y_13_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_13_end_0 = const()[name = tensor<string, []>("y_13_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 256])];
            tensor<bool, [4]> y_13_end_mask_0 = const()[name = tensor<string, []>("y_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 256]> y_13_cast_fp16 = slice_by_index(begin = y_13_begin_0, end = y_13_end_0, end_mask = y_13_end_mask_0, x = x_283_cast_fp16)[name = tensor<string, []>("y_13_cast_fp16")];
            tensor<int32, [4]> transpose_50_perm_0 = const()[name = tensor<string, []>("transpose_50_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_254 = const()[name = tensor<string, []>("concat_254"), val = tensor<int32, [3]>([1, 256, 256])];
            tensor<fp16, [1, 1, 256, 256]> transpose_50_cast_fp16 = transpose(perm = transpose_50_perm_0, x = var_2345_cast_fp16)[name = tensor<string, []>("transpose_244")];
            tensor<fp16, [1, 256, 256]> reshape_75_cast_fp16 = reshape(shape = concat_254, x = transpose_50_cast_fp16)[name = tensor<string, []>("reshape_75_cast_fp16")];
            tensor<int32, [4]> transpose_51_perm_0 = const()[name = tensor<string, []>("transpose_51_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_255 = const()[name = tensor<string, []>("concat_255"), val = tensor<int32, [3]>([1, 256, 384])];
            tensor<fp16, [1, 1, 256, 384]> transpose_51_cast_fp16 = transpose(perm = transpose_51_perm_0, x = var_2360_cast_fp16)[name = tensor<string, []>("transpose_243")];
            tensor<fp16, [1, 256, 384]> reshape_76_cast_fp16 = reshape(shape = concat_255, x = transpose_51_cast_fp16)[name = tensor<string, []>("reshape_76_cast_fp16")];
            tensor<bool, []> matmul_25_transpose_x_0 = const()[name = tensor<string, []>("matmul_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_25_transpose_y_0 = const()[name = tensor<string, []>("matmul_25_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 256, 384]> matmul_25_cast_fp16 = matmul(transpose_x = matmul_25_transpose_x_0, transpose_y = matmul_25_transpose_y_0, x = reshape_75_cast_fp16, y = reshape_76_cast_fp16)[name = tensor<string, []>("matmul_25_cast_fp16")];
            tensor<int32, [4]> concat_259 = const()[name = tensor<string, []>("concat_259"), val = tensor<int32, [4]>([1, 1, 256, 384])];
            tensor<fp16, [1, 1, 256, 384]> reshape_77_cast_fp16 = reshape(shape = concat_259, x = matmul_25_cast_fp16)[name = tensor<string, []>("reshape_77_cast_fp16")];
            tensor<int32, [4]> v_13_perm_0 = const()[name = tensor<string, []>("v_13_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 256]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = reshape_77_cast_fp16)[name = tensor<string, []>("transpose_242")];
            tensor<fp16, [1, 384, 1, 256]> input_209_cast_fp16 = mul(x = v_13_cast_fp16, y = y_13_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
            tensor<string, []> var_2372_pad_type_0 = const()[name = tensor<string, []>("op_2372_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2372_strides_0 = const()[name = tensor<string, []>("op_2372_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2372_pad_0 = const()[name = tensor<string, []>("op_2372_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2372_dilations_0 = const()[name = tensor<string, []>("op_2372_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2372_groups_0 = const()[name = tensor<string, []>("op_2372_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_2_stack_layers_2_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182185728)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_2_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182579008)))];
            tensor<fp16, [1, 512, 1, 256]> var_2372_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_nonlin_attention_out_proj_bias_to_fp16, dilations = var_2372_dilations_0, groups = var_2372_groups_0, pad = var_2372_pad_0, pad_type = var_2372_pad_type_0, strides = var_2372_strides_0, weight = core_stacks_2_stack_layers_2_nonlin_attention_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor<string, []>("op_2372_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_285_cast_fp16 = add(x = input_207_cast_fp16, y = var_2372_cast_fp16)[name = tensor<string, []>("x_285_cast_fp16")];
            tensor<string, []> var_2383_pad_type_0 = const()[name = tensor<string, []>("op_2383_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2383_strides_0 = const()[name = tensor<string, []>("op_2383_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2383_pad_0 = const()[name = tensor<string, []>("op_2383_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2383_dilations_0 = const()[name = tensor<string, []>("op_2383_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2383_groups_0 = const()[name = tensor<string, []>("op_2383_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_2_stack_layers_2_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182580096)))];
            tensor<fp16, [48]> core_stacks_2_stack_layers_2_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182629312)))];
            tensor<fp16, [1, 48, 1, 256]> var_2383_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_self_attn1_in_proj_bias_to_fp16, dilations = var_2383_dilations_0, groups = var_2383_groups_0, pad = var_2383_pad_0, pad_type = var_2383_pad_type_0, strides = var_2383_strides_0, weight = core_stacks_2_stack_layers_2_self_attn1_in_proj_weight_to_fp16, x = x_285_cast_fp16)[name = tensor<string, []>("op_2383_cast_fp16")];
            tensor<int32, [4]> var_2384 = const()[name = tensor<string, []>("op_2384"), val = tensor<int32, [4]>([4, 12, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> var_2385_cast_fp16 = reshape(shape = var_2384, x = var_2383_cast_fp16)[name = tensor<string, []>("op_2385_cast_fp16")];
            tensor<int32, [4]> transpose_52_perm_0 = const()[name = tensor<string, []>("transpose_52_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_264 = const()[name = tensor<string, []>("concat_264"), val = tensor<int32, [3]>([4, 256, 256])];
            tensor<fp16, [4, 1, 256, 256]> transpose_52_cast_fp16 = transpose(perm = transpose_52_perm_0, x = attn_weights_13_cast_fp16)[name = tensor<string, []>("transpose_241")];
            tensor<fp16, [4, 256, 256]> reshape_78_cast_fp16 = reshape(shape = concat_264, x = transpose_52_cast_fp16)[name = tensor<string, []>("reshape_78_cast_fp16")];
            tensor<int32, [4]> transpose_53_perm_0 = const()[name = tensor<string, []>("transpose_53_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_265 = const()[name = tensor<string, []>("concat_265"), val = tensor<int32, [3]>([4, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = var_2385_cast_fp16)[name = tensor<string, []>("transpose_240")];
            tensor<fp16, [4, 256, 12]> reshape_79_cast_fp16 = reshape(shape = concat_265, x = transpose_53_cast_fp16)[name = tensor<string, []>("reshape_79_cast_fp16")];
            tensor<bool, []> matmul_26_transpose_x_0 = const()[name = tensor<string, []>("matmul_26_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_26_transpose_y_0 = const()[name = tensor<string, []>("matmul_26_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 12]> matmul_26_cast_fp16 = matmul(transpose_x = matmul_26_transpose_x_0, transpose_y = matmul_26_transpose_y_0, x = reshape_78_cast_fp16, y = reshape_79_cast_fp16)[name = tensor<string, []>("matmul_26_cast_fp16")];
            tensor<int32, [4]> concat_269 = const()[name = tensor<string, []>("concat_269"), val = tensor<int32, [4]>([4, 1, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> reshape_80_cast_fp16 = reshape(shape = concat_269, x = matmul_26_cast_fp16)[name = tensor<string, []>("reshape_80_cast_fp16")];
            tensor<int32, [4]> out_25_perm_0 = const()[name = tensor<string, []>("out_25_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_2388 = const()[name = tensor<string, []>("op_2388"), val = tensor<int32, [4]>([1, 48, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> out_25_cast_fp16 = transpose(perm = out_25_perm_0, x = reshape_80_cast_fp16)[name = tensor<string, []>("transpose_239")];
            tensor<fp16, [1, 48, 1, 256]> input_211_cast_fp16 = reshape(shape = var_2388, x = out_25_cast_fp16)[name = tensor<string, []>("input_211_cast_fp16")];
            tensor<string, []> var_2396_pad_type_0 = const()[name = tensor<string, []>("op_2396_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2396_strides_0 = const()[name = tensor<string, []>("op_2396_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2396_pad_0 = const()[name = tensor<string, []>("op_2396_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2396_dilations_0 = const()[name = tensor<string, []>("op_2396_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2396_groups_0 = const()[name = tensor<string, []>("op_2396_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_2_stack_layers_2_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182629504)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_2_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182678720)))];
            tensor<fp16, [1, 512, 1, 256]> var_2396_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_self_attn1_out_proj_bias_to_fp16, dilations = var_2396_dilations_0, groups = var_2396_groups_0, pad = var_2396_pad_0, pad_type = var_2396_pad_type_0, strides = var_2396_strides_0, weight = core_stacks_2_stack_layers_2_self_attn1_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("op_2396_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_287_cast_fp16 = add(x = x_285_cast_fp16, y = var_2396_cast_fp16)[name = tensor<string, []>("x_287_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_289_cast_fp16 = add(x = x_287_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("x_289_cast_fp16")];
            tensor<string, []> x_291_pad_type_0 = const()[name = tensor<string, []>("x_291_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_291_strides_0 = const()[name = tensor<string, []>("x_291_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_291_pad_0 = const()[name = tensor<string, []>("x_291_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_291_dilations_0 = const()[name = tensor<string, []>("x_291_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_291_groups_0 = const()[name = tensor<string, []>("x_291_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_2_stack_layers_2_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182679808)))];
            tensor<fp16, [1024]> core_stacks_2_stack_layers_2_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183728448)))];
            tensor<fp16, [1, 1024, 1, 256]> x_291_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_conv_module1_in_proj_bias_to_fp16, dilations = x_291_dilations_0, groups = x_291_groups_0, pad = x_291_pad_0, pad_type = x_291_pad_type_0, strides = x_291_strides_0, weight = core_stacks_2_stack_layers_2_conv_module1_in_proj_weight_to_fp16, x = x_289_cast_fp16)[name = tensor<string, []>("x_291_cast_fp16")];
            tensor<int32, [4]> var_2410_begin_0 = const()[name = tensor<string, []>("op_2410_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2410_end_0 = const()[name = tensor<string, []>("op_2410_end_0"), val = tensor<int32, [4]>([1, 512, 1, 256])];
            tensor<bool, [4]> var_2410_end_mask_0 = const()[name = tensor<string, []>("op_2410_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2410_cast_fp16 = slice_by_index(begin = var_2410_begin_0, end = var_2410_end_0, end_mask = var_2410_end_mask_0, x = x_291_cast_fp16)[name = tensor<string, []>("op_2410_cast_fp16")];
            tensor<int32, [4]> var_2412_begin_0 = const()[name = tensor<string, []>("op_2412_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_2412_end_0 = const()[name = tensor<string, []>("op_2412_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 256])];
            tensor<bool, [4]> var_2412_end_mask_0 = const()[name = tensor<string, []>("op_2412_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2412_cast_fp16 = slice_by_index(begin = var_2412_begin_0, end = var_2412_end_0, end_mask = var_2412_end_mask_0, x = x_291_cast_fp16)[name = tensor<string, []>("op_2412_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2413_cast_fp16 = sigmoid(x = var_2412_cast_fp16)[name = tensor<string, []>("op_2413_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_293_cast_fp16 = mul(x = var_2410_cast_fp16, y = var_2413_cast_fp16)[name = tensor<string, []>("x_293_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_213_cast_fp16 = mul(x = x_293_cast_fp16, y = conv_keep_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
            tensor<string, []> x_295_pad_type_0 = const()[name = tensor<string, []>("x_295_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_295_pad_0 = const()[name = tensor<string, []>("x_295_pad_0"), val = tensor<int32, [4]>([0, 0, 3, 3])];
            tensor<int32, []> x_295_groups_0 = const()[name = tensor<string, []>("x_295_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_295_strides_0 = const()[name = tensor<string, []>("x_295_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_295_dilations_0 = const()[name = tensor<string, []>("x_295_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 7]> core_stacks_2_stack_layers_2_conv_module1_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module1_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 7]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183730560)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_2_conv_module1_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module1_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183737792)))];
            tensor<fp16, [1, 512, 1, 256]> x_295_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_conv_module1_dw_bias_to_fp16, dilations = x_295_dilations_0, groups = x_295_groups_0, pad = x_295_pad_0, pad_type = x_295_pad_type_0, strides = x_295_strides_0, weight = core_stacks_2_stack_layers_2_conv_module1_dw_weight_to_fp16, x = input_213_cast_fp16)[name = tensor<string, []>("x_295_cast_fp16")];
            tensor<fp16, []> var_2423_to_fp16 = const()[name = tensor<string, []>("op_2423_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> x_297_cast_fp16 = sub(x = x_295_cast_fp16, y = var_2423_to_fp16)[name = tensor<string, []>("x_297_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2425_cast_fp16 = relu(x = x_297_cast_fp16)[name = tensor<string, []>("op_2425_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2426_cast_fp16 = abs(x = x_297_cast_fp16)[name = tensor<string, []>("op_2426_cast_fp16")];
            tensor<fp16, []> const_52_promoted_to_fp16 = const()[name = tensor<string, []>("const_52_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2427_cast_fp16 = mul(x = var_2426_cast_fp16, y = const_52_promoted_to_fp16)[name = tensor<string, []>("op_2427_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2428_cast_fp16 = exp(x = var_2427_cast_fp16)[name = tensor<string, []>("op_2428_cast_fp16")];
            tensor<fp32, []> var_2429_epsilon_0 = const()[name = tensor<string, []>("op_2429_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2429_cast_fp16 = log(epsilon = var_2429_epsilon_0, x = var_2428_cast_fp16)[name = tensor<string, []>("op_2429_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2430_cast_fp16 = add(x = var_2425_cast_fp16, y = var_2429_cast_fp16)[name = tensor<string, []>("op_2430_cast_fp16")];
            tensor<fp16, []> var_2431_to_fp16 = const()[name = tensor<string, []>("op_2431_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 256]> var_2432_cast_fp16 = mul(x = x_295_cast_fp16, y = var_2431_to_fp16)[name = tensor<string, []>("op_2432_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2433_cast_fp16 = sub(x = var_2430_cast_fp16, y = var_2432_cast_fp16)[name = tensor<string, []>("op_2433_cast_fp16")];
            tensor<fp16, []> var_2434_to_fp16 = const()[name = tensor<string, []>("op_2434_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 256]> input_215_cast_fp16 = sub(x = var_2433_cast_fp16, y = var_2434_to_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
            tensor<string, []> var_2442_pad_type_0 = const()[name = tensor<string, []>("op_2442_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2442_strides_0 = const()[name = tensor<string, []>("op_2442_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2442_pad_0 = const()[name = tensor<string, []>("op_2442_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2442_dilations_0 = const()[name = tensor<string, []>("op_2442_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2442_groups_0 = const()[name = tensor<string, []>("op_2442_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_2_stack_layers_2_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183738880)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_2_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184263232)))];
            tensor<fp16, [1, 512, 1, 256]> var_2442_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_conv_module1_out_proj_bias_to_fp16, dilations = var_2442_dilations_0, groups = var_2442_groups_0, pad = var_2442_pad_0, pad_type = var_2442_pad_type_0, strides = var_2442_strides_0, weight = core_stacks_2_stack_layers_2_conv_module1_out_proj_weight_to_fp16, x = input_215_cast_fp16)[name = tensor<string, []>("op_2442_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_217_cast_fp16 = add(x = x_289_cast_fp16, y = var_2442_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
            tensor<string, []> x_299_pad_type_0 = const()[name = tensor<string, []>("x_299_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_299_strides_0 = const()[name = tensor<string, []>("x_299_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_299_pad_0 = const()[name = tensor<string, []>("x_299_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_299_dilations_0 = const()[name = tensor<string, []>("x_299_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_299_groups_0 = const()[name = tensor<string, []>("x_299_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_2_stack_layers_2_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184264320)))];
            tensor<fp16, [1536]> core_stacks_2_stack_layers_2_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185837248)))];
            tensor<fp16, [1, 1536, 1, 256]> x_299_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_feed_forward2_in_proj_bias_to_fp16, dilations = x_299_dilations_0, groups = x_299_groups_0, pad = x_299_pad_0, pad_type = x_299_pad_type_0, strides = x_299_strides_0, weight = core_stacks_2_stack_layers_2_feed_forward2_in_proj_weight_to_fp16, x = input_217_cast_fp16)[name = tensor<string, []>("x_299_cast_fp16")];
            tensor<fp16, []> var_2453_to_fp16 = const()[name = tensor<string, []>("op_2453_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 256]> x_301_cast_fp16 = sub(x = x_299_cast_fp16, y = var_2453_to_fp16)[name = tensor<string, []>("x_301_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2455_cast_fp16 = relu(x = x_301_cast_fp16)[name = tensor<string, []>("op_2455_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2456_cast_fp16 = abs(x = x_301_cast_fp16)[name = tensor<string, []>("op_2456_cast_fp16")];
            tensor<fp16, []> const_53_promoted_to_fp16 = const()[name = tensor<string, []>("const_53_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 256]> var_2457_cast_fp16 = mul(x = var_2456_cast_fp16, y = const_53_promoted_to_fp16)[name = tensor<string, []>("op_2457_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2458_cast_fp16 = exp(x = var_2457_cast_fp16)[name = tensor<string, []>("op_2458_cast_fp16")];
            tensor<fp32, []> var_2459_epsilon_0 = const()[name = tensor<string, []>("op_2459_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 256]> var_2459_cast_fp16 = log(epsilon = var_2459_epsilon_0, x = var_2458_cast_fp16)[name = tensor<string, []>("op_2459_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2460_cast_fp16 = add(x = var_2455_cast_fp16, y = var_2459_cast_fp16)[name = tensor<string, []>("op_2460_cast_fp16")];
            tensor<fp16, []> var_2461_to_fp16 = const()[name = tensor<string, []>("op_2461_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 256]> var_2462_cast_fp16 = mul(x = x_299_cast_fp16, y = var_2461_to_fp16)[name = tensor<string, []>("op_2462_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2463_cast_fp16 = sub(x = var_2460_cast_fp16, y = var_2462_cast_fp16)[name = tensor<string, []>("op_2463_cast_fp16")];
            tensor<fp16, []> var_2464_to_fp16 = const()[name = tensor<string, []>("op_2464_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 256]> input_219_cast_fp16 = sub(x = var_2463_cast_fp16, y = var_2464_to_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
            tensor<string, []> var_2472_pad_type_0 = const()[name = tensor<string, []>("op_2472_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2472_strides_0 = const()[name = tensor<string, []>("op_2472_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2472_pad_0 = const()[name = tensor<string, []>("op_2472_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2472_dilations_0 = const()[name = tensor<string, []>("op_2472_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2472_groups_0 = const()[name = tensor<string, []>("op_2472_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_2_stack_layers_2_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185840384)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_2_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187413312)))];
            tensor<fp16, [1, 512, 1, 256]> var_2472_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_feed_forward2_out_proj_bias_to_fp16, dilations = var_2472_dilations_0, groups = var_2472_groups_0, pad = var_2472_pad_0, pad_type = var_2472_pad_type_0, strides = var_2472_strides_0, weight = core_stacks_2_stack_layers_2_feed_forward2_out_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("op_2472_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> src_27_cast_fp16 = add(x = input_217_cast_fp16, y = var_2472_cast_fp16)[name = tensor<string, []>("src_27_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2475_cast_fp16 = sub(x = src_27_cast_fp16, y = input_201_cast_fp16)[name = tensor<string, []>("op_2475_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_2_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187414400)))];
            tensor<fp16, [1, 512, 1, 256]> var_2476_cast_fp16 = mul(x = var_2475_cast_fp16, y = core_stacks_2_stack_layers_2_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_2476_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_303_cast_fp16 = add(x = input_201_cast_fp16, y = var_2476_cast_fp16)[name = tensor<string, []>("x_303_cast_fp16")];
            tensor<string, []> var_2487_pad_type_0 = const()[name = tensor<string, []>("op_2487_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2487_strides_0 = const()[name = tensor<string, []>("op_2487_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2487_pad_0 = const()[name = tensor<string, []>("op_2487_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2487_dilations_0 = const()[name = tensor<string, []>("op_2487_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2487_groups_0 = const()[name = tensor<string, []>("op_2487_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_2_stack_layers_2_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187415488)))];
            tensor<fp16, [48]> core_stacks_2_stack_layers_2_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187464704)))];
            tensor<fp16, [1, 48, 1, 256]> var_2487_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_self_attn2_in_proj_bias_to_fp16, dilations = var_2487_dilations_0, groups = var_2487_groups_0, pad = var_2487_pad_0, pad_type = var_2487_pad_type_0, strides = var_2487_strides_0, weight = core_stacks_2_stack_layers_2_self_attn2_in_proj_weight_to_fp16, x = x_303_cast_fp16)[name = tensor<string, []>("op_2487_cast_fp16")];
            tensor<int32, [4]> var_2488 = const()[name = tensor<string, []>("op_2488"), val = tensor<int32, [4]>([4, 12, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> var_2489_cast_fp16 = reshape(shape = var_2488, x = var_2487_cast_fp16)[name = tensor<string, []>("op_2489_cast_fp16")];
            tensor<int32, [4]> transpose_55_perm_0 = const()[name = tensor<string, []>("transpose_55_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_275 = const()[name = tensor<string, []>("concat_275"), val = tensor<int32, [3]>([4, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> transpose_55_cast_fp16 = transpose(perm = transpose_55_perm_0, x = var_2489_cast_fp16)[name = tensor<string, []>("transpose_238")];
            tensor<fp16, [4, 256, 12]> reshape_82_cast_fp16 = reshape(shape = concat_275, x = transpose_55_cast_fp16)[name = tensor<string, []>("reshape_82_cast_fp16")];
            tensor<bool, []> matmul_27_transpose_x_0 = const()[name = tensor<string, []>("matmul_27_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_27_transpose_y_0 = const()[name = tensor<string, []>("matmul_27_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 12]> matmul_27_cast_fp16 = matmul(transpose_x = matmul_27_transpose_x_0, transpose_y = matmul_27_transpose_y_0, x = reshape_78_cast_fp16, y = reshape_82_cast_fp16)[name = tensor<string, []>("matmul_27_cast_fp16")];
            tensor<int32, [4]> concat_279 = const()[name = tensor<string, []>("concat_279"), val = tensor<int32, [4]>([4, 1, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> reshape_83_cast_fp16 = reshape(shape = concat_279, x = matmul_27_cast_fp16)[name = tensor<string, []>("reshape_83_cast_fp16")];
            tensor<int32, [4]> out_27_perm_0 = const()[name = tensor<string, []>("out_27_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_2492 = const()[name = tensor<string, []>("op_2492"), val = tensor<int32, [4]>([1, 48, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> out_27_cast_fp16 = transpose(perm = out_27_perm_0, x = reshape_83_cast_fp16)[name = tensor<string, []>("transpose_237")];
            tensor<fp16, [1, 48, 1, 256]> input_221_cast_fp16 = reshape(shape = var_2492, x = out_27_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
            tensor<string, []> var_2500_pad_type_0 = const()[name = tensor<string, []>("op_2500_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2500_strides_0 = const()[name = tensor<string, []>("op_2500_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2500_pad_0 = const()[name = tensor<string, []>("op_2500_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2500_dilations_0 = const()[name = tensor<string, []>("op_2500_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2500_groups_0 = const()[name = tensor<string, []>("op_2500_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_2_stack_layers_2_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187464896)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_2_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187514112)))];
            tensor<fp16, [1, 512, 1, 256]> var_2500_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_self_attn2_out_proj_bias_to_fp16, dilations = var_2500_dilations_0, groups = var_2500_groups_0, pad = var_2500_pad_0, pad_type = var_2500_pad_type_0, strides = var_2500_strides_0, weight = core_stacks_2_stack_layers_2_self_attn2_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor<string, []>("op_2500_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_305_cast_fp16 = add(x = x_303_cast_fp16, y = var_2500_cast_fp16)[name = tensor<string, []>("x_305_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_307_cast_fp16 = add(x = x_305_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("x_307_cast_fp16")];
            tensor<string, []> x_309_pad_type_0 = const()[name = tensor<string, []>("x_309_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_309_strides_0 = const()[name = tensor<string, []>("x_309_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_309_pad_0 = const()[name = tensor<string, []>("x_309_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_309_dilations_0 = const()[name = tensor<string, []>("x_309_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_309_groups_0 = const()[name = tensor<string, []>("x_309_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_2_stack_layers_2_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187515200)))];
            tensor<fp16, [1024]> core_stacks_2_stack_layers_2_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188563840)))];
            tensor<fp16, [1, 1024, 1, 256]> x_309_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_conv_module2_in_proj_bias_to_fp16, dilations = x_309_dilations_0, groups = x_309_groups_0, pad = x_309_pad_0, pad_type = x_309_pad_type_0, strides = x_309_strides_0, weight = core_stacks_2_stack_layers_2_conv_module2_in_proj_weight_to_fp16, x = x_307_cast_fp16)[name = tensor<string, []>("x_309_cast_fp16")];
            tensor<int32, [4]> var_2514_begin_0 = const()[name = tensor<string, []>("op_2514_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2514_end_0 = const()[name = tensor<string, []>("op_2514_end_0"), val = tensor<int32, [4]>([1, 512, 1, 256])];
            tensor<bool, [4]> var_2514_end_mask_0 = const()[name = tensor<string, []>("op_2514_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2514_cast_fp16 = slice_by_index(begin = var_2514_begin_0, end = var_2514_end_0, end_mask = var_2514_end_mask_0, x = x_309_cast_fp16)[name = tensor<string, []>("op_2514_cast_fp16")];
            tensor<int32, [4]> var_2516_begin_0 = const()[name = tensor<string, []>("op_2516_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_2516_end_0 = const()[name = tensor<string, []>("op_2516_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 256])];
            tensor<bool, [4]> var_2516_end_mask_0 = const()[name = tensor<string, []>("op_2516_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2516_cast_fp16 = slice_by_index(begin = var_2516_begin_0, end = var_2516_end_0, end_mask = var_2516_end_mask_0, x = x_309_cast_fp16)[name = tensor<string, []>("op_2516_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2517_cast_fp16 = sigmoid(x = var_2516_cast_fp16)[name = tensor<string, []>("op_2517_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_311_cast_fp16 = mul(x = var_2514_cast_fp16, y = var_2517_cast_fp16)[name = tensor<string, []>("x_311_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_223_cast_fp16 = mul(x = x_311_cast_fp16, y = conv_keep_cast_fp16)[name = tensor<string, []>("input_223_cast_fp16")];
            tensor<string, []> x_313_pad_type_0 = const()[name = tensor<string, []>("x_313_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_313_pad_0 = const()[name = tensor<string, []>("x_313_pad_0"), val = tensor<int32, [4]>([0, 0, 3, 3])];
            tensor<int32, []> x_313_groups_0 = const()[name = tensor<string, []>("x_313_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_313_strides_0 = const()[name = tensor<string, []>("x_313_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_313_dilations_0 = const()[name = tensor<string, []>("x_313_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 7]> core_stacks_2_stack_layers_2_conv_module2_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module2_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 7]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188565952)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_2_conv_module2_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module2_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188573184)))];
            tensor<fp16, [1, 512, 1, 256]> x_313_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_conv_module2_dw_bias_to_fp16, dilations = x_313_dilations_0, groups = x_313_groups_0, pad = x_313_pad_0, pad_type = x_313_pad_type_0, strides = x_313_strides_0, weight = core_stacks_2_stack_layers_2_conv_module2_dw_weight_to_fp16, x = input_223_cast_fp16)[name = tensor<string, []>("x_313_cast_fp16")];
            tensor<fp16, []> var_2527_to_fp16 = const()[name = tensor<string, []>("op_2527_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> x_315_cast_fp16 = sub(x = x_313_cast_fp16, y = var_2527_to_fp16)[name = tensor<string, []>("x_315_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2529_cast_fp16 = relu(x = x_315_cast_fp16)[name = tensor<string, []>("op_2529_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2530_cast_fp16 = abs(x = x_315_cast_fp16)[name = tensor<string, []>("op_2530_cast_fp16")];
            tensor<fp16, []> const_55_promoted_to_fp16 = const()[name = tensor<string, []>("const_55_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2531_cast_fp16 = mul(x = var_2530_cast_fp16, y = const_55_promoted_to_fp16)[name = tensor<string, []>("op_2531_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2532_cast_fp16 = exp(x = var_2531_cast_fp16)[name = tensor<string, []>("op_2532_cast_fp16")];
            tensor<fp32, []> var_2533_epsilon_0 = const()[name = tensor<string, []>("op_2533_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2533_cast_fp16 = log(epsilon = var_2533_epsilon_0, x = var_2532_cast_fp16)[name = tensor<string, []>("op_2533_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2534_cast_fp16 = add(x = var_2529_cast_fp16, y = var_2533_cast_fp16)[name = tensor<string, []>("op_2534_cast_fp16")];
            tensor<fp16, []> var_2535_to_fp16 = const()[name = tensor<string, []>("op_2535_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 256]> var_2536_cast_fp16 = mul(x = x_313_cast_fp16, y = var_2535_to_fp16)[name = tensor<string, []>("op_2536_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2537_cast_fp16 = sub(x = var_2534_cast_fp16, y = var_2536_cast_fp16)[name = tensor<string, []>("op_2537_cast_fp16")];
            tensor<fp16, []> var_2538_to_fp16 = const()[name = tensor<string, []>("op_2538_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 256]> input_225_cast_fp16 = sub(x = var_2537_cast_fp16, y = var_2538_to_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
            tensor<string, []> var_2546_pad_type_0 = const()[name = tensor<string, []>("op_2546_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2546_strides_0 = const()[name = tensor<string, []>("op_2546_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2546_pad_0 = const()[name = tensor<string, []>("op_2546_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2546_dilations_0 = const()[name = tensor<string, []>("op_2546_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2546_groups_0 = const()[name = tensor<string, []>("op_2546_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_2_stack_layers_2_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188574272)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_2_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189098624)))];
            tensor<fp16, [1, 512, 1, 256]> var_2546_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_conv_module2_out_proj_bias_to_fp16, dilations = var_2546_dilations_0, groups = var_2546_groups_0, pad = var_2546_pad_0, pad_type = var_2546_pad_type_0, strides = var_2546_strides_0, weight = core_stacks_2_stack_layers_2_conv_module2_out_proj_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("op_2546_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_227_cast_fp16 = add(x = x_307_cast_fp16, y = var_2546_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
            tensor<string, []> x_317_pad_type_0 = const()[name = tensor<string, []>("x_317_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_317_strides_0 = const()[name = tensor<string, []>("x_317_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_317_pad_0 = const()[name = tensor<string, []>("x_317_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_317_dilations_0 = const()[name = tensor<string, []>("x_317_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_317_groups_0 = const()[name = tensor<string, []>("x_317_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_2_stack_layers_2_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189099712)))];
            tensor<fp16, [1920]> core_stacks_2_stack_layers_2_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191065856)))];
            tensor<fp16, [1, 1920, 1, 256]> x_317_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_feed_forward3_in_proj_bias_to_fp16, dilations = x_317_dilations_0, groups = x_317_groups_0, pad = x_317_pad_0, pad_type = x_317_pad_type_0, strides = x_317_strides_0, weight = core_stacks_2_stack_layers_2_feed_forward3_in_proj_weight_to_fp16, x = input_227_cast_fp16)[name = tensor<string, []>("x_317_cast_fp16")];
            tensor<fp16, []> var_2557_to_fp16 = const()[name = tensor<string, []>("op_2557_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 256]> x_319_cast_fp16 = sub(x = x_317_cast_fp16, y = var_2557_to_fp16)[name = tensor<string, []>("x_319_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2559_cast_fp16 = relu(x = x_319_cast_fp16)[name = tensor<string, []>("op_2559_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2560_cast_fp16 = abs(x = x_319_cast_fp16)[name = tensor<string, []>("op_2560_cast_fp16")];
            tensor<fp16, []> const_56_promoted_to_fp16 = const()[name = tensor<string, []>("const_56_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 256]> var_2561_cast_fp16 = mul(x = var_2560_cast_fp16, y = const_56_promoted_to_fp16)[name = tensor<string, []>("op_2561_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2562_cast_fp16 = exp(x = var_2561_cast_fp16)[name = tensor<string, []>("op_2562_cast_fp16")];
            tensor<fp32, []> var_2563_epsilon_0 = const()[name = tensor<string, []>("op_2563_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 256]> var_2563_cast_fp16 = log(epsilon = var_2563_epsilon_0, x = var_2562_cast_fp16)[name = tensor<string, []>("op_2563_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2564_cast_fp16 = add(x = var_2559_cast_fp16, y = var_2563_cast_fp16)[name = tensor<string, []>("op_2564_cast_fp16")];
            tensor<fp16, []> var_2565_to_fp16 = const()[name = tensor<string, []>("op_2565_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 256]> var_2566_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2565_to_fp16)[name = tensor<string, []>("op_2566_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2567_cast_fp16 = sub(x = var_2564_cast_fp16, y = var_2566_cast_fp16)[name = tensor<string, []>("op_2567_cast_fp16")];
            tensor<fp16, []> var_2568_to_fp16 = const()[name = tensor<string, []>("op_2568_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 256]> input_229_cast_fp16 = sub(x = var_2567_cast_fp16, y = var_2568_to_fp16)[name = tensor<string, []>("input_229_cast_fp16")];
            tensor<string, []> var_2576_pad_type_0 = const()[name = tensor<string, []>("op_2576_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2576_strides_0 = const()[name = tensor<string, []>("op_2576_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2576_pad_0 = const()[name = tensor<string, []>("op_2576_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2576_dilations_0 = const()[name = tensor<string, []>("op_2576_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2576_groups_0 = const()[name = tensor<string, []>("op_2576_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_2_stack_layers_2_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191069760)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_2_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193035904)))];
            tensor<fp16, [1, 512, 1, 256]> var_2576_cast_fp16 = conv(bias = core_stacks_2_stack_layers_2_feed_forward3_out_proj_bias_to_fp16, dilations = var_2576_dilations_0, groups = var_2576_groups_0, pad = var_2576_pad_0, pad_type = var_2576_pad_type_0, strides = var_2576_strides_0, weight = core_stacks_2_stack_layers_2_feed_forward3_out_proj_weight_to_fp16, x = input_229_cast_fp16)[name = tensor<string, []>("op_2576_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_321_cast_fp16 = add(x = input_227_cast_fp16, y = var_2576_cast_fp16)[name = tensor<string, []>("x_321_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_2_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193036992)))];
            tensor<fp16, [1, 512, 1, 256]> d_13_cast_fp16 = sub(x = x_321_cast_fp16, y = core_stacks_2_stack_layers_2_norm_bias_to_fp16)[name = tensor<string, []>("d_13_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2581_cast_fp16 = mul(x = d_13_cast_fp16, y = d_13_cast_fp16)[name = tensor<string, []>("op_2581_cast_fp16")];
            tensor<int32, [1]> var_2583_axes_0 = const()[name = tensor<string, []>("op_2583_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_2583_keep_dims_0 = const()[name = tensor<string, []>("op_2583_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 256]> var_2583_cast_fp16 = reduce_mean(axes = var_2583_axes_0, keep_dims = var_2583_keep_dims_0, x = var_2581_cast_fp16)[name = tensor<string, []>("op_2583_cast_fp16")];
            tensor<fp32, []> scales_13_epsilon_0 = const()[name = tensor<string, []>("scales_13_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 256]> scales_13_cast_fp16 = rsqrt(epsilon = scales_13_epsilon_0, x = var_2583_cast_fp16)[name = tensor<string, []>("scales_13_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2585_cast_fp16 = mul(x = x_321_cast_fp16, y = scales_13_cast_fp16)[name = tensor<string, []>("op_2585_cast_fp16")];
            tensor<fp16, []> core_stacks_2_stack_layers_2_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.e84p-3)];
            tensor<fp16, [1, 512, 1, 256]> src_29_cast_fp16 = mul(x = var_2585_cast_fp16, y = core_stacks_2_stack_layers_2_norm_scale_to_fp16)[name = tensor<string, []>("src_29_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2588_cast_fp16 = sub(x = src_29_cast_fp16, y = input_201_cast_fp16)[name = tensor<string, []>("op_2588_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_2_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_2_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193038080)))];
            tensor<fp16, [1, 512, 1, 256]> var_2589_cast_fp16 = mul(x = var_2588_cast_fp16, y = core_stacks_2_stack_layers_2_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_2589_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_231_cast_fp16 = add(x = input_201_cast_fp16, y = var_2589_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
            tensor<string, []> proj_15_pad_type_0 = const()[name = tensor<string, []>("proj_15_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_15_strides_0 = const()[name = tensor<string, []>("proj_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_15_pad_0 = const()[name = tensor<string, []>("proj_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_15_dilations_0 = const()[name = tensor<string, []>("proj_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_15_groups_0 = const()[name = tensor<string, []>("proj_15_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_2_stack_layers_3_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193039168)))];
            tensor<fp16, [384]> core_stacks_2_stack_layers_3_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193432448)))];
            tensor<fp16, [1, 384, 1, 256]> proj_15_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_15_dilations_0, groups = proj_15_groups_0, pad = proj_15_pad_0, pad_type = proj_15_pad_type_0, strides = proj_15_strides_0, weight = core_stacks_2_stack_layers_3_self_attn_weights_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("proj_15_cast_fp16")];
            tensor<int32, [4]> var_2612_begin_0 = const()[name = tensor<string, []>("op_2612_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2612_end_0 = const()[name = tensor<string, []>("op_2612_end_0"), val = tensor<int32, [4]>([1, 128, 1, 256])];
            tensor<bool, [4]> var_2612_end_mask_0 = const()[name = tensor<string, []>("op_2612_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_2612_cast_fp16 = slice_by_index(begin = var_2612_begin_0, end = var_2612_end_0, end_mask = var_2612_end_mask_0, x = proj_15_cast_fp16)[name = tensor<string, []>("op_2612_cast_fp16")];
            tensor<int32, [4]> var_2613 = const()[name = tensor<string, []>("op_2613"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> var_2614_cast_fp16 = reshape(shape = var_2613, x = var_2612_cast_fp16)[name = tensor<string, []>("op_2614_cast_fp16")];
            tensor<int32, [4]> var_2616_begin_0 = const()[name = tensor<string, []>("op_2616_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_2616_end_0 = const()[name = tensor<string, []>("op_2616_end_0"), val = tensor<int32, [4]>([1, 256, 1, 256])];
            tensor<bool, [4]> var_2616_end_mask_0 = const()[name = tensor<string, []>("op_2616_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_2616_cast_fp16 = slice_by_index(begin = var_2616_begin_0, end = var_2616_end_0, end_mask = var_2616_end_mask_0, x = proj_15_cast_fp16)[name = tensor<string, []>("op_2616_cast_fp16")];
            tensor<int32, [4]> var_2617 = const()[name = tensor<string, []>("op_2617"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> var_2618_cast_fp16 = reshape(shape = var_2617, x = var_2616_cast_fp16)[name = tensor<string, []>("op_2618_cast_fp16")];
            tensor<int32, [4]> var_2620_begin_0 = const()[name = tensor<string, []>("op_2620_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_2620_end_0 = const()[name = tensor<string, []>("op_2620_end_0"), val = tensor<int32, [4]>([1, 384, 1, 256])];
            tensor<bool, [4]> var_2620_end_mask_0 = const()[name = tensor<string, []>("op_2620_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 256]> var_2620_cast_fp16 = slice_by_index(begin = var_2620_begin_0, end = var_2620_end_0, end_mask = var_2620_end_mask_0, x = proj_15_cast_fp16)[name = tensor<string, []>("op_2620_cast_fp16")];
            tensor<int32, [4]> var_2621 = const()[name = tensor<string, []>("op_2621"), val = tensor<int32, [4]>([4, 32, 1, 256])];
            tensor<fp16, [4, 32, 1, 256]> pt_15_cast_fp16 = reshape(shape = var_2621, x = var_2620_cast_fp16)[name = tensor<string, []>("pt_15_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = tensor<string, []>("transpose_56_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_284 = const()[name = tensor<string, []>("concat_284"), val = tensor<int32, [3]>([4, 256, 32])];
            tensor<fp16, [4, 1, 256, 32]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = var_2614_cast_fp16)[name = tensor<string, []>("transpose_236")];
            tensor<fp16, [4, 256, 32]> reshape_84_cast_fp16 = reshape(shape = concat_284, x = transpose_56_cast_fp16)[name = tensor<string, []>("reshape_84_cast_fp16")];
            tensor<int32, [4]> transpose_57_perm_0 = const()[name = tensor<string, []>("transpose_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_285 = const()[name = tensor<string, []>("concat_285"), val = tensor<int32, [3]>([4, 32, 256])];
            tensor<fp16, [4, 1, 32, 256]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = var_2618_cast_fp16)[name = tensor<string, []>("transpose_235")];
            tensor<fp16, [4, 32, 256]> reshape_85_cast_fp16 = reshape(shape = concat_285, x = transpose_57_cast_fp16)[name = tensor<string, []>("reshape_85_cast_fp16")];
            tensor<bool, []> matmul_28_transpose_x_0 = const()[name = tensor<string, []>("matmul_28_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_28_transpose_y_0 = const()[name = tensor<string, []>("matmul_28_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 256]> matmul_28_cast_fp16 = matmul(transpose_x = matmul_28_transpose_x_0, transpose_y = matmul_28_transpose_y_0, x = reshape_84_cast_fp16, y = reshape_85_cast_fp16)[name = tensor<string, []>("matmul_28_cast_fp16")];
            tensor<int32, [4]> concat_289 = const()[name = tensor<string, []>("concat_289"), val = tensor<int32, [4]>([4, 1, 256, 256])];
            tensor<fp16, [4, 1, 256, 256]> reshape_86_cast_fp16 = reshape(shape = concat_289, x = matmul_28_cast_fp16)[name = tensor<string, []>("reshape_86_cast_fp16")];
            tensor<int32, [4]> attn_scores_43_perm_0 = const()[name = tensor<string, []>("attn_scores_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2625 = const()[name = tensor<string, []>("op_2625"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 256, 1]> var_2626_cast_fp16 = transpose(perm = var_2625, x = pt_15_cast_fp16)[name = tensor<string, []>("transpose_233")];
            tensor<fp16, [4, 32, 256, 256]> var_2627_cast_fp16 = mul(x = var_2626_cast_fp16, y = core_pos_basis_256_to_fp16)[name = tensor<string, []>("op_2627_cast_fp16")];
            tensor<int32, [1]> pos_scores_15_axes_0 = const()[name = tensor<string, []>("pos_scores_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_15_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_15_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 256, 256]> pos_scores_15_cast_fp16 = reduce_sum(axes = pos_scores_15_axes_0, keep_dims = pos_scores_15_keep_dims_0, x = var_2627_cast_fp16)[name = tensor<string, []>("pos_scores_15_cast_fp16")];
            tensor<int32, [4]> var_2630 = const()[name = tensor<string, []>("op_2630"), val = tensor<int32, [4]>([4, 256, 1, 256])];
            tensor<fp16, [4, 256, 1, 256]> var_2631_cast_fp16 = reshape(shape = var_2630, x = pos_scores_15_cast_fp16)[name = tensor<string, []>("op_2631_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_43_cast_fp16 = transpose(perm = attn_scores_43_perm_0, x = reshape_86_cast_fp16)[name = tensor<string, []>("transpose_234")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_45_cast_fp16 = add(x = attn_scores_43_cast_fp16, y = var_2631_cast_fp16)[name = tensor<string, []>("attn_scores_45_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_scores_47_cast_fp16 = add(x = attn_scores_45_cast_fp16, y = attn_bias_cast_fp16)[name = tensor<string, []>("attn_scores_47_cast_fp16")];
            tensor<fp16, [4, 256, 1, 256]> attn_weights_15_cast_fp16 = softmax(axis = var_41, x = attn_scores_47_cast_fp16)[name = tensor<string, []>("attn_weights_15_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_233_cast_fp16 = add(x = input_231_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
            tensor<string, []> x_323_pad_type_0 = const()[name = tensor<string, []>("x_323_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_323_strides_0 = const()[name = tensor<string, []>("x_323_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_323_pad_0 = const()[name = tensor<string, []>("x_323_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_323_dilations_0 = const()[name = tensor<string, []>("x_323_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_323_groups_0 = const()[name = tensor<string, []>("x_323_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_2_stack_layers_3_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(193433280)))];
            tensor<fp16, [1152]> core_stacks_2_stack_layers_3_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194612992)))];
            tensor<fp16, [1, 1152, 1, 256]> x_323_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_feed_forward1_in_proj_bias_to_fp16, dilations = x_323_dilations_0, groups = x_323_groups_0, pad = x_323_pad_0, pad_type = x_323_pad_type_0, strides = x_323_strides_0, weight = core_stacks_2_stack_layers_3_feed_forward1_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("x_323_cast_fp16")];
            tensor<fp16, []> var_2645_to_fp16 = const()[name = tensor<string, []>("op_2645_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 256]> x_325_cast_fp16 = sub(x = x_323_cast_fp16, y = var_2645_to_fp16)[name = tensor<string, []>("x_325_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2647_cast_fp16 = relu(x = x_325_cast_fp16)[name = tensor<string, []>("op_2647_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2648_cast_fp16 = abs(x = x_325_cast_fp16)[name = tensor<string, []>("op_2648_cast_fp16")];
            tensor<fp16, []> const_57_promoted_to_fp16 = const()[name = tensor<string, []>("const_57_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 256]> var_2649_cast_fp16 = mul(x = var_2648_cast_fp16, y = const_57_promoted_to_fp16)[name = tensor<string, []>("op_2649_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2650_cast_fp16 = exp(x = var_2649_cast_fp16)[name = tensor<string, []>("op_2650_cast_fp16")];
            tensor<fp32, []> var_2651_epsilon_0 = const()[name = tensor<string, []>("op_2651_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 256]> var_2651_cast_fp16 = log(epsilon = var_2651_epsilon_0, x = var_2650_cast_fp16)[name = tensor<string, []>("op_2651_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2652_cast_fp16 = add(x = var_2647_cast_fp16, y = var_2651_cast_fp16)[name = tensor<string, []>("op_2652_cast_fp16")];
            tensor<fp16, []> var_2653_to_fp16 = const()[name = tensor<string, []>("op_2653_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 256]> var_2654_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2653_to_fp16)[name = tensor<string, []>("op_2654_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 256]> var_2655_cast_fp16 = sub(x = var_2652_cast_fp16, y = var_2654_cast_fp16)[name = tensor<string, []>("op_2655_cast_fp16")];
            tensor<fp16, []> var_2656_to_fp16 = const()[name = tensor<string, []>("op_2656_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 256]> input_235_cast_fp16 = sub(x = var_2655_cast_fp16, y = var_2656_to_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
            tensor<string, []> var_2664_pad_type_0 = const()[name = tensor<string, []>("op_2664_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2664_strides_0 = const()[name = tensor<string, []>("op_2664_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2664_pad_0 = const()[name = tensor<string, []>("op_2664_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2664_dilations_0 = const()[name = tensor<string, []>("op_2664_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2664_groups_0 = const()[name = tensor<string, []>("op_2664_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_2_stack_layers_3_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194615360)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_3_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195795072)))];
            tensor<fp16, [1, 512, 1, 256]> var_2664_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_feed_forward1_out_proj_bias_to_fp16, dilations = var_2664_dilations_0, groups = var_2664_groups_0, pad = var_2664_pad_0, pad_type = var_2664_pad_type_0, strides = var_2664_strides_0, weight = core_stacks_2_stack_layers_3_feed_forward1_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("op_2664_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_237_cast_fp16 = add(x = input_233_cast_fp16, y = var_2664_cast_fp16)[name = tensor<string, []>("input_237_cast_fp16")];
            tensor<int32, [4]> var_2666_begin_0 = const()[name = tensor<string, []>("op_2666_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2666_end_0 = const()[name = tensor<string, []>("op_2666_end_0"), val = tensor<int32, [4]>([1, 256, 1, 256])];
            tensor<bool, [4]> var_2666_end_mask_0 = const()[name = tensor<string, []>("op_2666_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 256, 1, 256]> var_2666_cast_fp16 = slice_by_index(begin = var_2666_begin_0, end = var_2666_end_0, end_mask = var_2666_end_mask_0, x = attn_weights_15_cast_fp16)[name = tensor<string, []>("op_2666_cast_fp16")];
            tensor<string, []> x_327_pad_type_0 = const()[name = tensor<string, []>("x_327_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_327_strides_0 = const()[name = tensor<string, []>("x_327_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_327_pad_0 = const()[name = tensor<string, []>("x_327_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_327_dilations_0 = const()[name = tensor<string, []>("x_327_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_327_groups_0 = const()[name = tensor<string, []>("x_327_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_2_stack_layers_3_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195796160)))];
            tensor<fp16, [1152]> core_stacks_2_stack_layers_3_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196975872)))];
            tensor<fp16, [1, 1152, 1, 256]> x_327_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_nonlin_attention_in_proj_bias_to_fp16, dilations = x_327_dilations_0, groups = x_327_groups_0, pad = x_327_pad_0, pad_type = x_327_pad_type_0, strides = x_327_strides_0, weight = core_stacks_2_stack_layers_3_nonlin_attention_in_proj_weight_to_fp16, x = input_237_cast_fp16)[name = tensor<string, []>("x_327_cast_fp16")];
            tensor<int32, [4]> var_2677_begin_0 = const()[name = tensor<string, []>("op_2677_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2677_end_0 = const()[name = tensor<string, []>("op_2677_end_0"), val = tensor<int32, [4]>([1, 384, 1, 256])];
            tensor<bool, [4]> var_2677_end_mask_0 = const()[name = tensor<string, []>("op_2677_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 256]> var_2677_cast_fp16 = slice_by_index(begin = var_2677_begin_0, end = var_2677_end_0, end_mask = var_2677_end_mask_0, x = x_327_cast_fp16)[name = tensor<string, []>("op_2677_cast_fp16")];
            tensor<fp16, [1, 384, 1, 256]> s_gate_15_cast_fp16 = tanh(x = var_2677_cast_fp16)[name = tensor<string, []>("s_gate_15_cast_fp16")];
            tensor<int32, [4]> var_2680_begin_0 = const()[name = tensor<string, []>("op_2680_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_2680_end_0 = const()[name = tensor<string, []>("op_2680_end_0"), val = tensor<int32, [4]>([1, 768, 1, 256])];
            tensor<bool, [4]> var_2680_end_mask_0 = const()[name = tensor<string, []>("op_2680_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 256]> var_2680_cast_fp16 = slice_by_index(begin = var_2680_begin_0, end = var_2680_end_0, end_mask = var_2680_end_mask_0, x = x_327_cast_fp16)[name = tensor<string, []>("op_2680_cast_fp16")];
            tensor<fp16, [1, 384, 1, 256]> var_2681_cast_fp16 = mul(x = var_2680_cast_fp16, y = s_gate_15_cast_fp16)[name = tensor<string, []>("op_2681_cast_fp16")];
            tensor<int32, [4]> y_15_begin_0 = const()[name = tensor<string, []>("y_15_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_15_end_0 = const()[name = tensor<string, []>("y_15_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 256])];
            tensor<bool, [4]> y_15_end_mask_0 = const()[name = tensor<string, []>("y_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 256]> y_15_cast_fp16 = slice_by_index(begin = y_15_begin_0, end = y_15_end_0, end_mask = y_15_end_mask_0, x = x_327_cast_fp16)[name = tensor<string, []>("y_15_cast_fp16")];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = tensor<string, []>("transpose_58_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_294 = const()[name = tensor<string, []>("concat_294"), val = tensor<int32, [3]>([1, 256, 256])];
            tensor<fp16, [1, 1, 256, 256]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = var_2666_cast_fp16)[name = tensor<string, []>("transpose_232")];
            tensor<fp16, [1, 256, 256]> reshape_87_cast_fp16 = reshape(shape = concat_294, x = transpose_58_cast_fp16)[name = tensor<string, []>("reshape_87_cast_fp16")];
            tensor<int32, [4]> transpose_59_perm_0 = const()[name = tensor<string, []>("transpose_59_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_295 = const()[name = tensor<string, []>("concat_295"), val = tensor<int32, [3]>([1, 256, 384])];
            tensor<fp16, [1, 1, 256, 384]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = var_2681_cast_fp16)[name = tensor<string, []>("transpose_231")];
            tensor<fp16, [1, 256, 384]> reshape_88_cast_fp16 = reshape(shape = concat_295, x = transpose_59_cast_fp16)[name = tensor<string, []>("reshape_88_cast_fp16")];
            tensor<bool, []> matmul_29_transpose_x_0 = const()[name = tensor<string, []>("matmul_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_29_transpose_y_0 = const()[name = tensor<string, []>("matmul_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 256, 384]> matmul_29_cast_fp16 = matmul(transpose_x = matmul_29_transpose_x_0, transpose_y = matmul_29_transpose_y_0, x = reshape_87_cast_fp16, y = reshape_88_cast_fp16)[name = tensor<string, []>("matmul_29_cast_fp16")];
            tensor<int32, [4]> concat_299 = const()[name = tensor<string, []>("concat_299"), val = tensor<int32, [4]>([1, 1, 256, 384])];
            tensor<fp16, [1, 1, 256, 384]> reshape_89_cast_fp16 = reshape(shape = concat_299, x = matmul_29_cast_fp16)[name = tensor<string, []>("reshape_89_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = tensor<string, []>("v_15_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 256]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = reshape_89_cast_fp16)[name = tensor<string, []>("transpose_230")];
            tensor<fp16, [1, 384, 1, 256]> input_239_cast_fp16 = mul(x = v_15_cast_fp16, y = y_15_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
            tensor<string, []> var_2693_pad_type_0 = const()[name = tensor<string, []>("op_2693_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2693_strides_0 = const()[name = tensor<string, []>("op_2693_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2693_pad_0 = const()[name = tensor<string, []>("op_2693_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2693_dilations_0 = const()[name = tensor<string, []>("op_2693_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2693_groups_0 = const()[name = tensor<string, []>("op_2693_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_2_stack_layers_3_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(196978240)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_3_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197371520)))];
            tensor<fp16, [1, 512, 1, 256]> var_2693_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_nonlin_attention_out_proj_bias_to_fp16, dilations = var_2693_dilations_0, groups = var_2693_groups_0, pad = var_2693_pad_0, pad_type = var_2693_pad_type_0, strides = var_2693_strides_0, weight = core_stacks_2_stack_layers_3_nonlin_attention_out_proj_weight_to_fp16, x = input_239_cast_fp16)[name = tensor<string, []>("op_2693_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_329_cast_fp16 = add(x = input_237_cast_fp16, y = var_2693_cast_fp16)[name = tensor<string, []>("x_329_cast_fp16")];
            tensor<string, []> var_2704_pad_type_0 = const()[name = tensor<string, []>("op_2704_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2704_strides_0 = const()[name = tensor<string, []>("op_2704_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2704_pad_0 = const()[name = tensor<string, []>("op_2704_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2704_dilations_0 = const()[name = tensor<string, []>("op_2704_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2704_groups_0 = const()[name = tensor<string, []>("op_2704_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_2_stack_layers_3_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197372608)))];
            tensor<fp16, [48]> core_stacks_2_stack_layers_3_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197421824)))];
            tensor<fp16, [1, 48, 1, 256]> var_2704_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_self_attn1_in_proj_bias_to_fp16, dilations = var_2704_dilations_0, groups = var_2704_groups_0, pad = var_2704_pad_0, pad_type = var_2704_pad_type_0, strides = var_2704_strides_0, weight = core_stacks_2_stack_layers_3_self_attn1_in_proj_weight_to_fp16, x = x_329_cast_fp16)[name = tensor<string, []>("op_2704_cast_fp16")];
            tensor<int32, [4]> var_2705 = const()[name = tensor<string, []>("op_2705"), val = tensor<int32, [4]>([4, 12, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> var_2706_cast_fp16 = reshape(shape = var_2705, x = var_2704_cast_fp16)[name = tensor<string, []>("op_2706_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = tensor<string, []>("transpose_60_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_304 = const()[name = tensor<string, []>("concat_304"), val = tensor<int32, [3]>([4, 256, 256])];
            tensor<fp16, [4, 1, 256, 256]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = attn_weights_15_cast_fp16)[name = tensor<string, []>("transpose_229")];
            tensor<fp16, [4, 256, 256]> reshape_90_cast_fp16 = reshape(shape = concat_304, x = transpose_60_cast_fp16)[name = tensor<string, []>("reshape_90_cast_fp16")];
            tensor<int32, [4]> transpose_61_perm_0 = const()[name = tensor<string, []>("transpose_61_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_305 = const()[name = tensor<string, []>("concat_305"), val = tensor<int32, [3]>([4, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = var_2706_cast_fp16)[name = tensor<string, []>("transpose_228")];
            tensor<fp16, [4, 256, 12]> reshape_91_cast_fp16 = reshape(shape = concat_305, x = transpose_61_cast_fp16)[name = tensor<string, []>("reshape_91_cast_fp16")];
            tensor<bool, []> matmul_30_transpose_x_0 = const()[name = tensor<string, []>("matmul_30_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_30_transpose_y_0 = const()[name = tensor<string, []>("matmul_30_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 12]> matmul_30_cast_fp16 = matmul(transpose_x = matmul_30_transpose_x_0, transpose_y = matmul_30_transpose_y_0, x = reshape_90_cast_fp16, y = reshape_91_cast_fp16)[name = tensor<string, []>("matmul_30_cast_fp16")];
            tensor<int32, [4]> concat_309 = const()[name = tensor<string, []>("concat_309"), val = tensor<int32, [4]>([4, 1, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> reshape_92_cast_fp16 = reshape(shape = concat_309, x = matmul_30_cast_fp16)[name = tensor<string, []>("reshape_92_cast_fp16")];
            tensor<int32, [4]> out_29_perm_0 = const()[name = tensor<string, []>("out_29_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_2709 = const()[name = tensor<string, []>("op_2709"), val = tensor<int32, [4]>([1, 48, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> out_29_cast_fp16 = transpose(perm = out_29_perm_0, x = reshape_92_cast_fp16)[name = tensor<string, []>("transpose_227")];
            tensor<fp16, [1, 48, 1, 256]> input_241_cast_fp16 = reshape(shape = var_2709, x = out_29_cast_fp16)[name = tensor<string, []>("input_241_cast_fp16")];
            tensor<string, []> var_2717_pad_type_0 = const()[name = tensor<string, []>("op_2717_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2717_strides_0 = const()[name = tensor<string, []>("op_2717_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2717_pad_0 = const()[name = tensor<string, []>("op_2717_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2717_dilations_0 = const()[name = tensor<string, []>("op_2717_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2717_groups_0 = const()[name = tensor<string, []>("op_2717_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_2_stack_layers_3_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197422016)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_3_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197471232)))];
            tensor<fp16, [1, 512, 1, 256]> var_2717_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_self_attn1_out_proj_bias_to_fp16, dilations = var_2717_dilations_0, groups = var_2717_groups_0, pad = var_2717_pad_0, pad_type = var_2717_pad_type_0, strides = var_2717_strides_0, weight = core_stacks_2_stack_layers_3_self_attn1_out_proj_weight_to_fp16, x = input_241_cast_fp16)[name = tensor<string, []>("op_2717_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_331_cast_fp16 = add(x = x_329_cast_fp16, y = var_2717_cast_fp16)[name = tensor<string, []>("x_331_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_333_cast_fp16 = add(x = x_331_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("x_333_cast_fp16")];
            tensor<string, []> x_335_pad_type_0 = const()[name = tensor<string, []>("x_335_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_335_strides_0 = const()[name = tensor<string, []>("x_335_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_335_pad_0 = const()[name = tensor<string, []>("x_335_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_335_dilations_0 = const()[name = tensor<string, []>("x_335_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_335_groups_0 = const()[name = tensor<string, []>("x_335_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_2_stack_layers_3_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197472320)))];
            tensor<fp16, [1024]> core_stacks_2_stack_layers_3_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198520960)))];
            tensor<fp16, [1, 1024, 1, 256]> x_335_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_conv_module1_in_proj_bias_to_fp16, dilations = x_335_dilations_0, groups = x_335_groups_0, pad = x_335_pad_0, pad_type = x_335_pad_type_0, strides = x_335_strides_0, weight = core_stacks_2_stack_layers_3_conv_module1_in_proj_weight_to_fp16, x = x_333_cast_fp16)[name = tensor<string, []>("x_335_cast_fp16")];
            tensor<int32, [4]> var_2731_begin_0 = const()[name = tensor<string, []>("op_2731_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2731_end_0 = const()[name = tensor<string, []>("op_2731_end_0"), val = tensor<int32, [4]>([1, 512, 1, 256])];
            tensor<bool, [4]> var_2731_end_mask_0 = const()[name = tensor<string, []>("op_2731_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2731_cast_fp16 = slice_by_index(begin = var_2731_begin_0, end = var_2731_end_0, end_mask = var_2731_end_mask_0, x = x_335_cast_fp16)[name = tensor<string, []>("op_2731_cast_fp16")];
            tensor<int32, [4]> var_2733_begin_0 = const()[name = tensor<string, []>("op_2733_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_2733_end_0 = const()[name = tensor<string, []>("op_2733_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 256])];
            tensor<bool, [4]> var_2733_end_mask_0 = const()[name = tensor<string, []>("op_2733_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2733_cast_fp16 = slice_by_index(begin = var_2733_begin_0, end = var_2733_end_0, end_mask = var_2733_end_mask_0, x = x_335_cast_fp16)[name = tensor<string, []>("op_2733_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2734_cast_fp16 = sigmoid(x = var_2733_cast_fp16)[name = tensor<string, []>("op_2734_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_337_cast_fp16 = mul(x = var_2731_cast_fp16, y = var_2734_cast_fp16)[name = tensor<string, []>("x_337_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_243_cast_fp16 = mul(x = x_337_cast_fp16, y = conv_keep_cast_fp16)[name = tensor<string, []>("input_243_cast_fp16")];
            tensor<string, []> x_339_pad_type_0 = const()[name = tensor<string, []>("x_339_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_339_pad_0 = const()[name = tensor<string, []>("x_339_pad_0"), val = tensor<int32, [4]>([0, 0, 3, 3])];
            tensor<int32, []> x_339_groups_0 = const()[name = tensor<string, []>("x_339_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_339_strides_0 = const()[name = tensor<string, []>("x_339_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_339_dilations_0 = const()[name = tensor<string, []>("x_339_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 7]> core_stacks_2_stack_layers_3_conv_module1_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module1_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 7]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198523072)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_3_conv_module1_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module1_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198530304)))];
            tensor<fp16, [1, 512, 1, 256]> x_339_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_conv_module1_dw_bias_to_fp16, dilations = x_339_dilations_0, groups = x_339_groups_0, pad = x_339_pad_0, pad_type = x_339_pad_type_0, strides = x_339_strides_0, weight = core_stacks_2_stack_layers_3_conv_module1_dw_weight_to_fp16, x = input_243_cast_fp16)[name = tensor<string, []>("x_339_cast_fp16")];
            tensor<fp16, []> var_2744_to_fp16 = const()[name = tensor<string, []>("op_2744_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> x_341_cast_fp16 = sub(x = x_339_cast_fp16, y = var_2744_to_fp16)[name = tensor<string, []>("x_341_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2746_cast_fp16 = relu(x = x_341_cast_fp16)[name = tensor<string, []>("op_2746_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2747_cast_fp16 = abs(x = x_341_cast_fp16)[name = tensor<string, []>("op_2747_cast_fp16")];
            tensor<fp16, []> const_59_promoted_to_fp16 = const()[name = tensor<string, []>("const_59_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2748_cast_fp16 = mul(x = var_2747_cast_fp16, y = const_59_promoted_to_fp16)[name = tensor<string, []>("op_2748_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2749_cast_fp16 = exp(x = var_2748_cast_fp16)[name = tensor<string, []>("op_2749_cast_fp16")];
            tensor<fp32, []> var_2750_epsilon_0 = const()[name = tensor<string, []>("op_2750_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2750_cast_fp16 = log(epsilon = var_2750_epsilon_0, x = var_2749_cast_fp16)[name = tensor<string, []>("op_2750_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2751_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2750_cast_fp16)[name = tensor<string, []>("op_2751_cast_fp16")];
            tensor<fp16, []> var_2752_to_fp16 = const()[name = tensor<string, []>("op_2752_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 256]> var_2753_cast_fp16 = mul(x = x_339_cast_fp16, y = var_2752_to_fp16)[name = tensor<string, []>("op_2753_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2754_cast_fp16 = sub(x = var_2751_cast_fp16, y = var_2753_cast_fp16)[name = tensor<string, []>("op_2754_cast_fp16")];
            tensor<fp16, []> var_2755_to_fp16 = const()[name = tensor<string, []>("op_2755_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 256]> input_245_cast_fp16 = sub(x = var_2754_cast_fp16, y = var_2755_to_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
            tensor<string, []> var_2763_pad_type_0 = const()[name = tensor<string, []>("op_2763_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2763_strides_0 = const()[name = tensor<string, []>("op_2763_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2763_pad_0 = const()[name = tensor<string, []>("op_2763_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2763_dilations_0 = const()[name = tensor<string, []>("op_2763_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2763_groups_0 = const()[name = tensor<string, []>("op_2763_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_2_stack_layers_3_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198531392)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_3_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199055744)))];
            tensor<fp16, [1, 512, 1, 256]> var_2763_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_conv_module1_out_proj_bias_to_fp16, dilations = var_2763_dilations_0, groups = var_2763_groups_0, pad = var_2763_pad_0, pad_type = var_2763_pad_type_0, strides = var_2763_strides_0, weight = core_stacks_2_stack_layers_3_conv_module1_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("op_2763_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_247_cast_fp16 = add(x = x_333_cast_fp16, y = var_2763_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
            tensor<string, []> x_343_pad_type_0 = const()[name = tensor<string, []>("x_343_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_343_strides_0 = const()[name = tensor<string, []>("x_343_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_343_pad_0 = const()[name = tensor<string, []>("x_343_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_343_dilations_0 = const()[name = tensor<string, []>("x_343_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_343_groups_0 = const()[name = tensor<string, []>("x_343_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_2_stack_layers_3_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199056832)))];
            tensor<fp16, [1536]> core_stacks_2_stack_layers_3_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200629760)))];
            tensor<fp16, [1, 1536, 1, 256]> x_343_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_feed_forward2_in_proj_bias_to_fp16, dilations = x_343_dilations_0, groups = x_343_groups_0, pad = x_343_pad_0, pad_type = x_343_pad_type_0, strides = x_343_strides_0, weight = core_stacks_2_stack_layers_3_feed_forward2_in_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("x_343_cast_fp16")];
            tensor<fp16, []> var_2774_to_fp16 = const()[name = tensor<string, []>("op_2774_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 256]> x_345_cast_fp16 = sub(x = x_343_cast_fp16, y = var_2774_to_fp16)[name = tensor<string, []>("x_345_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2776_cast_fp16 = relu(x = x_345_cast_fp16)[name = tensor<string, []>("op_2776_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2777_cast_fp16 = abs(x = x_345_cast_fp16)[name = tensor<string, []>("op_2777_cast_fp16")];
            tensor<fp16, []> const_60_promoted_to_fp16 = const()[name = tensor<string, []>("const_60_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 256]> var_2778_cast_fp16 = mul(x = var_2777_cast_fp16, y = const_60_promoted_to_fp16)[name = tensor<string, []>("op_2778_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2779_cast_fp16 = exp(x = var_2778_cast_fp16)[name = tensor<string, []>("op_2779_cast_fp16")];
            tensor<fp32, []> var_2780_epsilon_0 = const()[name = tensor<string, []>("op_2780_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 256]> var_2780_cast_fp16 = log(epsilon = var_2780_epsilon_0, x = var_2779_cast_fp16)[name = tensor<string, []>("op_2780_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2781_cast_fp16 = add(x = var_2776_cast_fp16, y = var_2780_cast_fp16)[name = tensor<string, []>("op_2781_cast_fp16")];
            tensor<fp16, []> var_2782_to_fp16 = const()[name = tensor<string, []>("op_2782_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 256]> var_2783_cast_fp16 = mul(x = x_343_cast_fp16, y = var_2782_to_fp16)[name = tensor<string, []>("op_2783_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 256]> var_2784_cast_fp16 = sub(x = var_2781_cast_fp16, y = var_2783_cast_fp16)[name = tensor<string, []>("op_2784_cast_fp16")];
            tensor<fp16, []> var_2785_to_fp16 = const()[name = tensor<string, []>("op_2785_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 256]> input_249_cast_fp16 = sub(x = var_2784_cast_fp16, y = var_2785_to_fp16)[name = tensor<string, []>("input_249_cast_fp16")];
            tensor<string, []> var_2793_pad_type_0 = const()[name = tensor<string, []>("op_2793_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2793_strides_0 = const()[name = tensor<string, []>("op_2793_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2793_pad_0 = const()[name = tensor<string, []>("op_2793_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2793_dilations_0 = const()[name = tensor<string, []>("op_2793_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2793_groups_0 = const()[name = tensor<string, []>("op_2793_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_2_stack_layers_3_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200632896)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_3_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202205824)))];
            tensor<fp16, [1, 512, 1, 256]> var_2793_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_feed_forward2_out_proj_bias_to_fp16, dilations = var_2793_dilations_0, groups = var_2793_groups_0, pad = var_2793_pad_0, pad_type = var_2793_pad_type_0, strides = var_2793_strides_0, weight = core_stacks_2_stack_layers_3_feed_forward2_out_proj_weight_to_fp16, x = input_249_cast_fp16)[name = tensor<string, []>("op_2793_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> src_31_cast_fp16 = add(x = input_247_cast_fp16, y = var_2793_cast_fp16)[name = tensor<string, []>("src_31_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2796_cast_fp16 = sub(x = src_31_cast_fp16, y = input_231_cast_fp16)[name = tensor<string, []>("op_2796_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_3_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202206912)))];
            tensor<fp16, [1, 512, 1, 256]> var_2797_cast_fp16 = mul(x = var_2796_cast_fp16, y = core_stacks_2_stack_layers_3_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_2797_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_347_cast_fp16 = add(x = input_231_cast_fp16, y = var_2797_cast_fp16)[name = tensor<string, []>("x_347_cast_fp16")];
            tensor<string, []> var_2808_pad_type_0 = const()[name = tensor<string, []>("op_2808_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2808_strides_0 = const()[name = tensor<string, []>("op_2808_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2808_pad_0 = const()[name = tensor<string, []>("op_2808_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2808_dilations_0 = const()[name = tensor<string, []>("op_2808_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2808_groups_0 = const()[name = tensor<string, []>("op_2808_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_2_stack_layers_3_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202208000)))];
            tensor<fp16, [48]> core_stacks_2_stack_layers_3_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202257216)))];
            tensor<fp16, [1, 48, 1, 256]> var_2808_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_self_attn2_in_proj_bias_to_fp16, dilations = var_2808_dilations_0, groups = var_2808_groups_0, pad = var_2808_pad_0, pad_type = var_2808_pad_type_0, strides = var_2808_strides_0, weight = core_stacks_2_stack_layers_3_self_attn2_in_proj_weight_to_fp16, x = x_347_cast_fp16)[name = tensor<string, []>("op_2808_cast_fp16")];
            tensor<int32, [4]> var_2809 = const()[name = tensor<string, []>("op_2809"), val = tensor<int32, [4]>([4, 12, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> var_2810_cast_fp16 = reshape(shape = var_2809, x = var_2808_cast_fp16)[name = tensor<string, []>("op_2810_cast_fp16")];
            tensor<int32, [4]> transpose_63_perm_0 = const()[name = tensor<string, []>("transpose_63_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_315 = const()[name = tensor<string, []>("concat_315"), val = tensor<int32, [3]>([4, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = var_2810_cast_fp16)[name = tensor<string, []>("transpose_226")];
            tensor<fp16, [4, 256, 12]> reshape_94_cast_fp16 = reshape(shape = concat_315, x = transpose_63_cast_fp16)[name = tensor<string, []>("reshape_94_cast_fp16")];
            tensor<bool, []> matmul_31_transpose_x_0 = const()[name = tensor<string, []>("matmul_31_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_31_transpose_y_0 = const()[name = tensor<string, []>("matmul_31_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 256, 12]> matmul_31_cast_fp16 = matmul(transpose_x = matmul_31_transpose_x_0, transpose_y = matmul_31_transpose_y_0, x = reshape_90_cast_fp16, y = reshape_94_cast_fp16)[name = tensor<string, []>("matmul_31_cast_fp16")];
            tensor<int32, [4]> concat_319 = const()[name = tensor<string, []>("concat_319"), val = tensor<int32, [4]>([4, 1, 256, 12])];
            tensor<fp16, [4, 1, 256, 12]> reshape_95_cast_fp16 = reshape(shape = concat_319, x = matmul_31_cast_fp16)[name = tensor<string, []>("reshape_95_cast_fp16")];
            tensor<int32, [4]> out_31_perm_0 = const()[name = tensor<string, []>("out_31_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_2813 = const()[name = tensor<string, []>("op_2813"), val = tensor<int32, [4]>([1, 48, 1, 256])];
            tensor<fp16, [4, 12, 1, 256]> out_31_cast_fp16 = transpose(perm = out_31_perm_0, x = reshape_95_cast_fp16)[name = tensor<string, []>("transpose_225")];
            tensor<fp16, [1, 48, 1, 256]> input_251_cast_fp16 = reshape(shape = var_2813, x = out_31_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
            tensor<string, []> var_2821_pad_type_0 = const()[name = tensor<string, []>("op_2821_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2821_strides_0 = const()[name = tensor<string, []>("op_2821_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2821_pad_0 = const()[name = tensor<string, []>("op_2821_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2821_dilations_0 = const()[name = tensor<string, []>("op_2821_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2821_groups_0 = const()[name = tensor<string, []>("op_2821_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_2_stack_layers_3_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202257408)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_3_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202306624)))];
            tensor<fp16, [1, 512, 1, 256]> var_2821_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_self_attn2_out_proj_bias_to_fp16, dilations = var_2821_dilations_0, groups = var_2821_groups_0, pad = var_2821_pad_0, pad_type = var_2821_pad_type_0, strides = var_2821_strides_0, weight = core_stacks_2_stack_layers_3_self_attn2_out_proj_weight_to_fp16, x = input_251_cast_fp16)[name = tensor<string, []>("op_2821_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_349_cast_fp16 = add(x = x_347_cast_fp16, y = var_2821_cast_fp16)[name = tensor<string, []>("x_349_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_351_cast_fp16 = add(x = x_349_cast_fp16, y = time_emb_5_cast_fp16)[name = tensor<string, []>("x_351_cast_fp16")];
            tensor<string, []> x_353_pad_type_0 = const()[name = tensor<string, []>("x_353_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_353_strides_0 = const()[name = tensor<string, []>("x_353_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_353_pad_0 = const()[name = tensor<string, []>("x_353_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_353_dilations_0 = const()[name = tensor<string, []>("x_353_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_353_groups_0 = const()[name = tensor<string, []>("x_353_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_2_stack_layers_3_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(202307712)))];
            tensor<fp16, [1024]> core_stacks_2_stack_layers_3_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203356352)))];
            tensor<fp16, [1, 1024, 1, 256]> x_353_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_conv_module2_in_proj_bias_to_fp16, dilations = x_353_dilations_0, groups = x_353_groups_0, pad = x_353_pad_0, pad_type = x_353_pad_type_0, strides = x_353_strides_0, weight = core_stacks_2_stack_layers_3_conv_module2_in_proj_weight_to_fp16, x = x_351_cast_fp16)[name = tensor<string, []>("x_353_cast_fp16")];
            tensor<int32, [4]> var_2835_begin_0 = const()[name = tensor<string, []>("op_2835_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2835_end_0 = const()[name = tensor<string, []>("op_2835_end_0"), val = tensor<int32, [4]>([1, 512, 1, 256])];
            tensor<bool, [4]> var_2835_end_mask_0 = const()[name = tensor<string, []>("op_2835_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2835_cast_fp16 = slice_by_index(begin = var_2835_begin_0, end = var_2835_end_0, end_mask = var_2835_end_mask_0, x = x_353_cast_fp16)[name = tensor<string, []>("op_2835_cast_fp16")];
            tensor<int32, [4]> var_2837_begin_0 = const()[name = tensor<string, []>("op_2837_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_2837_end_0 = const()[name = tensor<string, []>("op_2837_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 256])];
            tensor<bool, [4]> var_2837_end_mask_0 = const()[name = tensor<string, []>("op_2837_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 256]> var_2837_cast_fp16 = slice_by_index(begin = var_2837_begin_0, end = var_2837_end_0, end_mask = var_2837_end_mask_0, x = x_353_cast_fp16)[name = tensor<string, []>("op_2837_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2838_cast_fp16 = sigmoid(x = var_2837_cast_fp16)[name = tensor<string, []>("op_2838_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_355_cast_fp16 = mul(x = var_2835_cast_fp16, y = var_2838_cast_fp16)[name = tensor<string, []>("x_355_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_253_cast_fp16 = mul(x = x_355_cast_fp16, y = conv_keep_cast_fp16)[name = tensor<string, []>("input_253_cast_fp16")];
            tensor<string, []> x_357_pad_type_0 = const()[name = tensor<string, []>("x_357_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_357_pad_0 = const()[name = tensor<string, []>("x_357_pad_0"), val = tensor<int32, [4]>([0, 0, 3, 3])];
            tensor<int32, []> x_357_groups_0 = const()[name = tensor<string, []>("x_357_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_357_strides_0 = const()[name = tensor<string, []>("x_357_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_357_dilations_0 = const()[name = tensor<string, []>("x_357_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 7]> core_stacks_2_stack_layers_3_conv_module2_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module2_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 7]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203358464)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_3_conv_module2_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module2_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203365696)))];
            tensor<fp16, [1, 512, 1, 256]> x_357_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_conv_module2_dw_bias_to_fp16, dilations = x_357_dilations_0, groups = x_357_groups_0, pad = x_357_pad_0, pad_type = x_357_pad_type_0, strides = x_357_strides_0, weight = core_stacks_2_stack_layers_3_conv_module2_dw_weight_to_fp16, x = input_253_cast_fp16)[name = tensor<string, []>("x_357_cast_fp16")];
            tensor<fp16, []> var_2848_to_fp16 = const()[name = tensor<string, []>("op_2848_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> x_359_cast_fp16 = sub(x = x_357_cast_fp16, y = var_2848_to_fp16)[name = tensor<string, []>("x_359_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2850_cast_fp16 = relu(x = x_359_cast_fp16)[name = tensor<string, []>("op_2850_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2851_cast_fp16 = abs(x = x_359_cast_fp16)[name = tensor<string, []>("op_2851_cast_fp16")];
            tensor<fp16, []> const_62_promoted_to_fp16 = const()[name = tensor<string, []>("const_62_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2852_cast_fp16 = mul(x = var_2851_cast_fp16, y = const_62_promoted_to_fp16)[name = tensor<string, []>("op_2852_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2853_cast_fp16 = exp(x = var_2852_cast_fp16)[name = tensor<string, []>("op_2853_cast_fp16")];
            tensor<fp32, []> var_2854_epsilon_0 = const()[name = tensor<string, []>("op_2854_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 256]> var_2854_cast_fp16 = log(epsilon = var_2854_epsilon_0, x = var_2853_cast_fp16)[name = tensor<string, []>("op_2854_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2855_cast_fp16 = add(x = var_2850_cast_fp16, y = var_2854_cast_fp16)[name = tensor<string, []>("op_2855_cast_fp16")];
            tensor<fp16, []> var_2856_to_fp16 = const()[name = tensor<string, []>("op_2856_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 256]> var_2857_cast_fp16 = mul(x = x_357_cast_fp16, y = var_2856_to_fp16)[name = tensor<string, []>("op_2857_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2858_cast_fp16 = sub(x = var_2855_cast_fp16, y = var_2857_cast_fp16)[name = tensor<string, []>("op_2858_cast_fp16")];
            tensor<fp16, []> var_2859_to_fp16 = const()[name = tensor<string, []>("op_2859_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 256]> input_255_cast_fp16 = sub(x = var_2858_cast_fp16, y = var_2859_to_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
            tensor<string, []> var_2867_pad_type_0 = const()[name = tensor<string, []>("op_2867_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2867_strides_0 = const()[name = tensor<string, []>("op_2867_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2867_pad_0 = const()[name = tensor<string, []>("op_2867_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2867_dilations_0 = const()[name = tensor<string, []>("op_2867_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2867_groups_0 = const()[name = tensor<string, []>("op_2867_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_2_stack_layers_3_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203366784)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_3_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203891136)))];
            tensor<fp16, [1, 512, 1, 256]> var_2867_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_conv_module2_out_proj_bias_to_fp16, dilations = var_2867_dilations_0, groups = var_2867_groups_0, pad = var_2867_pad_0, pad_type = var_2867_pad_type_0, strides = var_2867_strides_0, weight = core_stacks_2_stack_layers_3_conv_module2_out_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor<string, []>("op_2867_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_257_cast_fp16 = add(x = x_351_cast_fp16, y = var_2867_cast_fp16)[name = tensor<string, []>("input_257_cast_fp16")];
            tensor<string, []> x_361_pad_type_0 = const()[name = tensor<string, []>("x_361_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_361_strides_0 = const()[name = tensor<string, []>("x_361_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_361_pad_0 = const()[name = tensor<string, []>("x_361_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_361_dilations_0 = const()[name = tensor<string, []>("x_361_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_361_groups_0 = const()[name = tensor<string, []>("x_361_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_2_stack_layers_3_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203892224)))];
            tensor<fp16, [1920]> core_stacks_2_stack_layers_3_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205858368)))];
            tensor<fp16, [1, 1920, 1, 256]> x_361_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_feed_forward3_in_proj_bias_to_fp16, dilations = x_361_dilations_0, groups = x_361_groups_0, pad = x_361_pad_0, pad_type = x_361_pad_type_0, strides = x_361_strides_0, weight = core_stacks_2_stack_layers_3_feed_forward3_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor<string, []>("x_361_cast_fp16")];
            tensor<fp16, []> var_2878_to_fp16 = const()[name = tensor<string, []>("op_2878_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 256]> x_363_cast_fp16 = sub(x = x_361_cast_fp16, y = var_2878_to_fp16)[name = tensor<string, []>("x_363_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2880_cast_fp16 = relu(x = x_363_cast_fp16)[name = tensor<string, []>("op_2880_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2881_cast_fp16 = abs(x = x_363_cast_fp16)[name = tensor<string, []>("op_2881_cast_fp16")];
            tensor<fp16, []> const_63_promoted_to_fp16 = const()[name = tensor<string, []>("const_63_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 256]> var_2882_cast_fp16 = mul(x = var_2881_cast_fp16, y = const_63_promoted_to_fp16)[name = tensor<string, []>("op_2882_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2883_cast_fp16 = exp(x = var_2882_cast_fp16)[name = tensor<string, []>("op_2883_cast_fp16")];
            tensor<fp32, []> var_2884_epsilon_0 = const()[name = tensor<string, []>("op_2884_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 256]> var_2884_cast_fp16 = log(epsilon = var_2884_epsilon_0, x = var_2883_cast_fp16)[name = tensor<string, []>("op_2884_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2885_cast_fp16 = add(x = var_2880_cast_fp16, y = var_2884_cast_fp16)[name = tensor<string, []>("op_2885_cast_fp16")];
            tensor<fp16, []> var_2886_to_fp16 = const()[name = tensor<string, []>("op_2886_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 256]> var_2887_cast_fp16 = mul(x = x_361_cast_fp16, y = var_2886_to_fp16)[name = tensor<string, []>("op_2887_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 256]> var_2888_cast_fp16 = sub(x = var_2885_cast_fp16, y = var_2887_cast_fp16)[name = tensor<string, []>("op_2888_cast_fp16")];
            tensor<fp16, []> var_2889_to_fp16 = const()[name = tensor<string, []>("op_2889_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 256]> input_259_cast_fp16 = sub(x = var_2888_cast_fp16, y = var_2889_to_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
            tensor<string, []> var_2897_pad_type_0 = const()[name = tensor<string, []>("op_2897_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_2897_strides_0 = const()[name = tensor<string, []>("op_2897_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2897_pad_0 = const()[name = tensor<string, []>("op_2897_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2897_dilations_0 = const()[name = tensor<string, []>("op_2897_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_2897_groups_0 = const()[name = tensor<string, []>("op_2897_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_2_stack_layers_3_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(205862272)))];
            tensor<fp16, [512]> core_stacks_2_stack_layers_3_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207828416)))];
            tensor<fp16, [1, 512, 1, 256]> var_2897_cast_fp16 = conv(bias = core_stacks_2_stack_layers_3_feed_forward3_out_proj_bias_to_fp16, dilations = var_2897_dilations_0, groups = var_2897_groups_0, pad = var_2897_pad_0, pad_type = var_2897_pad_type_0, strides = var_2897_strides_0, weight = core_stacks_2_stack_layers_3_feed_forward3_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor<string, []>("op_2897_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x_365_cast_fp16 = add(x = input_257_cast_fp16, y = var_2897_cast_fp16)[name = tensor<string, []>("x_365_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_3_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207829504)))];
            tensor<fp16, [1, 512, 1, 256]> d_15_cast_fp16 = sub(x = x_365_cast_fp16, y = core_stacks_2_stack_layers_3_norm_bias_to_fp16)[name = tensor<string, []>("d_15_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2902_cast_fp16 = mul(x = d_15_cast_fp16, y = d_15_cast_fp16)[name = tensor<string, []>("op_2902_cast_fp16")];
            tensor<int32, [1]> var_2904_axes_0 = const()[name = tensor<string, []>("op_2904_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_2904_keep_dims_0 = const()[name = tensor<string, []>("op_2904_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 256]> var_2904_cast_fp16 = reduce_mean(axes = var_2904_axes_0, keep_dims = var_2904_keep_dims_0, x = var_2902_cast_fp16)[name = tensor<string, []>("op_2904_cast_fp16")];
            tensor<fp32, []> scales_15_epsilon_0 = const()[name = tensor<string, []>("scales_15_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 256]> scales_15_cast_fp16 = rsqrt(epsilon = scales_15_epsilon_0, x = var_2904_cast_fp16)[name = tensor<string, []>("scales_15_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2906_cast_fp16 = mul(x = x_365_cast_fp16, y = scales_15_cast_fp16)[name = tensor<string, []>("op_2906_cast_fp16")];
            tensor<fp16, []> core_stacks_2_stack_layers_3_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.b4p-2)];
            tensor<fp16, [1, 512, 1, 256]> src_33_cast_fp16 = mul(x = var_2906_cast_fp16, y = core_stacks_2_stack_layers_3_norm_scale_to_fp16)[name = tensor<string, []>("src_33_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_2909_cast_fp16 = sub(x = src_33_cast_fp16, y = input_231_cast_fp16)[name = tensor<string, []>("op_2909_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_stack_layers_3_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_stack_layers_3_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207830592)))];
            tensor<fp16, [1, 512, 1, 256]> var_2910_cast_fp16 = mul(x = var_2909_cast_fp16, y = core_stacks_2_stack_layers_3_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_2910_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> input_261_cast_fp16 = add(x = input_231_cast_fp16, y = var_2910_cast_fp16)[name = tensor<string, []>("input_261_cast_fp16")];
            tensor<fp32, []> src_35_scale_factor_height_0 = const()[name = tensor<string, []>("src_35_scale_factor_height_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, []> src_35_scale_factor_width_0 = const()[name = tensor<string, []>("src_35_scale_factor_width_0"), val = tensor<fp32, []>(0x1p+2)];
            tensor<fp16, [1, 512, 1, 1024]> src_35_cast_fp16 = upsample_nearest_neighbor(scale_factor_height = src_35_scale_factor_height_0, scale_factor_width = src_35_scale_factor_width_0, x = input_261_cast_fp16)[name = tensor<string, []>("src_35_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_2915_cast_fp16 = sub(x = src_35_cast_fp16, y = input_137_cast_fp16)[name = tensor<string, []>("op_2915_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_2_out_combiner_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_2_out_combiner_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207831680)))];
            tensor<fp16, [1, 512, 1, 1024]> var_2916_cast_fp16 = mul(x = var_2915_cast_fp16, y = core_stacks_2_out_combiner_bypass_scale_to_fp16)[name = tensor<string, []>("op_2916_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_263_cast_fp16 = add(x = input_137_cast_fp16, y = var_2916_cast_fp16)[name = tensor<string, []>("input_263_cast_fp16")];
            tensor<string, []> input_267_pad_type_0 = const()[name = tensor<string, []>("input_267_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> input_267_strides_0 = const()[name = tensor<string, []>("input_267_strides_0"), val = tensor<int32, [2]>([1, 2])];
            tensor<int32, []> input_267_groups_0 = const()[name = tensor<string, []>("input_267_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [4]> input_267_pad_0 = const()[name = tensor<string, []>("input_267_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_267_dilations_0 = const()[name = tensor<string, []>("input_267_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 2]> core_stacks_3_down_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_down_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 2]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207832768)))];
            tensor<fp16, [1, 512, 1, 512]> input_267_cast_fp16 = conv(dilations = input_267_dilations_0, groups = input_267_groups_0, pad = input_267_pad_0, pad_type = input_267_pad_type_0, strides = input_267_strides_0, weight = core_stacks_3_down_weight_to_fp16, x = input_263_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
            tensor<string, []> time_emb_7_pad_type_0 = const()[name = tensor<string, []>("time_emb_7_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> time_emb_7_strides_0 = const()[name = tensor<string, []>("time_emb_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> time_emb_7_pad_0 = const()[name = tensor<string, []>("time_emb_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> time_emb_7_dilations_0 = const()[name = tensor<string, []>("time_emb_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> time_emb_7_groups_0 = const()[name = tensor<string, []>("time_emb_7_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 192, 1, 1]> core_stacks_3_stack_time_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_time_proj_weight_to_fp16"), val = tensor<fp16, [512, 192, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207834880)))];
            tensor<fp16, [512]> core_stacks_3_stack_time_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_time_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208031552)))];
            tensor<fp16, [1, 512, 1, 1]> time_emb_7_cast_fp16 = conv(bias = core_stacks_3_stack_time_proj_bias_to_fp16, dilations = time_emb_7_dilations_0, groups = time_emb_7_groups_0, pad = time_emb_7_pad_0, pad_type = time_emb_7_pad_type_0, strides = time_emb_7_strides_0, weight = core_stacks_3_stack_time_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("time_emb_7_cast_fp16")];
            tensor<string, []> proj_17_pad_type_0 = const()[name = tensor<string, []>("proj_17_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_17_strides_0 = const()[name = tensor<string, []>("proj_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_17_pad_0 = const()[name = tensor<string, []>("proj_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_17_dilations_0 = const()[name = tensor<string, []>("proj_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_17_groups_0 = const()[name = tensor<string, []>("proj_17_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_3_stack_layers_0_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208032640)))];
            tensor<fp16, [384]> core_stacks_3_stack_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208425920)))];
            tensor<fp16, [1, 384, 1, 512]> proj_17_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_17_dilations_0, groups = proj_17_groups_0, pad = proj_17_pad_0, pad_type = proj_17_pad_type_0, strides = proj_17_strides_0, weight = core_stacks_3_stack_layers_0_self_attn_weights_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor<string, []>("proj_17_cast_fp16")];
            tensor<int32, [4]> var_2977_begin_0 = const()[name = tensor<string, []>("op_2977_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2977_end_0 = const()[name = tensor<string, []>("op_2977_end_0"), val = tensor<int32, [4]>([1, 128, 1, 512])];
            tensor<bool, [4]> var_2977_end_mask_0 = const()[name = tensor<string, []>("op_2977_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_2977_cast_fp16 = slice_by_index(begin = var_2977_begin_0, end = var_2977_end_0, end_mask = var_2977_end_mask_0, x = proj_17_cast_fp16)[name = tensor<string, []>("op_2977_cast_fp16")];
            tensor<int32, [4]> var_2978 = const()[name = tensor<string, []>("op_2978"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_2979_cast_fp16 = reshape(shape = var_2978, x = var_2977_cast_fp16)[name = tensor<string, []>("op_2979_cast_fp16")];
            tensor<int32, [4]> var_2981_begin_0 = const()[name = tensor<string, []>("op_2981_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_2981_end_0 = const()[name = tensor<string, []>("op_2981_end_0"), val = tensor<int32, [4]>([1, 256, 1, 512])];
            tensor<bool, [4]> var_2981_end_mask_0 = const()[name = tensor<string, []>("op_2981_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_2981_cast_fp16 = slice_by_index(begin = var_2981_begin_0, end = var_2981_end_0, end_mask = var_2981_end_mask_0, x = proj_17_cast_fp16)[name = tensor<string, []>("op_2981_cast_fp16")];
            tensor<int32, [4]> var_2982 = const()[name = tensor<string, []>("op_2982"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_2983_cast_fp16 = reshape(shape = var_2982, x = var_2981_cast_fp16)[name = tensor<string, []>("op_2983_cast_fp16")];
            tensor<int32, [4]> var_2985_begin_0 = const()[name = tensor<string, []>("op_2985_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_2985_end_0 = const()[name = tensor<string, []>("op_2985_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_2985_end_mask_0 = const()[name = tensor<string, []>("op_2985_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_2985_cast_fp16 = slice_by_index(begin = var_2985_begin_0, end = var_2985_end_0, end_mask = var_2985_end_mask_0, x = proj_17_cast_fp16)[name = tensor<string, []>("op_2985_cast_fp16")];
            tensor<int32, [4]> var_2986 = const()[name = tensor<string, []>("op_2986"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> pt_17_cast_fp16 = reshape(shape = var_2986, x = var_2985_cast_fp16)[name = tensor<string, []>("pt_17_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = tensor<string, []>("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_324 = const()[name = tensor<string, []>("concat_324"), val = tensor<int32, [3]>([4, 512, 32])];
            tensor<fp16, [4, 1, 512, 32]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = var_2979_cast_fp16)[name = tensor<string, []>("transpose_224")];
            tensor<fp16, [4, 512, 32]> reshape_96_cast_fp16 = reshape(shape = concat_324, x = transpose_64_cast_fp16)[name = tensor<string, []>("reshape_96_cast_fp16")];
            tensor<int32, [4]> transpose_65_perm_0 = const()[name = tensor<string, []>("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_325 = const()[name = tensor<string, []>("concat_325"), val = tensor<int32, [3]>([4, 32, 512])];
            tensor<fp16, [4, 1, 32, 512]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = var_2983_cast_fp16)[name = tensor<string, []>("transpose_223")];
            tensor<fp16, [4, 32, 512]> reshape_97_cast_fp16 = reshape(shape = concat_325, x = transpose_65_cast_fp16)[name = tensor<string, []>("reshape_97_cast_fp16")];
            tensor<bool, []> matmul_32_transpose_x_0 = const()[name = tensor<string, []>("matmul_32_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_32_transpose_y_0 = const()[name = tensor<string, []>("matmul_32_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 512]> matmul_32_cast_fp16 = matmul(transpose_x = matmul_32_transpose_x_0, transpose_y = matmul_32_transpose_y_0, x = reshape_96_cast_fp16, y = reshape_97_cast_fp16)[name = tensor<string, []>("matmul_32_cast_fp16")];
            tensor<int32, [4]> concat_329 = const()[name = tensor<string, []>("concat_329"), val = tensor<int32, [4]>([4, 1, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> reshape_98_cast_fp16 = reshape(shape = concat_329, x = matmul_32_cast_fp16)[name = tensor<string, []>("reshape_98_cast_fp16")];
            tensor<int32, [4]> attn_scores_49_perm_0 = const()[name = tensor<string, []>("attn_scores_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_2990 = const()[name = tensor<string, []>("op_2990"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 512, 1]> var_2991_cast_fp16 = transpose(perm = var_2990, x = pt_17_cast_fp16)[name = tensor<string, []>("transpose_221")];
            tensor<fp16, [4, 32, 512, 512]> var_2992_cast_fp16 = mul(x = var_2991_cast_fp16, y = core_pos_basis_512_to_fp16)[name = tensor<string, []>("op_2992_cast_fp16")];
            tensor<int32, [1]> pos_scores_17_axes_0 = const()[name = tensor<string, []>("pos_scores_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_17_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_17_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 512, 512]> pos_scores_17_cast_fp16 = reduce_sum(axes = pos_scores_17_axes_0, keep_dims = pos_scores_17_keep_dims_0, x = var_2992_cast_fp16)[name = tensor<string, []>("pos_scores_17_cast_fp16")];
            tensor<int32, [4]> var_2995 = const()[name = tensor<string, []>("op_2995"), val = tensor<int32, [4]>([4, 512, 1, 512])];
            tensor<fp16, [4, 512, 1, 512]> var_2996_cast_fp16 = reshape(shape = var_2995, x = pos_scores_17_cast_fp16)[name = tensor<string, []>("op_2996_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_49_cast_fp16 = transpose(perm = attn_scores_49_perm_0, x = reshape_98_cast_fp16)[name = tensor<string, []>("transpose_222")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_51_cast_fp16 = add(x = attn_scores_49_cast_fp16, y = var_2996_cast_fp16)[name = tensor<string, []>("attn_scores_51_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_53_cast_fp16 = add(x = attn_scores_51_cast_fp16, y = attn_bias_3_cast_fp16)[name = tensor<string, []>("attn_scores_53_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_weights_17_cast_fp16 = softmax(axis = var_41, x = attn_scores_53_cast_fp16)[name = tensor<string, []>("attn_weights_17_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_269_cast_fp16 = add(x = input_267_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("input_269_cast_fp16")];
            tensor<string, []> x_369_pad_type_0 = const()[name = tensor<string, []>("x_369_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_369_strides_0 = const()[name = tensor<string, []>("x_369_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_369_pad_0 = const()[name = tensor<string, []>("x_369_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_369_dilations_0 = const()[name = tensor<string, []>("x_369_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_369_groups_0 = const()[name = tensor<string, []>("x_369_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_3_stack_layers_0_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208426752)))];
            tensor<fp16, [1152]> core_stacks_3_stack_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209606464)))];
            tensor<fp16, [1, 1152, 1, 512]> x_369_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_feed_forward1_in_proj_bias_to_fp16, dilations = x_369_dilations_0, groups = x_369_groups_0, pad = x_369_pad_0, pad_type = x_369_pad_type_0, strides = x_369_strides_0, weight = core_stacks_3_stack_layers_0_feed_forward1_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor<string, []>("x_369_cast_fp16")];
            tensor<fp16, []> var_3010_to_fp16 = const()[name = tensor<string, []>("op_3010_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 512]> x_371_cast_fp16 = sub(x = x_369_cast_fp16, y = var_3010_to_fp16)[name = tensor<string, []>("x_371_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3012_cast_fp16 = relu(x = x_371_cast_fp16)[name = tensor<string, []>("op_3012_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3013_cast_fp16 = abs(x = x_371_cast_fp16)[name = tensor<string, []>("op_3013_cast_fp16")];
            tensor<fp16, []> const_65_promoted_to_fp16 = const()[name = tensor<string, []>("const_65_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_3014_cast_fp16 = mul(x = var_3013_cast_fp16, y = const_65_promoted_to_fp16)[name = tensor<string, []>("op_3014_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3015_cast_fp16 = exp(x = var_3014_cast_fp16)[name = tensor<string, []>("op_3015_cast_fp16")];
            tensor<fp32, []> var_3016_epsilon_0 = const()[name = tensor<string, []>("op_3016_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_3016_cast_fp16 = log(epsilon = var_3016_epsilon_0, x = var_3015_cast_fp16)[name = tensor<string, []>("op_3016_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3017_cast_fp16 = add(x = var_3012_cast_fp16, y = var_3016_cast_fp16)[name = tensor<string, []>("op_3017_cast_fp16")];
            tensor<fp16, []> var_3018_to_fp16 = const()[name = tensor<string, []>("op_3018_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 512]> var_3019_cast_fp16 = mul(x = x_369_cast_fp16, y = var_3018_to_fp16)[name = tensor<string, []>("op_3019_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3020_cast_fp16 = sub(x = var_3017_cast_fp16, y = var_3019_cast_fp16)[name = tensor<string, []>("op_3020_cast_fp16")];
            tensor<fp16, []> var_3021_to_fp16 = const()[name = tensor<string, []>("op_3021_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 512]> input_271_cast_fp16 = sub(x = var_3020_cast_fp16, y = var_3021_to_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
            tensor<string, []> var_3029_pad_type_0 = const()[name = tensor<string, []>("op_3029_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3029_strides_0 = const()[name = tensor<string, []>("op_3029_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3029_pad_0 = const()[name = tensor<string, []>("op_3029_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3029_dilations_0 = const()[name = tensor<string, []>("op_3029_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3029_groups_0 = const()[name = tensor<string, []>("op_3029_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_3_stack_layers_0_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209608832)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210788544)))];
            tensor<fp16, [1, 512, 1, 512]> var_3029_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_feed_forward1_out_proj_bias_to_fp16, dilations = var_3029_dilations_0, groups = var_3029_groups_0, pad = var_3029_pad_0, pad_type = var_3029_pad_type_0, strides = var_3029_strides_0, weight = core_stacks_3_stack_layers_0_feed_forward1_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor<string, []>("op_3029_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_273_cast_fp16 = add(x = input_269_cast_fp16, y = var_3029_cast_fp16)[name = tensor<string, []>("input_273_cast_fp16")];
            tensor<int32, [4]> var_3031_begin_0 = const()[name = tensor<string, []>("op_3031_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3031_end_0 = const()[name = tensor<string, []>("op_3031_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_3031_end_mask_0 = const()[name = tensor<string, []>("op_3031_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3031_cast_fp16 = slice_by_index(begin = var_3031_begin_0, end = var_3031_end_0, end_mask = var_3031_end_mask_0, x = attn_weights_17_cast_fp16)[name = tensor<string, []>("op_3031_cast_fp16")];
            tensor<string, []> x_373_pad_type_0 = const()[name = tensor<string, []>("x_373_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_373_strides_0 = const()[name = tensor<string, []>("x_373_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_373_pad_0 = const()[name = tensor<string, []>("x_373_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_373_dilations_0 = const()[name = tensor<string, []>("x_373_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_373_groups_0 = const()[name = tensor<string, []>("x_373_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_3_stack_layers_0_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210789632)))];
            tensor<fp16, [1152]> core_stacks_3_stack_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211969344)))];
            tensor<fp16, [1, 1152, 1, 512]> x_373_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_nonlin_attention_in_proj_bias_to_fp16, dilations = x_373_dilations_0, groups = x_373_groups_0, pad = x_373_pad_0, pad_type = x_373_pad_type_0, strides = x_373_strides_0, weight = core_stacks_3_stack_layers_0_nonlin_attention_in_proj_weight_to_fp16, x = input_273_cast_fp16)[name = tensor<string, []>("x_373_cast_fp16")];
            tensor<int32, [4]> var_3042_begin_0 = const()[name = tensor<string, []>("op_3042_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3042_end_0 = const()[name = tensor<string, []>("op_3042_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_3042_end_mask_0 = const()[name = tensor<string, []>("op_3042_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_3042_cast_fp16 = slice_by_index(begin = var_3042_begin_0, end = var_3042_end_0, end_mask = var_3042_end_mask_0, x = x_373_cast_fp16)[name = tensor<string, []>("op_3042_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> s_gate_17_cast_fp16 = tanh(x = var_3042_cast_fp16)[name = tensor<string, []>("s_gate_17_cast_fp16")];
            tensor<int32, [4]> var_3045_begin_0 = const()[name = tensor<string, []>("op_3045_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_3045_end_0 = const()[name = tensor<string, []>("op_3045_end_0"), val = tensor<int32, [4]>([1, 768, 1, 512])];
            tensor<bool, [4]> var_3045_end_mask_0 = const()[name = tensor<string, []>("op_3045_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_3045_cast_fp16 = slice_by_index(begin = var_3045_begin_0, end = var_3045_end_0, end_mask = var_3045_end_mask_0, x = x_373_cast_fp16)[name = tensor<string, []>("op_3045_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> var_3046_cast_fp16 = mul(x = var_3045_cast_fp16, y = s_gate_17_cast_fp16)[name = tensor<string, []>("op_3046_cast_fp16")];
            tensor<int32, [4]> y_17_begin_0 = const()[name = tensor<string, []>("y_17_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_17_end_0 = const()[name = tensor<string, []>("y_17_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 512])];
            tensor<bool, [4]> y_17_end_mask_0 = const()[name = tensor<string, []>("y_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 512]> y_17_cast_fp16 = slice_by_index(begin = y_17_begin_0, end = y_17_end_0, end_mask = y_17_end_mask_0, x = x_373_cast_fp16)[name = tensor<string, []>("y_17_cast_fp16")];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = tensor<string, []>("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_334 = const()[name = tensor<string, []>("concat_334"), val = tensor<int32, [3]>([1, 512, 512])];
            tensor<fp16, [1, 1, 512, 512]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = var_3031_cast_fp16)[name = tensor<string, []>("transpose_220")];
            tensor<fp16, [1, 512, 512]> reshape_99_cast_fp16 = reshape(shape = concat_334, x = transpose_66_cast_fp16)[name = tensor<string, []>("reshape_99_cast_fp16")];
            tensor<int32, [4]> transpose_67_perm_0 = const()[name = tensor<string, []>("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_335 = const()[name = tensor<string, []>("concat_335"), val = tensor<int32, [3]>([1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> transpose_67_cast_fp16 = transpose(perm = transpose_67_perm_0, x = var_3046_cast_fp16)[name = tensor<string, []>("transpose_219")];
            tensor<fp16, [1, 512, 384]> reshape_100_cast_fp16 = reshape(shape = concat_335, x = transpose_67_cast_fp16)[name = tensor<string, []>("reshape_100_cast_fp16")];
            tensor<bool, []> matmul_33_transpose_x_0 = const()[name = tensor<string, []>("matmul_33_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_33_transpose_y_0 = const()[name = tensor<string, []>("matmul_33_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 384]> matmul_33_cast_fp16 = matmul(transpose_x = matmul_33_transpose_x_0, transpose_y = matmul_33_transpose_y_0, x = reshape_99_cast_fp16, y = reshape_100_cast_fp16)[name = tensor<string, []>("matmul_33_cast_fp16")];
            tensor<int32, [4]> concat_339 = const()[name = tensor<string, []>("concat_339"), val = tensor<int32, [4]>([1, 1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> reshape_101_cast_fp16 = reshape(shape = concat_339, x = matmul_33_cast_fp16)[name = tensor<string, []>("reshape_101_cast_fp16")];
            tensor<int32, [4]> v_17_perm_0 = const()[name = tensor<string, []>("v_17_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 512]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = reshape_101_cast_fp16)[name = tensor<string, []>("transpose_218")];
            tensor<fp16, [1, 384, 1, 512]> input_275_cast_fp16 = mul(x = v_17_cast_fp16, y = y_17_cast_fp16)[name = tensor<string, []>("input_275_cast_fp16")];
            tensor<string, []> var_3058_pad_type_0 = const()[name = tensor<string, []>("op_3058_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3058_strides_0 = const()[name = tensor<string, []>("op_3058_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3058_pad_0 = const()[name = tensor<string, []>("op_3058_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3058_dilations_0 = const()[name = tensor<string, []>("op_3058_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3058_groups_0 = const()[name = tensor<string, []>("op_3058_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_3_stack_layers_0_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211971712)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212364992)))];
            tensor<fp16, [1, 512, 1, 512]> var_3058_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_nonlin_attention_out_proj_bias_to_fp16, dilations = var_3058_dilations_0, groups = var_3058_groups_0, pad = var_3058_pad_0, pad_type = var_3058_pad_type_0, strides = var_3058_strides_0, weight = core_stacks_3_stack_layers_0_nonlin_attention_out_proj_weight_to_fp16, x = input_275_cast_fp16)[name = tensor<string, []>("op_3058_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_375_cast_fp16 = add(x = input_273_cast_fp16, y = var_3058_cast_fp16)[name = tensor<string, []>("x_375_cast_fp16")];
            tensor<string, []> var_3069_pad_type_0 = const()[name = tensor<string, []>("op_3069_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3069_strides_0 = const()[name = tensor<string, []>("op_3069_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3069_pad_0 = const()[name = tensor<string, []>("op_3069_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3069_dilations_0 = const()[name = tensor<string, []>("op_3069_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3069_groups_0 = const()[name = tensor<string, []>("op_3069_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_3_stack_layers_0_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212366080)))];
            tensor<fp16, [48]> core_stacks_3_stack_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212415296)))];
            tensor<fp16, [1, 48, 1, 512]> var_3069_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_self_attn1_in_proj_bias_to_fp16, dilations = var_3069_dilations_0, groups = var_3069_groups_0, pad = var_3069_pad_0, pad_type = var_3069_pad_type_0, strides = var_3069_strides_0, weight = core_stacks_3_stack_layers_0_self_attn1_in_proj_weight_to_fp16, x = x_375_cast_fp16)[name = tensor<string, []>("op_3069_cast_fp16")];
            tensor<int32, [4]> var_3070 = const()[name = tensor<string, []>("op_3070"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_3071_cast_fp16 = reshape(shape = var_3070, x = var_3069_cast_fp16)[name = tensor<string, []>("op_3071_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = tensor<string, []>("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_344 = const()[name = tensor<string, []>("concat_344"), val = tensor<int32, [3]>([4, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = attn_weights_17_cast_fp16)[name = tensor<string, []>("transpose_217")];
            tensor<fp16, [4, 512, 512]> reshape_102_cast_fp16 = reshape(shape = concat_344, x = transpose_68_cast_fp16)[name = tensor<string, []>("reshape_102_cast_fp16")];
            tensor<int32, [4]> transpose_69_perm_0 = const()[name = tensor<string, []>("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_345 = const()[name = tensor<string, []>("concat_345"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = var_3071_cast_fp16)[name = tensor<string, []>("transpose_216")];
            tensor<fp16, [4, 512, 12]> reshape_103_cast_fp16 = reshape(shape = concat_345, x = transpose_69_cast_fp16)[name = tensor<string, []>("reshape_103_cast_fp16")];
            tensor<bool, []> matmul_34_transpose_x_0 = const()[name = tensor<string, []>("matmul_34_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_34_transpose_y_0 = const()[name = tensor<string, []>("matmul_34_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_34_cast_fp16 = matmul(transpose_x = matmul_34_transpose_x_0, transpose_y = matmul_34_transpose_y_0, x = reshape_102_cast_fp16, y = reshape_103_cast_fp16)[name = tensor<string, []>("matmul_34_cast_fp16")];
            tensor<int32, [4]> concat_349 = const()[name = tensor<string, []>("concat_349"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_104_cast_fp16 = reshape(shape = concat_349, x = matmul_34_cast_fp16)[name = tensor<string, []>("reshape_104_cast_fp16")];
            tensor<int32, [4]> out_33_perm_0 = const()[name = tensor<string, []>("out_33_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_3074 = const()[name = tensor<string, []>("op_3074"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_33_cast_fp16 = transpose(perm = out_33_perm_0, x = reshape_104_cast_fp16)[name = tensor<string, []>("transpose_215")];
            tensor<fp16, [1, 48, 1, 512]> input_277_cast_fp16 = reshape(shape = var_3074, x = out_33_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
            tensor<string, []> var_3082_pad_type_0 = const()[name = tensor<string, []>("op_3082_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3082_strides_0 = const()[name = tensor<string, []>("op_3082_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3082_pad_0 = const()[name = tensor<string, []>("op_3082_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3082_dilations_0 = const()[name = tensor<string, []>("op_3082_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3082_groups_0 = const()[name = tensor<string, []>("op_3082_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_3_stack_layers_0_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212415488)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212464704)))];
            tensor<fp16, [1, 512, 1, 512]> var_3082_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_self_attn1_out_proj_bias_to_fp16, dilations = var_3082_dilations_0, groups = var_3082_groups_0, pad = var_3082_pad_0, pad_type = var_3082_pad_type_0, strides = var_3082_strides_0, weight = core_stacks_3_stack_layers_0_self_attn1_out_proj_weight_to_fp16, x = input_277_cast_fp16)[name = tensor<string, []>("op_3082_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_377_cast_fp16 = add(x = x_375_cast_fp16, y = var_3082_cast_fp16)[name = tensor<string, []>("x_377_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_379_cast_fp16 = add(x = x_377_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("x_379_cast_fp16")];
            tensor<string, []> x_381_pad_type_0 = const()[name = tensor<string, []>("x_381_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_381_strides_0 = const()[name = tensor<string, []>("x_381_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_381_pad_0 = const()[name = tensor<string, []>("x_381_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_381_dilations_0 = const()[name = tensor<string, []>("x_381_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_381_groups_0 = const()[name = tensor<string, []>("x_381_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_3_stack_layers_0_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212465792)))];
            tensor<fp16, [1024]> core_stacks_3_stack_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213514432)))];
            tensor<fp16, [1, 1024, 1, 512]> x_381_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_conv_module1_in_proj_bias_to_fp16, dilations = x_381_dilations_0, groups = x_381_groups_0, pad = x_381_pad_0, pad_type = x_381_pad_type_0, strides = x_381_strides_0, weight = core_stacks_3_stack_layers_0_conv_module1_in_proj_weight_to_fp16, x = x_379_cast_fp16)[name = tensor<string, []>("x_381_cast_fp16")];
            tensor<int32, [4]> var_3096_begin_0 = const()[name = tensor<string, []>("op_3096_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3096_end_0 = const()[name = tensor<string, []>("op_3096_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_3096_end_mask_0 = const()[name = tensor<string, []>("op_3096_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3096_cast_fp16 = slice_by_index(begin = var_3096_begin_0, end = var_3096_end_0, end_mask = var_3096_end_mask_0, x = x_381_cast_fp16)[name = tensor<string, []>("op_3096_cast_fp16")];
            tensor<int32, [4]> var_3098_begin_0 = const()[name = tensor<string, []>("op_3098_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_3098_end_0 = const()[name = tensor<string, []>("op_3098_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_3098_end_mask_0 = const()[name = tensor<string, []>("op_3098_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3098_cast_fp16 = slice_by_index(begin = var_3098_begin_0, end = var_3098_end_0, end_mask = var_3098_end_mask_0, x = x_381_cast_fp16)[name = tensor<string, []>("op_3098_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3099_cast_fp16 = sigmoid(x = var_3098_cast_fp16)[name = tensor<string, []>("op_3099_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_383_cast_fp16 = mul(x = var_3096_cast_fp16, y = var_3099_cast_fp16)[name = tensor<string, []>("x_383_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_279_cast_fp16 = mul(x = x_383_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_279_cast_fp16")];
            tensor<string, []> x_385_pad_type_0 = const()[name = tensor<string, []>("x_385_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_385_pad_0 = const()[name = tensor<string, []>("x_385_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_385_groups_0 = const()[name = tensor<string, []>("x_385_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_385_strides_0 = const()[name = tensor<string, []>("x_385_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_385_dilations_0 = const()[name = tensor<string, []>("x_385_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_3_stack_layers_0_conv_module1_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module1_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213516544)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_0_conv_module1_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module1_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213531968)))];
            tensor<fp16, [1, 512, 1, 512]> x_385_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_conv_module1_dw_bias_to_fp16, dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = core_stacks_3_stack_layers_0_conv_module1_dw_weight_to_fp16, x = input_279_cast_fp16)[name = tensor<string, []>("x_385_cast_fp16")];
            tensor<fp16, []> var_3109_to_fp16 = const()[name = tensor<string, []>("op_3109_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_387_cast_fp16 = sub(x = x_385_cast_fp16, y = var_3109_to_fp16)[name = tensor<string, []>("x_387_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3111_cast_fp16 = relu(x = x_387_cast_fp16)[name = tensor<string, []>("op_3111_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3112_cast_fp16 = abs(x = x_387_cast_fp16)[name = tensor<string, []>("op_3112_cast_fp16")];
            tensor<fp16, []> const_67_promoted_to_fp16 = const()[name = tensor<string, []>("const_67_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3113_cast_fp16 = mul(x = var_3112_cast_fp16, y = const_67_promoted_to_fp16)[name = tensor<string, []>("op_3113_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3114_cast_fp16 = exp(x = var_3113_cast_fp16)[name = tensor<string, []>("op_3114_cast_fp16")];
            tensor<fp32, []> var_3115_epsilon_0 = const()[name = tensor<string, []>("op_3115_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3115_cast_fp16 = log(epsilon = var_3115_epsilon_0, x = var_3114_cast_fp16)[name = tensor<string, []>("op_3115_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3116_cast_fp16 = add(x = var_3111_cast_fp16, y = var_3115_cast_fp16)[name = tensor<string, []>("op_3116_cast_fp16")];
            tensor<fp16, []> var_3117_to_fp16 = const()[name = tensor<string, []>("op_3117_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_3118_cast_fp16 = mul(x = x_385_cast_fp16, y = var_3117_to_fp16)[name = tensor<string, []>("op_3118_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3119_cast_fp16 = sub(x = var_3116_cast_fp16, y = var_3118_cast_fp16)[name = tensor<string, []>("op_3119_cast_fp16")];
            tensor<fp16, []> var_3120_to_fp16 = const()[name = tensor<string, []>("op_3120_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_281_cast_fp16 = sub(x = var_3119_cast_fp16, y = var_3120_to_fp16)[name = tensor<string, []>("input_281_cast_fp16")];
            tensor<string, []> var_3128_pad_type_0 = const()[name = tensor<string, []>("op_3128_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3128_strides_0 = const()[name = tensor<string, []>("op_3128_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3128_pad_0 = const()[name = tensor<string, []>("op_3128_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3128_dilations_0 = const()[name = tensor<string, []>("op_3128_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3128_groups_0 = const()[name = tensor<string, []>("op_3128_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_3_stack_layers_0_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213533056)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214057408)))];
            tensor<fp16, [1, 512, 1, 512]> var_3128_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_conv_module1_out_proj_bias_to_fp16, dilations = var_3128_dilations_0, groups = var_3128_groups_0, pad = var_3128_pad_0, pad_type = var_3128_pad_type_0, strides = var_3128_strides_0, weight = core_stacks_3_stack_layers_0_conv_module1_out_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor<string, []>("op_3128_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_283_cast_fp16 = add(x = x_379_cast_fp16, y = var_3128_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
            tensor<string, []> x_389_pad_type_0 = const()[name = tensor<string, []>("x_389_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_389_strides_0 = const()[name = tensor<string, []>("x_389_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_389_pad_0 = const()[name = tensor<string, []>("x_389_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_389_dilations_0 = const()[name = tensor<string, []>("x_389_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_389_groups_0 = const()[name = tensor<string, []>("x_389_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_3_stack_layers_0_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214058496)))];
            tensor<fp16, [1536]> core_stacks_3_stack_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215631424)))];
            tensor<fp16, [1, 1536, 1, 512]> x_389_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_feed_forward2_in_proj_bias_to_fp16, dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = core_stacks_3_stack_layers_0_feed_forward2_in_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor<string, []>("x_389_cast_fp16")];
            tensor<fp16, []> var_3139_to_fp16 = const()[name = tensor<string, []>("op_3139_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 512]> x_391_cast_fp16 = sub(x = x_389_cast_fp16, y = var_3139_to_fp16)[name = tensor<string, []>("x_391_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3141_cast_fp16 = relu(x = x_391_cast_fp16)[name = tensor<string, []>("op_3141_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3142_cast_fp16 = abs(x = x_391_cast_fp16)[name = tensor<string, []>("op_3142_cast_fp16")];
            tensor<fp16, []> const_68_promoted_to_fp16 = const()[name = tensor<string, []>("const_68_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_3143_cast_fp16 = mul(x = var_3142_cast_fp16, y = const_68_promoted_to_fp16)[name = tensor<string, []>("op_3143_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3144_cast_fp16 = exp(x = var_3143_cast_fp16)[name = tensor<string, []>("op_3144_cast_fp16")];
            tensor<fp32, []> var_3145_epsilon_0 = const()[name = tensor<string, []>("op_3145_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_3145_cast_fp16 = log(epsilon = var_3145_epsilon_0, x = var_3144_cast_fp16)[name = tensor<string, []>("op_3145_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3146_cast_fp16 = add(x = var_3141_cast_fp16, y = var_3145_cast_fp16)[name = tensor<string, []>("op_3146_cast_fp16")];
            tensor<fp16, []> var_3147_to_fp16 = const()[name = tensor<string, []>("op_3147_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 512]> var_3148_cast_fp16 = mul(x = x_389_cast_fp16, y = var_3147_to_fp16)[name = tensor<string, []>("op_3148_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3149_cast_fp16 = sub(x = var_3146_cast_fp16, y = var_3148_cast_fp16)[name = tensor<string, []>("op_3149_cast_fp16")];
            tensor<fp16, []> var_3150_to_fp16 = const()[name = tensor<string, []>("op_3150_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 512]> input_285_cast_fp16 = sub(x = var_3149_cast_fp16, y = var_3150_to_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
            tensor<string, []> var_3158_pad_type_0 = const()[name = tensor<string, []>("op_3158_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3158_strides_0 = const()[name = tensor<string, []>("op_3158_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3158_pad_0 = const()[name = tensor<string, []>("op_3158_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3158_dilations_0 = const()[name = tensor<string, []>("op_3158_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3158_groups_0 = const()[name = tensor<string, []>("op_3158_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_3_stack_layers_0_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215634560)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217207488)))];
            tensor<fp16, [1, 512, 1, 512]> var_3158_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_feed_forward2_out_proj_bias_to_fp16, dilations = var_3158_dilations_0, groups = var_3158_groups_0, pad = var_3158_pad_0, pad_type = var_3158_pad_type_0, strides = var_3158_strides_0, weight = core_stacks_3_stack_layers_0_feed_forward2_out_proj_weight_to_fp16, x = input_285_cast_fp16)[name = tensor<string, []>("op_3158_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> src_37_cast_fp16 = add(x = input_283_cast_fp16, y = var_3158_cast_fp16)[name = tensor<string, []>("src_37_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3161_cast_fp16 = sub(x = src_37_cast_fp16, y = input_267_cast_fp16)[name = tensor<string, []>("op_3161_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217208576)))];
            tensor<fp16, [1, 512, 1, 512]> var_3162_cast_fp16 = mul(x = var_3161_cast_fp16, y = core_stacks_3_stack_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_3162_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_393_cast_fp16 = add(x = input_267_cast_fp16, y = var_3162_cast_fp16)[name = tensor<string, []>("x_393_cast_fp16")];
            tensor<string, []> var_3173_pad_type_0 = const()[name = tensor<string, []>("op_3173_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3173_strides_0 = const()[name = tensor<string, []>("op_3173_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3173_pad_0 = const()[name = tensor<string, []>("op_3173_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3173_dilations_0 = const()[name = tensor<string, []>("op_3173_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3173_groups_0 = const()[name = tensor<string, []>("op_3173_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_3_stack_layers_0_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217209664)))];
            tensor<fp16, [48]> core_stacks_3_stack_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217258880)))];
            tensor<fp16, [1, 48, 1, 512]> var_3173_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_self_attn2_in_proj_bias_to_fp16, dilations = var_3173_dilations_0, groups = var_3173_groups_0, pad = var_3173_pad_0, pad_type = var_3173_pad_type_0, strides = var_3173_strides_0, weight = core_stacks_3_stack_layers_0_self_attn2_in_proj_weight_to_fp16, x = x_393_cast_fp16)[name = tensor<string, []>("op_3173_cast_fp16")];
            tensor<int32, [4]> var_3174 = const()[name = tensor<string, []>("op_3174"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_3175_cast_fp16 = reshape(shape = var_3174, x = var_3173_cast_fp16)[name = tensor<string, []>("op_3175_cast_fp16")];
            tensor<int32, [4]> transpose_71_perm_0 = const()[name = tensor<string, []>("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_355 = const()[name = tensor<string, []>("concat_355"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_71_cast_fp16 = transpose(perm = transpose_71_perm_0, x = var_3175_cast_fp16)[name = tensor<string, []>("transpose_214")];
            tensor<fp16, [4, 512, 12]> reshape_106_cast_fp16 = reshape(shape = concat_355, x = transpose_71_cast_fp16)[name = tensor<string, []>("reshape_106_cast_fp16")];
            tensor<bool, []> matmul_35_transpose_x_0 = const()[name = tensor<string, []>("matmul_35_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_35_transpose_y_0 = const()[name = tensor<string, []>("matmul_35_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_35_cast_fp16 = matmul(transpose_x = matmul_35_transpose_x_0, transpose_y = matmul_35_transpose_y_0, x = reshape_102_cast_fp16, y = reshape_106_cast_fp16)[name = tensor<string, []>("matmul_35_cast_fp16")];
            tensor<int32, [4]> concat_359 = const()[name = tensor<string, []>("concat_359"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_107_cast_fp16 = reshape(shape = concat_359, x = matmul_35_cast_fp16)[name = tensor<string, []>("reshape_107_cast_fp16")];
            tensor<int32, [4]> out_35_perm_0 = const()[name = tensor<string, []>("out_35_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_3178 = const()[name = tensor<string, []>("op_3178"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_35_cast_fp16 = transpose(perm = out_35_perm_0, x = reshape_107_cast_fp16)[name = tensor<string, []>("transpose_213")];
            tensor<fp16, [1, 48, 1, 512]> input_287_cast_fp16 = reshape(shape = var_3178, x = out_35_cast_fp16)[name = tensor<string, []>("input_287_cast_fp16")];
            tensor<string, []> var_3186_pad_type_0 = const()[name = tensor<string, []>("op_3186_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3186_strides_0 = const()[name = tensor<string, []>("op_3186_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3186_pad_0 = const()[name = tensor<string, []>("op_3186_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3186_dilations_0 = const()[name = tensor<string, []>("op_3186_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3186_groups_0 = const()[name = tensor<string, []>("op_3186_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_3_stack_layers_0_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217259072)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217308288)))];
            tensor<fp16, [1, 512, 1, 512]> var_3186_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_self_attn2_out_proj_bias_to_fp16, dilations = var_3186_dilations_0, groups = var_3186_groups_0, pad = var_3186_pad_0, pad_type = var_3186_pad_type_0, strides = var_3186_strides_0, weight = core_stacks_3_stack_layers_0_self_attn2_out_proj_weight_to_fp16, x = input_287_cast_fp16)[name = tensor<string, []>("op_3186_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_395_cast_fp16 = add(x = x_393_cast_fp16, y = var_3186_cast_fp16)[name = tensor<string, []>("x_395_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_397_cast_fp16 = add(x = x_395_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("x_397_cast_fp16")];
            tensor<string, []> x_399_pad_type_0 = const()[name = tensor<string, []>("x_399_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_399_strides_0 = const()[name = tensor<string, []>("x_399_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_399_pad_0 = const()[name = tensor<string, []>("x_399_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_399_dilations_0 = const()[name = tensor<string, []>("x_399_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_399_groups_0 = const()[name = tensor<string, []>("x_399_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_3_stack_layers_0_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(217309376)))];
            tensor<fp16, [1024]> core_stacks_3_stack_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218358016)))];
            tensor<fp16, [1, 1024, 1, 512]> x_399_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_conv_module2_in_proj_bias_to_fp16, dilations = x_399_dilations_0, groups = x_399_groups_0, pad = x_399_pad_0, pad_type = x_399_pad_type_0, strides = x_399_strides_0, weight = core_stacks_3_stack_layers_0_conv_module2_in_proj_weight_to_fp16, x = x_397_cast_fp16)[name = tensor<string, []>("x_399_cast_fp16")];
            tensor<int32, [4]> var_3200_begin_0 = const()[name = tensor<string, []>("op_3200_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3200_end_0 = const()[name = tensor<string, []>("op_3200_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_3200_end_mask_0 = const()[name = tensor<string, []>("op_3200_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3200_cast_fp16 = slice_by_index(begin = var_3200_begin_0, end = var_3200_end_0, end_mask = var_3200_end_mask_0, x = x_399_cast_fp16)[name = tensor<string, []>("op_3200_cast_fp16")];
            tensor<int32, [4]> var_3202_begin_0 = const()[name = tensor<string, []>("op_3202_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_3202_end_0 = const()[name = tensor<string, []>("op_3202_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_3202_end_mask_0 = const()[name = tensor<string, []>("op_3202_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3202_cast_fp16 = slice_by_index(begin = var_3202_begin_0, end = var_3202_end_0, end_mask = var_3202_end_mask_0, x = x_399_cast_fp16)[name = tensor<string, []>("op_3202_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3203_cast_fp16 = sigmoid(x = var_3202_cast_fp16)[name = tensor<string, []>("op_3203_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_401_cast_fp16 = mul(x = var_3200_cast_fp16, y = var_3203_cast_fp16)[name = tensor<string, []>("x_401_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_289_cast_fp16 = mul(x = x_401_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_289_cast_fp16")];
            tensor<string, []> x_403_pad_type_0 = const()[name = tensor<string, []>("x_403_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_403_pad_0 = const()[name = tensor<string, []>("x_403_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_403_groups_0 = const()[name = tensor<string, []>("x_403_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_403_strides_0 = const()[name = tensor<string, []>("x_403_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_403_dilations_0 = const()[name = tensor<string, []>("x_403_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_3_stack_layers_0_conv_module2_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module2_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218360128)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_0_conv_module2_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module2_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218375552)))];
            tensor<fp16, [1, 512, 1, 512]> x_403_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_conv_module2_dw_bias_to_fp16, dilations = x_403_dilations_0, groups = x_403_groups_0, pad = x_403_pad_0, pad_type = x_403_pad_type_0, strides = x_403_strides_0, weight = core_stacks_3_stack_layers_0_conv_module2_dw_weight_to_fp16, x = input_289_cast_fp16)[name = tensor<string, []>("x_403_cast_fp16")];
            tensor<fp16, []> var_3213_to_fp16 = const()[name = tensor<string, []>("op_3213_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_405_cast_fp16 = sub(x = x_403_cast_fp16, y = var_3213_to_fp16)[name = tensor<string, []>("x_405_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3215_cast_fp16 = relu(x = x_405_cast_fp16)[name = tensor<string, []>("op_3215_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3216_cast_fp16 = abs(x = x_405_cast_fp16)[name = tensor<string, []>("op_3216_cast_fp16")];
            tensor<fp16, []> const_70_promoted_to_fp16 = const()[name = tensor<string, []>("const_70_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3217_cast_fp16 = mul(x = var_3216_cast_fp16, y = const_70_promoted_to_fp16)[name = tensor<string, []>("op_3217_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3218_cast_fp16 = exp(x = var_3217_cast_fp16)[name = tensor<string, []>("op_3218_cast_fp16")];
            tensor<fp32, []> var_3219_epsilon_0 = const()[name = tensor<string, []>("op_3219_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3219_cast_fp16 = log(epsilon = var_3219_epsilon_0, x = var_3218_cast_fp16)[name = tensor<string, []>("op_3219_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3220_cast_fp16 = add(x = var_3215_cast_fp16, y = var_3219_cast_fp16)[name = tensor<string, []>("op_3220_cast_fp16")];
            tensor<fp16, []> var_3221_to_fp16 = const()[name = tensor<string, []>("op_3221_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_3222_cast_fp16 = mul(x = x_403_cast_fp16, y = var_3221_to_fp16)[name = tensor<string, []>("op_3222_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3223_cast_fp16 = sub(x = var_3220_cast_fp16, y = var_3222_cast_fp16)[name = tensor<string, []>("op_3223_cast_fp16")];
            tensor<fp16, []> var_3224_to_fp16 = const()[name = tensor<string, []>("op_3224_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_291_cast_fp16 = sub(x = var_3223_cast_fp16, y = var_3224_to_fp16)[name = tensor<string, []>("input_291_cast_fp16")];
            tensor<string, []> var_3232_pad_type_0 = const()[name = tensor<string, []>("op_3232_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3232_strides_0 = const()[name = tensor<string, []>("op_3232_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3232_pad_0 = const()[name = tensor<string, []>("op_3232_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3232_dilations_0 = const()[name = tensor<string, []>("op_3232_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3232_groups_0 = const()[name = tensor<string, []>("op_3232_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_3_stack_layers_0_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218376640)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218900992)))];
            tensor<fp16, [1, 512, 1, 512]> var_3232_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_conv_module2_out_proj_bias_to_fp16, dilations = var_3232_dilations_0, groups = var_3232_groups_0, pad = var_3232_pad_0, pad_type = var_3232_pad_type_0, strides = var_3232_strides_0, weight = core_stacks_3_stack_layers_0_conv_module2_out_proj_weight_to_fp16, x = input_291_cast_fp16)[name = tensor<string, []>("op_3232_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_293_cast_fp16 = add(x = x_397_cast_fp16, y = var_3232_cast_fp16)[name = tensor<string, []>("input_293_cast_fp16")];
            tensor<string, []> x_407_pad_type_0 = const()[name = tensor<string, []>("x_407_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_407_strides_0 = const()[name = tensor<string, []>("x_407_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_407_pad_0 = const()[name = tensor<string, []>("x_407_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_407_dilations_0 = const()[name = tensor<string, []>("x_407_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_407_groups_0 = const()[name = tensor<string, []>("x_407_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_3_stack_layers_0_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218902080)))];
            tensor<fp16, [1920]> core_stacks_3_stack_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220868224)))];
            tensor<fp16, [1, 1920, 1, 512]> x_407_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_feed_forward3_in_proj_bias_to_fp16, dilations = x_407_dilations_0, groups = x_407_groups_0, pad = x_407_pad_0, pad_type = x_407_pad_type_0, strides = x_407_strides_0, weight = core_stacks_3_stack_layers_0_feed_forward3_in_proj_weight_to_fp16, x = input_293_cast_fp16)[name = tensor<string, []>("x_407_cast_fp16")];
            tensor<fp16, []> var_3243_to_fp16 = const()[name = tensor<string, []>("op_3243_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 512]> x_409_cast_fp16 = sub(x = x_407_cast_fp16, y = var_3243_to_fp16)[name = tensor<string, []>("x_409_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3245_cast_fp16 = relu(x = x_409_cast_fp16)[name = tensor<string, []>("op_3245_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3246_cast_fp16 = abs(x = x_409_cast_fp16)[name = tensor<string, []>("op_3246_cast_fp16")];
            tensor<fp16, []> const_71_promoted_to_fp16 = const()[name = tensor<string, []>("const_71_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_3247_cast_fp16 = mul(x = var_3246_cast_fp16, y = const_71_promoted_to_fp16)[name = tensor<string, []>("op_3247_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3248_cast_fp16 = exp(x = var_3247_cast_fp16)[name = tensor<string, []>("op_3248_cast_fp16")];
            tensor<fp32, []> var_3249_epsilon_0 = const()[name = tensor<string, []>("op_3249_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_3249_cast_fp16 = log(epsilon = var_3249_epsilon_0, x = var_3248_cast_fp16)[name = tensor<string, []>("op_3249_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3250_cast_fp16 = add(x = var_3245_cast_fp16, y = var_3249_cast_fp16)[name = tensor<string, []>("op_3250_cast_fp16")];
            tensor<fp16, []> var_3251_to_fp16 = const()[name = tensor<string, []>("op_3251_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 512]> var_3252_cast_fp16 = mul(x = x_407_cast_fp16, y = var_3251_to_fp16)[name = tensor<string, []>("op_3252_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3253_cast_fp16 = sub(x = var_3250_cast_fp16, y = var_3252_cast_fp16)[name = tensor<string, []>("op_3253_cast_fp16")];
            tensor<fp16, []> var_3254_to_fp16 = const()[name = tensor<string, []>("op_3254_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 512]> input_295_cast_fp16 = sub(x = var_3253_cast_fp16, y = var_3254_to_fp16)[name = tensor<string, []>("input_295_cast_fp16")];
            tensor<string, []> var_3262_pad_type_0 = const()[name = tensor<string, []>("op_3262_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3262_strides_0 = const()[name = tensor<string, []>("op_3262_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3262_pad_0 = const()[name = tensor<string, []>("op_3262_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3262_dilations_0 = const()[name = tensor<string, []>("op_3262_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3262_groups_0 = const()[name = tensor<string, []>("op_3262_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_3_stack_layers_0_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220872128)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222838272)))];
            tensor<fp16, [1, 512, 1, 512]> var_3262_cast_fp16 = conv(bias = core_stacks_3_stack_layers_0_feed_forward3_out_proj_bias_to_fp16, dilations = var_3262_dilations_0, groups = var_3262_groups_0, pad = var_3262_pad_0, pad_type = var_3262_pad_type_0, strides = var_3262_strides_0, weight = core_stacks_3_stack_layers_0_feed_forward3_out_proj_weight_to_fp16, x = input_295_cast_fp16)[name = tensor<string, []>("op_3262_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_411_cast_fp16 = add(x = input_293_cast_fp16, y = var_3262_cast_fp16)[name = tensor<string, []>("x_411_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222839360)))];
            tensor<fp16, [1, 512, 1, 512]> d_17_cast_fp16 = sub(x = x_411_cast_fp16, y = core_stacks_3_stack_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("d_17_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3267_cast_fp16 = mul(x = d_17_cast_fp16, y = d_17_cast_fp16)[name = tensor<string, []>("op_3267_cast_fp16")];
            tensor<int32, [1]> var_3269_axes_0 = const()[name = tensor<string, []>("op_3269_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_3269_keep_dims_0 = const()[name = tensor<string, []>("op_3269_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 512]> var_3269_cast_fp16 = reduce_mean(axes = var_3269_axes_0, keep_dims = var_3269_keep_dims_0, x = var_3267_cast_fp16)[name = tensor<string, []>("op_3269_cast_fp16")];
            tensor<fp32, []> scales_17_epsilon_0 = const()[name = tensor<string, []>("scales_17_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 512]> scales_17_cast_fp16 = rsqrt(epsilon = scales_17_epsilon_0, x = var_3269_cast_fp16)[name = tensor<string, []>("scales_17_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3271_cast_fp16 = mul(x = x_411_cast_fp16, y = scales_17_cast_fp16)[name = tensor<string, []>("op_3271_cast_fp16")];
            tensor<fp16, []> core_stacks_3_stack_layers_0_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.6acp-2)];
            tensor<fp16, [1, 512, 1, 512]> src_39_cast_fp16 = mul(x = var_3271_cast_fp16, y = core_stacks_3_stack_layers_0_norm_scale_to_fp16)[name = tensor<string, []>("src_39_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3274_cast_fp16 = sub(x = src_39_cast_fp16, y = input_267_cast_fp16)[name = tensor<string, []>("op_3274_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222840448)))];
            tensor<fp16, [1, 512, 1, 512]> var_3275_cast_fp16 = mul(x = var_3274_cast_fp16, y = core_stacks_3_stack_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_3275_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_297_cast_fp16 = add(x = input_267_cast_fp16, y = var_3275_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
            tensor<string, []> proj_19_pad_type_0 = const()[name = tensor<string, []>("proj_19_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_19_strides_0 = const()[name = tensor<string, []>("proj_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_19_pad_0 = const()[name = tensor<string, []>("proj_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_19_dilations_0 = const()[name = tensor<string, []>("proj_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_19_groups_0 = const()[name = tensor<string, []>("proj_19_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_3_stack_layers_1_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222841536)))];
            tensor<fp16, [384]> core_stacks_3_stack_layers_1_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223234816)))];
            tensor<fp16, [1, 384, 1, 512]> proj_19_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_19_dilations_0, groups = proj_19_groups_0, pad = proj_19_pad_0, pad_type = proj_19_pad_type_0, strides = proj_19_strides_0, weight = core_stacks_3_stack_layers_1_self_attn_weights_in_proj_weight_to_fp16, x = input_297_cast_fp16)[name = tensor<string, []>("proj_19_cast_fp16")];
            tensor<int32, [4]> var_3298_begin_0 = const()[name = tensor<string, []>("op_3298_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3298_end_0 = const()[name = tensor<string, []>("op_3298_end_0"), val = tensor<int32, [4]>([1, 128, 1, 512])];
            tensor<bool, [4]> var_3298_end_mask_0 = const()[name = tensor<string, []>("op_3298_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_3298_cast_fp16 = slice_by_index(begin = var_3298_begin_0, end = var_3298_end_0, end_mask = var_3298_end_mask_0, x = proj_19_cast_fp16)[name = tensor<string, []>("op_3298_cast_fp16")];
            tensor<int32, [4]> var_3299 = const()[name = tensor<string, []>("op_3299"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_3300_cast_fp16 = reshape(shape = var_3299, x = var_3298_cast_fp16)[name = tensor<string, []>("op_3300_cast_fp16")];
            tensor<int32, [4]> var_3302_begin_0 = const()[name = tensor<string, []>("op_3302_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_3302_end_0 = const()[name = tensor<string, []>("op_3302_end_0"), val = tensor<int32, [4]>([1, 256, 1, 512])];
            tensor<bool, [4]> var_3302_end_mask_0 = const()[name = tensor<string, []>("op_3302_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_3302_cast_fp16 = slice_by_index(begin = var_3302_begin_0, end = var_3302_end_0, end_mask = var_3302_end_mask_0, x = proj_19_cast_fp16)[name = tensor<string, []>("op_3302_cast_fp16")];
            tensor<int32, [4]> var_3303 = const()[name = tensor<string, []>("op_3303"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_3304_cast_fp16 = reshape(shape = var_3303, x = var_3302_cast_fp16)[name = tensor<string, []>("op_3304_cast_fp16")];
            tensor<int32, [4]> var_3306_begin_0 = const()[name = tensor<string, []>("op_3306_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_3306_end_0 = const()[name = tensor<string, []>("op_3306_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_3306_end_mask_0 = const()[name = tensor<string, []>("op_3306_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_3306_cast_fp16 = slice_by_index(begin = var_3306_begin_0, end = var_3306_end_0, end_mask = var_3306_end_mask_0, x = proj_19_cast_fp16)[name = tensor<string, []>("op_3306_cast_fp16")];
            tensor<int32, [4]> var_3307 = const()[name = tensor<string, []>("op_3307"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> pt_19_cast_fp16 = reshape(shape = var_3307, x = var_3306_cast_fp16)[name = tensor<string, []>("pt_19_cast_fp16")];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = tensor<string, []>("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_364 = const()[name = tensor<string, []>("concat_364"), val = tensor<int32, [3]>([4, 512, 32])];
            tensor<fp16, [4, 1, 512, 32]> transpose_72_cast_fp16 = transpose(perm = transpose_72_perm_0, x = var_3300_cast_fp16)[name = tensor<string, []>("transpose_212")];
            tensor<fp16, [4, 512, 32]> reshape_108_cast_fp16 = reshape(shape = concat_364, x = transpose_72_cast_fp16)[name = tensor<string, []>("reshape_108_cast_fp16")];
            tensor<int32, [4]> transpose_73_perm_0 = const()[name = tensor<string, []>("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_365 = const()[name = tensor<string, []>("concat_365"), val = tensor<int32, [3]>([4, 32, 512])];
            tensor<fp16, [4, 1, 32, 512]> transpose_73_cast_fp16 = transpose(perm = transpose_73_perm_0, x = var_3304_cast_fp16)[name = tensor<string, []>("transpose_211")];
            tensor<fp16, [4, 32, 512]> reshape_109_cast_fp16 = reshape(shape = concat_365, x = transpose_73_cast_fp16)[name = tensor<string, []>("reshape_109_cast_fp16")];
            tensor<bool, []> matmul_36_transpose_x_0 = const()[name = tensor<string, []>("matmul_36_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_36_transpose_y_0 = const()[name = tensor<string, []>("matmul_36_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 512]> matmul_36_cast_fp16 = matmul(transpose_x = matmul_36_transpose_x_0, transpose_y = matmul_36_transpose_y_0, x = reshape_108_cast_fp16, y = reshape_109_cast_fp16)[name = tensor<string, []>("matmul_36_cast_fp16")];
            tensor<int32, [4]> concat_369 = const()[name = tensor<string, []>("concat_369"), val = tensor<int32, [4]>([4, 1, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> reshape_110_cast_fp16 = reshape(shape = concat_369, x = matmul_36_cast_fp16)[name = tensor<string, []>("reshape_110_cast_fp16")];
            tensor<int32, [4]> attn_scores_55_perm_0 = const()[name = tensor<string, []>("attn_scores_55_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3311 = const()[name = tensor<string, []>("op_3311"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 512, 1]> var_3312_cast_fp16 = transpose(perm = var_3311, x = pt_19_cast_fp16)[name = tensor<string, []>("transpose_209")];
            tensor<fp16, [4, 32, 512, 512]> var_3313_cast_fp16 = mul(x = var_3312_cast_fp16, y = core_pos_basis_512_to_fp16)[name = tensor<string, []>("op_3313_cast_fp16")];
            tensor<int32, [1]> pos_scores_19_axes_0 = const()[name = tensor<string, []>("pos_scores_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_19_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_19_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 512, 512]> pos_scores_19_cast_fp16 = reduce_sum(axes = pos_scores_19_axes_0, keep_dims = pos_scores_19_keep_dims_0, x = var_3313_cast_fp16)[name = tensor<string, []>("pos_scores_19_cast_fp16")];
            tensor<int32, [4]> var_3316 = const()[name = tensor<string, []>("op_3316"), val = tensor<int32, [4]>([4, 512, 1, 512])];
            tensor<fp16, [4, 512, 1, 512]> var_3317_cast_fp16 = reshape(shape = var_3316, x = pos_scores_19_cast_fp16)[name = tensor<string, []>("op_3317_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_55_cast_fp16 = transpose(perm = attn_scores_55_perm_0, x = reshape_110_cast_fp16)[name = tensor<string, []>("transpose_210")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_57_cast_fp16 = add(x = attn_scores_55_cast_fp16, y = var_3317_cast_fp16)[name = tensor<string, []>("attn_scores_57_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_59_cast_fp16 = add(x = attn_scores_57_cast_fp16, y = attn_bias_3_cast_fp16)[name = tensor<string, []>("attn_scores_59_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_weights_19_cast_fp16 = softmax(axis = var_41, x = attn_scores_59_cast_fp16)[name = tensor<string, []>("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_299_cast_fp16 = add(x = input_297_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("input_299_cast_fp16")];
            tensor<string, []> x_413_pad_type_0 = const()[name = tensor<string, []>("x_413_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_413_strides_0 = const()[name = tensor<string, []>("x_413_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_413_pad_0 = const()[name = tensor<string, []>("x_413_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_413_dilations_0 = const()[name = tensor<string, []>("x_413_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_413_groups_0 = const()[name = tensor<string, []>("x_413_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_3_stack_layers_1_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223235648)))];
            tensor<fp16, [1152]> core_stacks_3_stack_layers_1_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224415360)))];
            tensor<fp16, [1, 1152, 1, 512]> x_413_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_feed_forward1_in_proj_bias_to_fp16, dilations = x_413_dilations_0, groups = x_413_groups_0, pad = x_413_pad_0, pad_type = x_413_pad_type_0, strides = x_413_strides_0, weight = core_stacks_3_stack_layers_1_feed_forward1_in_proj_weight_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("x_413_cast_fp16")];
            tensor<fp16, []> var_3331_to_fp16 = const()[name = tensor<string, []>("op_3331_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 512]> x_415_cast_fp16 = sub(x = x_413_cast_fp16, y = var_3331_to_fp16)[name = tensor<string, []>("x_415_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3333_cast_fp16 = relu(x = x_415_cast_fp16)[name = tensor<string, []>("op_3333_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3334_cast_fp16 = abs(x = x_415_cast_fp16)[name = tensor<string, []>("op_3334_cast_fp16")];
            tensor<fp16, []> const_72_promoted_to_fp16 = const()[name = tensor<string, []>("const_72_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_3335_cast_fp16 = mul(x = var_3334_cast_fp16, y = const_72_promoted_to_fp16)[name = tensor<string, []>("op_3335_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3336_cast_fp16 = exp(x = var_3335_cast_fp16)[name = tensor<string, []>("op_3336_cast_fp16")];
            tensor<fp32, []> var_3337_epsilon_0 = const()[name = tensor<string, []>("op_3337_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_3337_cast_fp16 = log(epsilon = var_3337_epsilon_0, x = var_3336_cast_fp16)[name = tensor<string, []>("op_3337_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3338_cast_fp16 = add(x = var_3333_cast_fp16, y = var_3337_cast_fp16)[name = tensor<string, []>("op_3338_cast_fp16")];
            tensor<fp16, []> var_3339_to_fp16 = const()[name = tensor<string, []>("op_3339_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 512]> var_3340_cast_fp16 = mul(x = x_413_cast_fp16, y = var_3339_to_fp16)[name = tensor<string, []>("op_3340_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3341_cast_fp16 = sub(x = var_3338_cast_fp16, y = var_3340_cast_fp16)[name = tensor<string, []>("op_3341_cast_fp16")];
            tensor<fp16, []> var_3342_to_fp16 = const()[name = tensor<string, []>("op_3342_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 512]> input_301_cast_fp16 = sub(x = var_3341_cast_fp16, y = var_3342_to_fp16)[name = tensor<string, []>("input_301_cast_fp16")];
            tensor<string, []> var_3350_pad_type_0 = const()[name = tensor<string, []>("op_3350_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3350_strides_0 = const()[name = tensor<string, []>("op_3350_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3350_pad_0 = const()[name = tensor<string, []>("op_3350_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3350_dilations_0 = const()[name = tensor<string, []>("op_3350_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3350_groups_0 = const()[name = tensor<string, []>("op_3350_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_3_stack_layers_1_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224417728)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_1_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225597440)))];
            tensor<fp16, [1, 512, 1, 512]> var_3350_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_feed_forward1_out_proj_bias_to_fp16, dilations = var_3350_dilations_0, groups = var_3350_groups_0, pad = var_3350_pad_0, pad_type = var_3350_pad_type_0, strides = var_3350_strides_0, weight = core_stacks_3_stack_layers_1_feed_forward1_out_proj_weight_to_fp16, x = input_301_cast_fp16)[name = tensor<string, []>("op_3350_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_303_cast_fp16 = add(x = input_299_cast_fp16, y = var_3350_cast_fp16)[name = tensor<string, []>("input_303_cast_fp16")];
            tensor<int32, [4]> var_3352_begin_0 = const()[name = tensor<string, []>("op_3352_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3352_end_0 = const()[name = tensor<string, []>("op_3352_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_3352_end_mask_0 = const()[name = tensor<string, []>("op_3352_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3352_cast_fp16 = slice_by_index(begin = var_3352_begin_0, end = var_3352_end_0, end_mask = var_3352_end_mask_0, x = attn_weights_19_cast_fp16)[name = tensor<string, []>("op_3352_cast_fp16")];
            tensor<string, []> x_417_pad_type_0 = const()[name = tensor<string, []>("x_417_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_417_strides_0 = const()[name = tensor<string, []>("x_417_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_417_pad_0 = const()[name = tensor<string, []>("x_417_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_417_dilations_0 = const()[name = tensor<string, []>("x_417_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_417_groups_0 = const()[name = tensor<string, []>("x_417_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_3_stack_layers_1_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225598528)))];
            tensor<fp16, [1152]> core_stacks_3_stack_layers_1_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226778240)))];
            tensor<fp16, [1, 1152, 1, 512]> x_417_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_nonlin_attention_in_proj_bias_to_fp16, dilations = x_417_dilations_0, groups = x_417_groups_0, pad = x_417_pad_0, pad_type = x_417_pad_type_0, strides = x_417_strides_0, weight = core_stacks_3_stack_layers_1_nonlin_attention_in_proj_weight_to_fp16, x = input_303_cast_fp16)[name = tensor<string, []>("x_417_cast_fp16")];
            tensor<int32, [4]> var_3363_begin_0 = const()[name = tensor<string, []>("op_3363_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3363_end_0 = const()[name = tensor<string, []>("op_3363_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_3363_end_mask_0 = const()[name = tensor<string, []>("op_3363_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_3363_cast_fp16 = slice_by_index(begin = var_3363_begin_0, end = var_3363_end_0, end_mask = var_3363_end_mask_0, x = x_417_cast_fp16)[name = tensor<string, []>("op_3363_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> s_gate_19_cast_fp16 = tanh(x = var_3363_cast_fp16)[name = tensor<string, []>("s_gate_19_cast_fp16")];
            tensor<int32, [4]> var_3366_begin_0 = const()[name = tensor<string, []>("op_3366_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_3366_end_0 = const()[name = tensor<string, []>("op_3366_end_0"), val = tensor<int32, [4]>([1, 768, 1, 512])];
            tensor<bool, [4]> var_3366_end_mask_0 = const()[name = tensor<string, []>("op_3366_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_3366_cast_fp16 = slice_by_index(begin = var_3366_begin_0, end = var_3366_end_0, end_mask = var_3366_end_mask_0, x = x_417_cast_fp16)[name = tensor<string, []>("op_3366_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> var_3367_cast_fp16 = mul(x = var_3366_cast_fp16, y = s_gate_19_cast_fp16)[name = tensor<string, []>("op_3367_cast_fp16")];
            tensor<int32, [4]> y_19_begin_0 = const()[name = tensor<string, []>("y_19_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_19_end_0 = const()[name = tensor<string, []>("y_19_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 512])];
            tensor<bool, [4]> y_19_end_mask_0 = const()[name = tensor<string, []>("y_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 512]> y_19_cast_fp16 = slice_by_index(begin = y_19_begin_0, end = y_19_end_0, end_mask = y_19_end_mask_0, x = x_417_cast_fp16)[name = tensor<string, []>("y_19_cast_fp16")];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = tensor<string, []>("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_374 = const()[name = tensor<string, []>("concat_374"), val = tensor<int32, [3]>([1, 512, 512])];
            tensor<fp16, [1, 1, 512, 512]> transpose_74_cast_fp16 = transpose(perm = transpose_74_perm_0, x = var_3352_cast_fp16)[name = tensor<string, []>("transpose_208")];
            tensor<fp16, [1, 512, 512]> reshape_111_cast_fp16 = reshape(shape = concat_374, x = transpose_74_cast_fp16)[name = tensor<string, []>("reshape_111_cast_fp16")];
            tensor<int32, [4]> transpose_75_perm_0 = const()[name = tensor<string, []>("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_375 = const()[name = tensor<string, []>("concat_375"), val = tensor<int32, [3]>([1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> transpose_75_cast_fp16 = transpose(perm = transpose_75_perm_0, x = var_3367_cast_fp16)[name = tensor<string, []>("transpose_207")];
            tensor<fp16, [1, 512, 384]> reshape_112_cast_fp16 = reshape(shape = concat_375, x = transpose_75_cast_fp16)[name = tensor<string, []>("reshape_112_cast_fp16")];
            tensor<bool, []> matmul_37_transpose_x_0 = const()[name = tensor<string, []>("matmul_37_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_37_transpose_y_0 = const()[name = tensor<string, []>("matmul_37_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 384]> matmul_37_cast_fp16 = matmul(transpose_x = matmul_37_transpose_x_0, transpose_y = matmul_37_transpose_y_0, x = reshape_111_cast_fp16, y = reshape_112_cast_fp16)[name = tensor<string, []>("matmul_37_cast_fp16")];
            tensor<int32, [4]> concat_379 = const()[name = tensor<string, []>("concat_379"), val = tensor<int32, [4]>([1, 1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> reshape_113_cast_fp16 = reshape(shape = concat_379, x = matmul_37_cast_fp16)[name = tensor<string, []>("reshape_113_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = tensor<string, []>("v_19_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 512]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = reshape_113_cast_fp16)[name = tensor<string, []>("transpose_206")];
            tensor<fp16, [1, 384, 1, 512]> input_305_cast_fp16 = mul(x = v_19_cast_fp16, y = y_19_cast_fp16)[name = tensor<string, []>("input_305_cast_fp16")];
            tensor<string, []> var_3379_pad_type_0 = const()[name = tensor<string, []>("op_3379_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3379_strides_0 = const()[name = tensor<string, []>("op_3379_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3379_pad_0 = const()[name = tensor<string, []>("op_3379_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3379_dilations_0 = const()[name = tensor<string, []>("op_3379_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3379_groups_0 = const()[name = tensor<string, []>("op_3379_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_3_stack_layers_1_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226780608)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_1_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227173888)))];
            tensor<fp16, [1, 512, 1, 512]> var_3379_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_nonlin_attention_out_proj_bias_to_fp16, dilations = var_3379_dilations_0, groups = var_3379_groups_0, pad = var_3379_pad_0, pad_type = var_3379_pad_type_0, strides = var_3379_strides_0, weight = core_stacks_3_stack_layers_1_nonlin_attention_out_proj_weight_to_fp16, x = input_305_cast_fp16)[name = tensor<string, []>("op_3379_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_419_cast_fp16 = add(x = input_303_cast_fp16, y = var_3379_cast_fp16)[name = tensor<string, []>("x_419_cast_fp16")];
            tensor<string, []> var_3390_pad_type_0 = const()[name = tensor<string, []>("op_3390_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3390_strides_0 = const()[name = tensor<string, []>("op_3390_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3390_pad_0 = const()[name = tensor<string, []>("op_3390_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3390_dilations_0 = const()[name = tensor<string, []>("op_3390_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3390_groups_0 = const()[name = tensor<string, []>("op_3390_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_3_stack_layers_1_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227174976)))];
            tensor<fp16, [48]> core_stacks_3_stack_layers_1_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227224192)))];
            tensor<fp16, [1, 48, 1, 512]> var_3390_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_self_attn1_in_proj_bias_to_fp16, dilations = var_3390_dilations_0, groups = var_3390_groups_0, pad = var_3390_pad_0, pad_type = var_3390_pad_type_0, strides = var_3390_strides_0, weight = core_stacks_3_stack_layers_1_self_attn1_in_proj_weight_to_fp16, x = x_419_cast_fp16)[name = tensor<string, []>("op_3390_cast_fp16")];
            tensor<int32, [4]> var_3391 = const()[name = tensor<string, []>("op_3391"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_3392_cast_fp16 = reshape(shape = var_3391, x = var_3390_cast_fp16)[name = tensor<string, []>("op_3392_cast_fp16")];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = tensor<string, []>("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_384 = const()[name = tensor<string, []>("concat_384"), val = tensor<int32, [3]>([4, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> transpose_76_cast_fp16 = transpose(perm = transpose_76_perm_0, x = attn_weights_19_cast_fp16)[name = tensor<string, []>("transpose_205")];
            tensor<fp16, [4, 512, 512]> reshape_114_cast_fp16 = reshape(shape = concat_384, x = transpose_76_cast_fp16)[name = tensor<string, []>("reshape_114_cast_fp16")];
            tensor<int32, [4]> transpose_77_perm_0 = const()[name = tensor<string, []>("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_385 = const()[name = tensor<string, []>("concat_385"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_77_cast_fp16 = transpose(perm = transpose_77_perm_0, x = var_3392_cast_fp16)[name = tensor<string, []>("transpose_204")];
            tensor<fp16, [4, 512, 12]> reshape_115_cast_fp16 = reshape(shape = concat_385, x = transpose_77_cast_fp16)[name = tensor<string, []>("reshape_115_cast_fp16")];
            tensor<bool, []> matmul_38_transpose_x_0 = const()[name = tensor<string, []>("matmul_38_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_38_transpose_y_0 = const()[name = tensor<string, []>("matmul_38_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_38_cast_fp16 = matmul(transpose_x = matmul_38_transpose_x_0, transpose_y = matmul_38_transpose_y_0, x = reshape_114_cast_fp16, y = reshape_115_cast_fp16)[name = tensor<string, []>("matmul_38_cast_fp16")];
            tensor<int32, [4]> concat_389 = const()[name = tensor<string, []>("concat_389"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_116_cast_fp16 = reshape(shape = concat_389, x = matmul_38_cast_fp16)[name = tensor<string, []>("reshape_116_cast_fp16")];
            tensor<int32, [4]> out_37_perm_0 = const()[name = tensor<string, []>("out_37_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_3395 = const()[name = tensor<string, []>("op_3395"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_37_cast_fp16 = transpose(perm = out_37_perm_0, x = reshape_116_cast_fp16)[name = tensor<string, []>("transpose_203")];
            tensor<fp16, [1, 48, 1, 512]> input_307_cast_fp16 = reshape(shape = var_3395, x = out_37_cast_fp16)[name = tensor<string, []>("input_307_cast_fp16")];
            tensor<string, []> var_3403_pad_type_0 = const()[name = tensor<string, []>("op_3403_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3403_strides_0 = const()[name = tensor<string, []>("op_3403_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3403_pad_0 = const()[name = tensor<string, []>("op_3403_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3403_dilations_0 = const()[name = tensor<string, []>("op_3403_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3403_groups_0 = const()[name = tensor<string, []>("op_3403_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_3_stack_layers_1_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227224384)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_1_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227273600)))];
            tensor<fp16, [1, 512, 1, 512]> var_3403_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_self_attn1_out_proj_bias_to_fp16, dilations = var_3403_dilations_0, groups = var_3403_groups_0, pad = var_3403_pad_0, pad_type = var_3403_pad_type_0, strides = var_3403_strides_0, weight = core_stacks_3_stack_layers_1_self_attn1_out_proj_weight_to_fp16, x = input_307_cast_fp16)[name = tensor<string, []>("op_3403_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_421_cast_fp16 = add(x = x_419_cast_fp16, y = var_3403_cast_fp16)[name = tensor<string, []>("x_421_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_423_cast_fp16 = add(x = x_421_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("x_423_cast_fp16")];
            tensor<string, []> x_425_pad_type_0 = const()[name = tensor<string, []>("x_425_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_425_strides_0 = const()[name = tensor<string, []>("x_425_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_425_pad_0 = const()[name = tensor<string, []>("x_425_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_425_dilations_0 = const()[name = tensor<string, []>("x_425_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_425_groups_0 = const()[name = tensor<string, []>("x_425_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_3_stack_layers_1_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227274688)))];
            tensor<fp16, [1024]> core_stacks_3_stack_layers_1_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228323328)))];
            tensor<fp16, [1, 1024, 1, 512]> x_425_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_conv_module1_in_proj_bias_to_fp16, dilations = x_425_dilations_0, groups = x_425_groups_0, pad = x_425_pad_0, pad_type = x_425_pad_type_0, strides = x_425_strides_0, weight = core_stacks_3_stack_layers_1_conv_module1_in_proj_weight_to_fp16, x = x_423_cast_fp16)[name = tensor<string, []>("x_425_cast_fp16")];
            tensor<int32, [4]> var_3417_begin_0 = const()[name = tensor<string, []>("op_3417_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3417_end_0 = const()[name = tensor<string, []>("op_3417_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_3417_end_mask_0 = const()[name = tensor<string, []>("op_3417_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3417_cast_fp16 = slice_by_index(begin = var_3417_begin_0, end = var_3417_end_0, end_mask = var_3417_end_mask_0, x = x_425_cast_fp16)[name = tensor<string, []>("op_3417_cast_fp16")];
            tensor<int32, [4]> var_3419_begin_0 = const()[name = tensor<string, []>("op_3419_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_3419_end_0 = const()[name = tensor<string, []>("op_3419_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_3419_end_mask_0 = const()[name = tensor<string, []>("op_3419_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3419_cast_fp16 = slice_by_index(begin = var_3419_begin_0, end = var_3419_end_0, end_mask = var_3419_end_mask_0, x = x_425_cast_fp16)[name = tensor<string, []>("op_3419_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3420_cast_fp16 = sigmoid(x = var_3419_cast_fp16)[name = tensor<string, []>("op_3420_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_427_cast_fp16 = mul(x = var_3417_cast_fp16, y = var_3420_cast_fp16)[name = tensor<string, []>("x_427_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_309_cast_fp16 = mul(x = x_427_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_309_cast_fp16")];
            tensor<string, []> x_429_pad_type_0 = const()[name = tensor<string, []>("x_429_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_429_pad_0 = const()[name = tensor<string, []>("x_429_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_429_groups_0 = const()[name = tensor<string, []>("x_429_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_429_strides_0 = const()[name = tensor<string, []>("x_429_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_429_dilations_0 = const()[name = tensor<string, []>("x_429_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_3_stack_layers_1_conv_module1_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module1_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228325440)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_1_conv_module1_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module1_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228340864)))];
            tensor<fp16, [1, 512, 1, 512]> x_429_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_conv_module1_dw_bias_to_fp16, dilations = x_429_dilations_0, groups = x_429_groups_0, pad = x_429_pad_0, pad_type = x_429_pad_type_0, strides = x_429_strides_0, weight = core_stacks_3_stack_layers_1_conv_module1_dw_weight_to_fp16, x = input_309_cast_fp16)[name = tensor<string, []>("x_429_cast_fp16")];
            tensor<fp16, []> var_3430_to_fp16 = const()[name = tensor<string, []>("op_3430_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_431_cast_fp16 = sub(x = x_429_cast_fp16, y = var_3430_to_fp16)[name = tensor<string, []>("x_431_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3432_cast_fp16 = relu(x = x_431_cast_fp16)[name = tensor<string, []>("op_3432_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3433_cast_fp16 = abs(x = x_431_cast_fp16)[name = tensor<string, []>("op_3433_cast_fp16")];
            tensor<fp16, []> const_74_promoted_to_fp16 = const()[name = tensor<string, []>("const_74_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3434_cast_fp16 = mul(x = var_3433_cast_fp16, y = const_74_promoted_to_fp16)[name = tensor<string, []>("op_3434_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3435_cast_fp16 = exp(x = var_3434_cast_fp16)[name = tensor<string, []>("op_3435_cast_fp16")];
            tensor<fp32, []> var_3436_epsilon_0 = const()[name = tensor<string, []>("op_3436_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3436_cast_fp16 = log(epsilon = var_3436_epsilon_0, x = var_3435_cast_fp16)[name = tensor<string, []>("op_3436_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3437_cast_fp16 = add(x = var_3432_cast_fp16, y = var_3436_cast_fp16)[name = tensor<string, []>("op_3437_cast_fp16")];
            tensor<fp16, []> var_3438_to_fp16 = const()[name = tensor<string, []>("op_3438_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_3439_cast_fp16 = mul(x = x_429_cast_fp16, y = var_3438_to_fp16)[name = tensor<string, []>("op_3439_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3440_cast_fp16 = sub(x = var_3437_cast_fp16, y = var_3439_cast_fp16)[name = tensor<string, []>("op_3440_cast_fp16")];
            tensor<fp16, []> var_3441_to_fp16 = const()[name = tensor<string, []>("op_3441_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_311_cast_fp16 = sub(x = var_3440_cast_fp16, y = var_3441_to_fp16)[name = tensor<string, []>("input_311_cast_fp16")];
            tensor<string, []> var_3449_pad_type_0 = const()[name = tensor<string, []>("op_3449_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3449_strides_0 = const()[name = tensor<string, []>("op_3449_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3449_pad_0 = const()[name = tensor<string, []>("op_3449_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3449_dilations_0 = const()[name = tensor<string, []>("op_3449_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3449_groups_0 = const()[name = tensor<string, []>("op_3449_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_3_stack_layers_1_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228341952)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_1_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228866304)))];
            tensor<fp16, [1, 512, 1, 512]> var_3449_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_conv_module1_out_proj_bias_to_fp16, dilations = var_3449_dilations_0, groups = var_3449_groups_0, pad = var_3449_pad_0, pad_type = var_3449_pad_type_0, strides = var_3449_strides_0, weight = core_stacks_3_stack_layers_1_conv_module1_out_proj_weight_to_fp16, x = input_311_cast_fp16)[name = tensor<string, []>("op_3449_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_313_cast_fp16 = add(x = x_423_cast_fp16, y = var_3449_cast_fp16)[name = tensor<string, []>("input_313_cast_fp16")];
            tensor<string, []> x_433_pad_type_0 = const()[name = tensor<string, []>("x_433_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_433_strides_0 = const()[name = tensor<string, []>("x_433_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_433_pad_0 = const()[name = tensor<string, []>("x_433_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_433_dilations_0 = const()[name = tensor<string, []>("x_433_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_433_groups_0 = const()[name = tensor<string, []>("x_433_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_3_stack_layers_1_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228867392)))];
            tensor<fp16, [1536]> core_stacks_3_stack_layers_1_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(230440320)))];
            tensor<fp16, [1, 1536, 1, 512]> x_433_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_feed_forward2_in_proj_bias_to_fp16, dilations = x_433_dilations_0, groups = x_433_groups_0, pad = x_433_pad_0, pad_type = x_433_pad_type_0, strides = x_433_strides_0, weight = core_stacks_3_stack_layers_1_feed_forward2_in_proj_weight_to_fp16, x = input_313_cast_fp16)[name = tensor<string, []>("x_433_cast_fp16")];
            tensor<fp16, []> var_3460_to_fp16 = const()[name = tensor<string, []>("op_3460_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 512]> x_435_cast_fp16 = sub(x = x_433_cast_fp16, y = var_3460_to_fp16)[name = tensor<string, []>("x_435_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3462_cast_fp16 = relu(x = x_435_cast_fp16)[name = tensor<string, []>("op_3462_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3463_cast_fp16 = abs(x = x_435_cast_fp16)[name = tensor<string, []>("op_3463_cast_fp16")];
            tensor<fp16, []> const_75_promoted_to_fp16 = const()[name = tensor<string, []>("const_75_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_3464_cast_fp16 = mul(x = var_3463_cast_fp16, y = const_75_promoted_to_fp16)[name = tensor<string, []>("op_3464_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3465_cast_fp16 = exp(x = var_3464_cast_fp16)[name = tensor<string, []>("op_3465_cast_fp16")];
            tensor<fp32, []> var_3466_epsilon_0 = const()[name = tensor<string, []>("op_3466_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_3466_cast_fp16 = log(epsilon = var_3466_epsilon_0, x = var_3465_cast_fp16)[name = tensor<string, []>("op_3466_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3467_cast_fp16 = add(x = var_3462_cast_fp16, y = var_3466_cast_fp16)[name = tensor<string, []>("op_3467_cast_fp16")];
            tensor<fp16, []> var_3468_to_fp16 = const()[name = tensor<string, []>("op_3468_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 512]> var_3469_cast_fp16 = mul(x = x_433_cast_fp16, y = var_3468_to_fp16)[name = tensor<string, []>("op_3469_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3470_cast_fp16 = sub(x = var_3467_cast_fp16, y = var_3469_cast_fp16)[name = tensor<string, []>("op_3470_cast_fp16")];
            tensor<fp16, []> var_3471_to_fp16 = const()[name = tensor<string, []>("op_3471_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 512]> input_315_cast_fp16 = sub(x = var_3470_cast_fp16, y = var_3471_to_fp16)[name = tensor<string, []>("input_315_cast_fp16")];
            tensor<string, []> var_3479_pad_type_0 = const()[name = tensor<string, []>("op_3479_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3479_strides_0 = const()[name = tensor<string, []>("op_3479_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3479_pad_0 = const()[name = tensor<string, []>("op_3479_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3479_dilations_0 = const()[name = tensor<string, []>("op_3479_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3479_groups_0 = const()[name = tensor<string, []>("op_3479_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_3_stack_layers_1_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(230443456)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_1_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232016384)))];
            tensor<fp16, [1, 512, 1, 512]> var_3479_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_feed_forward2_out_proj_bias_to_fp16, dilations = var_3479_dilations_0, groups = var_3479_groups_0, pad = var_3479_pad_0, pad_type = var_3479_pad_type_0, strides = var_3479_strides_0, weight = core_stacks_3_stack_layers_1_feed_forward2_out_proj_weight_to_fp16, x = input_315_cast_fp16)[name = tensor<string, []>("op_3479_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> src_41_cast_fp16 = add(x = input_313_cast_fp16, y = var_3479_cast_fp16)[name = tensor<string, []>("src_41_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3482_cast_fp16 = sub(x = src_41_cast_fp16, y = input_297_cast_fp16)[name = tensor<string, []>("op_3482_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_1_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232017472)))];
            tensor<fp16, [1, 512, 1, 512]> var_3483_cast_fp16 = mul(x = var_3482_cast_fp16, y = core_stacks_3_stack_layers_1_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_3483_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_437_cast_fp16 = add(x = input_297_cast_fp16, y = var_3483_cast_fp16)[name = tensor<string, []>("x_437_cast_fp16")];
            tensor<string, []> var_3494_pad_type_0 = const()[name = tensor<string, []>("op_3494_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3494_strides_0 = const()[name = tensor<string, []>("op_3494_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3494_pad_0 = const()[name = tensor<string, []>("op_3494_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3494_dilations_0 = const()[name = tensor<string, []>("op_3494_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3494_groups_0 = const()[name = tensor<string, []>("op_3494_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_3_stack_layers_1_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232018560)))];
            tensor<fp16, [48]> core_stacks_3_stack_layers_1_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232067776)))];
            tensor<fp16, [1, 48, 1, 512]> var_3494_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_self_attn2_in_proj_bias_to_fp16, dilations = var_3494_dilations_0, groups = var_3494_groups_0, pad = var_3494_pad_0, pad_type = var_3494_pad_type_0, strides = var_3494_strides_0, weight = core_stacks_3_stack_layers_1_self_attn2_in_proj_weight_to_fp16, x = x_437_cast_fp16)[name = tensor<string, []>("op_3494_cast_fp16")];
            tensor<int32, [4]> var_3495 = const()[name = tensor<string, []>("op_3495"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_3496_cast_fp16 = reshape(shape = var_3495, x = var_3494_cast_fp16)[name = tensor<string, []>("op_3496_cast_fp16")];
            tensor<int32, [4]> transpose_79_perm_0 = const()[name = tensor<string, []>("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_395 = const()[name = tensor<string, []>("concat_395"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_79_cast_fp16 = transpose(perm = transpose_79_perm_0, x = var_3496_cast_fp16)[name = tensor<string, []>("transpose_202")];
            tensor<fp16, [4, 512, 12]> reshape_118_cast_fp16 = reshape(shape = concat_395, x = transpose_79_cast_fp16)[name = tensor<string, []>("reshape_118_cast_fp16")];
            tensor<bool, []> matmul_39_transpose_x_0 = const()[name = tensor<string, []>("matmul_39_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_39_transpose_y_0 = const()[name = tensor<string, []>("matmul_39_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_39_cast_fp16 = matmul(transpose_x = matmul_39_transpose_x_0, transpose_y = matmul_39_transpose_y_0, x = reshape_114_cast_fp16, y = reshape_118_cast_fp16)[name = tensor<string, []>("matmul_39_cast_fp16")];
            tensor<int32, [4]> concat_399 = const()[name = tensor<string, []>("concat_399"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_119_cast_fp16 = reshape(shape = concat_399, x = matmul_39_cast_fp16)[name = tensor<string, []>("reshape_119_cast_fp16")];
            tensor<int32, [4]> out_39_perm_0 = const()[name = tensor<string, []>("out_39_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_3499 = const()[name = tensor<string, []>("op_3499"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_39_cast_fp16 = transpose(perm = out_39_perm_0, x = reshape_119_cast_fp16)[name = tensor<string, []>("transpose_201")];
            tensor<fp16, [1, 48, 1, 512]> input_317_cast_fp16 = reshape(shape = var_3499, x = out_39_cast_fp16)[name = tensor<string, []>("input_317_cast_fp16")];
            tensor<string, []> var_3507_pad_type_0 = const()[name = tensor<string, []>("op_3507_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3507_strides_0 = const()[name = tensor<string, []>("op_3507_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3507_pad_0 = const()[name = tensor<string, []>("op_3507_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3507_dilations_0 = const()[name = tensor<string, []>("op_3507_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3507_groups_0 = const()[name = tensor<string, []>("op_3507_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_3_stack_layers_1_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232067968)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_1_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232117184)))];
            tensor<fp16, [1, 512, 1, 512]> var_3507_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_self_attn2_out_proj_bias_to_fp16, dilations = var_3507_dilations_0, groups = var_3507_groups_0, pad = var_3507_pad_0, pad_type = var_3507_pad_type_0, strides = var_3507_strides_0, weight = core_stacks_3_stack_layers_1_self_attn2_out_proj_weight_to_fp16, x = input_317_cast_fp16)[name = tensor<string, []>("op_3507_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_439_cast_fp16 = add(x = x_437_cast_fp16, y = var_3507_cast_fp16)[name = tensor<string, []>("x_439_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_441_cast_fp16 = add(x = x_439_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("x_441_cast_fp16")];
            tensor<string, []> x_443_pad_type_0 = const()[name = tensor<string, []>("x_443_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_443_strides_0 = const()[name = tensor<string, []>("x_443_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_443_pad_0 = const()[name = tensor<string, []>("x_443_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_443_dilations_0 = const()[name = tensor<string, []>("x_443_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_443_groups_0 = const()[name = tensor<string, []>("x_443_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_3_stack_layers_1_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232118272)))];
            tensor<fp16, [1024]> core_stacks_3_stack_layers_1_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233166912)))];
            tensor<fp16, [1, 1024, 1, 512]> x_443_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_conv_module2_in_proj_bias_to_fp16, dilations = x_443_dilations_0, groups = x_443_groups_0, pad = x_443_pad_0, pad_type = x_443_pad_type_0, strides = x_443_strides_0, weight = core_stacks_3_stack_layers_1_conv_module2_in_proj_weight_to_fp16, x = x_441_cast_fp16)[name = tensor<string, []>("x_443_cast_fp16")];
            tensor<int32, [4]> var_3521_begin_0 = const()[name = tensor<string, []>("op_3521_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3521_end_0 = const()[name = tensor<string, []>("op_3521_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_3521_end_mask_0 = const()[name = tensor<string, []>("op_3521_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3521_cast_fp16 = slice_by_index(begin = var_3521_begin_0, end = var_3521_end_0, end_mask = var_3521_end_mask_0, x = x_443_cast_fp16)[name = tensor<string, []>("op_3521_cast_fp16")];
            tensor<int32, [4]> var_3523_begin_0 = const()[name = tensor<string, []>("op_3523_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_3523_end_0 = const()[name = tensor<string, []>("op_3523_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_3523_end_mask_0 = const()[name = tensor<string, []>("op_3523_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3523_cast_fp16 = slice_by_index(begin = var_3523_begin_0, end = var_3523_end_0, end_mask = var_3523_end_mask_0, x = x_443_cast_fp16)[name = tensor<string, []>("op_3523_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3524_cast_fp16 = sigmoid(x = var_3523_cast_fp16)[name = tensor<string, []>("op_3524_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_445_cast_fp16 = mul(x = var_3521_cast_fp16, y = var_3524_cast_fp16)[name = tensor<string, []>("x_445_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_319_cast_fp16 = mul(x = x_445_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
            tensor<string, []> x_447_pad_type_0 = const()[name = tensor<string, []>("x_447_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_447_pad_0 = const()[name = tensor<string, []>("x_447_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_447_groups_0 = const()[name = tensor<string, []>("x_447_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_447_strides_0 = const()[name = tensor<string, []>("x_447_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_447_dilations_0 = const()[name = tensor<string, []>("x_447_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_3_stack_layers_1_conv_module2_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module2_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233169024)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_1_conv_module2_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module2_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233184448)))];
            tensor<fp16, [1, 512, 1, 512]> x_447_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_conv_module2_dw_bias_to_fp16, dilations = x_447_dilations_0, groups = x_447_groups_0, pad = x_447_pad_0, pad_type = x_447_pad_type_0, strides = x_447_strides_0, weight = core_stacks_3_stack_layers_1_conv_module2_dw_weight_to_fp16, x = input_319_cast_fp16)[name = tensor<string, []>("x_447_cast_fp16")];
            tensor<fp16, []> var_3534_to_fp16 = const()[name = tensor<string, []>("op_3534_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_449_cast_fp16 = sub(x = x_447_cast_fp16, y = var_3534_to_fp16)[name = tensor<string, []>("x_449_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3536_cast_fp16 = relu(x = x_449_cast_fp16)[name = tensor<string, []>("op_3536_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3537_cast_fp16 = abs(x = x_449_cast_fp16)[name = tensor<string, []>("op_3537_cast_fp16")];
            tensor<fp16, []> const_77_promoted_to_fp16 = const()[name = tensor<string, []>("const_77_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3538_cast_fp16 = mul(x = var_3537_cast_fp16, y = const_77_promoted_to_fp16)[name = tensor<string, []>("op_3538_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3539_cast_fp16 = exp(x = var_3538_cast_fp16)[name = tensor<string, []>("op_3539_cast_fp16")];
            tensor<fp32, []> var_3540_epsilon_0 = const()[name = tensor<string, []>("op_3540_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3540_cast_fp16 = log(epsilon = var_3540_epsilon_0, x = var_3539_cast_fp16)[name = tensor<string, []>("op_3540_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3541_cast_fp16 = add(x = var_3536_cast_fp16, y = var_3540_cast_fp16)[name = tensor<string, []>("op_3541_cast_fp16")];
            tensor<fp16, []> var_3542_to_fp16 = const()[name = tensor<string, []>("op_3542_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_3543_cast_fp16 = mul(x = x_447_cast_fp16, y = var_3542_to_fp16)[name = tensor<string, []>("op_3543_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3544_cast_fp16 = sub(x = var_3541_cast_fp16, y = var_3543_cast_fp16)[name = tensor<string, []>("op_3544_cast_fp16")];
            tensor<fp16, []> var_3545_to_fp16 = const()[name = tensor<string, []>("op_3545_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_321_cast_fp16 = sub(x = var_3544_cast_fp16, y = var_3545_to_fp16)[name = tensor<string, []>("input_321_cast_fp16")];
            tensor<string, []> var_3553_pad_type_0 = const()[name = tensor<string, []>("op_3553_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3553_strides_0 = const()[name = tensor<string, []>("op_3553_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3553_pad_0 = const()[name = tensor<string, []>("op_3553_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3553_dilations_0 = const()[name = tensor<string, []>("op_3553_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3553_groups_0 = const()[name = tensor<string, []>("op_3553_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_3_stack_layers_1_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233185536)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_1_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233709888)))];
            tensor<fp16, [1, 512, 1, 512]> var_3553_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_conv_module2_out_proj_bias_to_fp16, dilations = var_3553_dilations_0, groups = var_3553_groups_0, pad = var_3553_pad_0, pad_type = var_3553_pad_type_0, strides = var_3553_strides_0, weight = core_stacks_3_stack_layers_1_conv_module2_out_proj_weight_to_fp16, x = input_321_cast_fp16)[name = tensor<string, []>("op_3553_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_323_cast_fp16 = add(x = x_441_cast_fp16, y = var_3553_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
            tensor<string, []> x_451_pad_type_0 = const()[name = tensor<string, []>("x_451_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_451_strides_0 = const()[name = tensor<string, []>("x_451_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_451_pad_0 = const()[name = tensor<string, []>("x_451_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_451_dilations_0 = const()[name = tensor<string, []>("x_451_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_451_groups_0 = const()[name = tensor<string, []>("x_451_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_3_stack_layers_1_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233710976)))];
            tensor<fp16, [1920]> core_stacks_3_stack_layers_1_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235677120)))];
            tensor<fp16, [1, 1920, 1, 512]> x_451_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_feed_forward3_in_proj_bias_to_fp16, dilations = x_451_dilations_0, groups = x_451_groups_0, pad = x_451_pad_0, pad_type = x_451_pad_type_0, strides = x_451_strides_0, weight = core_stacks_3_stack_layers_1_feed_forward3_in_proj_weight_to_fp16, x = input_323_cast_fp16)[name = tensor<string, []>("x_451_cast_fp16")];
            tensor<fp16, []> var_3564_to_fp16 = const()[name = tensor<string, []>("op_3564_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 512]> x_453_cast_fp16 = sub(x = x_451_cast_fp16, y = var_3564_to_fp16)[name = tensor<string, []>("x_453_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3566_cast_fp16 = relu(x = x_453_cast_fp16)[name = tensor<string, []>("op_3566_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3567_cast_fp16 = abs(x = x_453_cast_fp16)[name = tensor<string, []>("op_3567_cast_fp16")];
            tensor<fp16, []> const_78_promoted_to_fp16 = const()[name = tensor<string, []>("const_78_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_3568_cast_fp16 = mul(x = var_3567_cast_fp16, y = const_78_promoted_to_fp16)[name = tensor<string, []>("op_3568_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3569_cast_fp16 = exp(x = var_3568_cast_fp16)[name = tensor<string, []>("op_3569_cast_fp16")];
            tensor<fp32, []> var_3570_epsilon_0 = const()[name = tensor<string, []>("op_3570_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_3570_cast_fp16 = log(epsilon = var_3570_epsilon_0, x = var_3569_cast_fp16)[name = tensor<string, []>("op_3570_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3571_cast_fp16 = add(x = var_3566_cast_fp16, y = var_3570_cast_fp16)[name = tensor<string, []>("op_3571_cast_fp16")];
            tensor<fp16, []> var_3572_to_fp16 = const()[name = tensor<string, []>("op_3572_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 512]> var_3573_cast_fp16 = mul(x = x_451_cast_fp16, y = var_3572_to_fp16)[name = tensor<string, []>("op_3573_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3574_cast_fp16 = sub(x = var_3571_cast_fp16, y = var_3573_cast_fp16)[name = tensor<string, []>("op_3574_cast_fp16")];
            tensor<fp16, []> var_3575_to_fp16 = const()[name = tensor<string, []>("op_3575_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 512]> input_325_cast_fp16 = sub(x = var_3574_cast_fp16, y = var_3575_to_fp16)[name = tensor<string, []>("input_325_cast_fp16")];
            tensor<string, []> var_3583_pad_type_0 = const()[name = tensor<string, []>("op_3583_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3583_strides_0 = const()[name = tensor<string, []>("op_3583_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3583_pad_0 = const()[name = tensor<string, []>("op_3583_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3583_dilations_0 = const()[name = tensor<string, []>("op_3583_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3583_groups_0 = const()[name = tensor<string, []>("op_3583_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_3_stack_layers_1_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235681024)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_1_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237647168)))];
            tensor<fp16, [1, 512, 1, 512]> var_3583_cast_fp16 = conv(bias = core_stacks_3_stack_layers_1_feed_forward3_out_proj_bias_to_fp16, dilations = var_3583_dilations_0, groups = var_3583_groups_0, pad = var_3583_pad_0, pad_type = var_3583_pad_type_0, strides = var_3583_strides_0, weight = core_stacks_3_stack_layers_1_feed_forward3_out_proj_weight_to_fp16, x = input_325_cast_fp16)[name = tensor<string, []>("op_3583_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_455_cast_fp16 = add(x = input_323_cast_fp16, y = var_3583_cast_fp16)[name = tensor<string, []>("x_455_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_1_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237648256)))];
            tensor<fp16, [1, 512, 1, 512]> d_19_cast_fp16 = sub(x = x_455_cast_fp16, y = core_stacks_3_stack_layers_1_norm_bias_to_fp16)[name = tensor<string, []>("d_19_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3588_cast_fp16 = mul(x = d_19_cast_fp16, y = d_19_cast_fp16)[name = tensor<string, []>("op_3588_cast_fp16")];
            tensor<int32, [1]> var_3590_axes_0 = const()[name = tensor<string, []>("op_3590_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_3590_keep_dims_0 = const()[name = tensor<string, []>("op_3590_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 512]> var_3590_cast_fp16 = reduce_mean(axes = var_3590_axes_0, keep_dims = var_3590_keep_dims_0, x = var_3588_cast_fp16)[name = tensor<string, []>("op_3590_cast_fp16")];
            tensor<fp32, []> scales_19_epsilon_0 = const()[name = tensor<string, []>("scales_19_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 512]> scales_19_cast_fp16 = rsqrt(epsilon = scales_19_epsilon_0, x = var_3590_cast_fp16)[name = tensor<string, []>("scales_19_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3592_cast_fp16 = mul(x = x_455_cast_fp16, y = scales_19_cast_fp16)[name = tensor<string, []>("op_3592_cast_fp16")];
            tensor<fp16, []> core_stacks_3_stack_layers_1_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.fc4p-3)];
            tensor<fp16, [1, 512, 1, 512]> src_43_cast_fp16 = mul(x = var_3592_cast_fp16, y = core_stacks_3_stack_layers_1_norm_scale_to_fp16)[name = tensor<string, []>("src_43_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3595_cast_fp16 = sub(x = src_43_cast_fp16, y = input_297_cast_fp16)[name = tensor<string, []>("op_3595_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_1_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_1_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237649344)))];
            tensor<fp16, [1, 512, 1, 512]> var_3596_cast_fp16 = mul(x = var_3595_cast_fp16, y = core_stacks_3_stack_layers_1_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_3596_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_327_cast_fp16 = add(x = input_297_cast_fp16, y = var_3596_cast_fp16)[name = tensor<string, []>("input_327_cast_fp16")];
            tensor<string, []> proj_21_pad_type_0 = const()[name = tensor<string, []>("proj_21_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_21_strides_0 = const()[name = tensor<string, []>("proj_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_21_pad_0 = const()[name = tensor<string, []>("proj_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_21_dilations_0 = const()[name = tensor<string, []>("proj_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_21_groups_0 = const()[name = tensor<string, []>("proj_21_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_3_stack_layers_2_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237650432)))];
            tensor<fp16, [384]> core_stacks_3_stack_layers_2_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238043712)))];
            tensor<fp16, [1, 384, 1, 512]> proj_21_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_21_dilations_0, groups = proj_21_groups_0, pad = proj_21_pad_0, pad_type = proj_21_pad_type_0, strides = proj_21_strides_0, weight = core_stacks_3_stack_layers_2_self_attn_weights_in_proj_weight_to_fp16, x = input_327_cast_fp16)[name = tensor<string, []>("proj_21_cast_fp16")];
            tensor<int32, [4]> var_3619_begin_0 = const()[name = tensor<string, []>("op_3619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3619_end_0 = const()[name = tensor<string, []>("op_3619_end_0"), val = tensor<int32, [4]>([1, 128, 1, 512])];
            tensor<bool, [4]> var_3619_end_mask_0 = const()[name = tensor<string, []>("op_3619_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_3619_cast_fp16 = slice_by_index(begin = var_3619_begin_0, end = var_3619_end_0, end_mask = var_3619_end_mask_0, x = proj_21_cast_fp16)[name = tensor<string, []>("op_3619_cast_fp16")];
            tensor<int32, [4]> var_3620 = const()[name = tensor<string, []>("op_3620"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_3621_cast_fp16 = reshape(shape = var_3620, x = var_3619_cast_fp16)[name = tensor<string, []>("op_3621_cast_fp16")];
            tensor<int32, [4]> var_3623_begin_0 = const()[name = tensor<string, []>("op_3623_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_3623_end_0 = const()[name = tensor<string, []>("op_3623_end_0"), val = tensor<int32, [4]>([1, 256, 1, 512])];
            tensor<bool, [4]> var_3623_end_mask_0 = const()[name = tensor<string, []>("op_3623_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_3623_cast_fp16 = slice_by_index(begin = var_3623_begin_0, end = var_3623_end_0, end_mask = var_3623_end_mask_0, x = proj_21_cast_fp16)[name = tensor<string, []>("op_3623_cast_fp16")];
            tensor<int32, [4]> var_3624 = const()[name = tensor<string, []>("op_3624"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_3625_cast_fp16 = reshape(shape = var_3624, x = var_3623_cast_fp16)[name = tensor<string, []>("op_3625_cast_fp16")];
            tensor<int32, [4]> var_3627_begin_0 = const()[name = tensor<string, []>("op_3627_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_3627_end_0 = const()[name = tensor<string, []>("op_3627_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_3627_end_mask_0 = const()[name = tensor<string, []>("op_3627_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_3627_cast_fp16 = slice_by_index(begin = var_3627_begin_0, end = var_3627_end_0, end_mask = var_3627_end_mask_0, x = proj_21_cast_fp16)[name = tensor<string, []>("op_3627_cast_fp16")];
            tensor<int32, [4]> var_3628 = const()[name = tensor<string, []>("op_3628"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> pt_21_cast_fp16 = reshape(shape = var_3628, x = var_3627_cast_fp16)[name = tensor<string, []>("pt_21_cast_fp16")];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = tensor<string, []>("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_404 = const()[name = tensor<string, []>("concat_404"), val = tensor<int32, [3]>([4, 512, 32])];
            tensor<fp16, [4, 1, 512, 32]> transpose_80_cast_fp16 = transpose(perm = transpose_80_perm_0, x = var_3621_cast_fp16)[name = tensor<string, []>("transpose_200")];
            tensor<fp16, [4, 512, 32]> reshape_120_cast_fp16 = reshape(shape = concat_404, x = transpose_80_cast_fp16)[name = tensor<string, []>("reshape_120_cast_fp16")];
            tensor<int32, [4]> transpose_81_perm_0 = const()[name = tensor<string, []>("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_405 = const()[name = tensor<string, []>("concat_405"), val = tensor<int32, [3]>([4, 32, 512])];
            tensor<fp16, [4, 1, 32, 512]> transpose_81_cast_fp16 = transpose(perm = transpose_81_perm_0, x = var_3625_cast_fp16)[name = tensor<string, []>("transpose_199")];
            tensor<fp16, [4, 32, 512]> reshape_121_cast_fp16 = reshape(shape = concat_405, x = transpose_81_cast_fp16)[name = tensor<string, []>("reshape_121_cast_fp16")];
            tensor<bool, []> matmul_40_transpose_x_0 = const()[name = tensor<string, []>("matmul_40_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_40_transpose_y_0 = const()[name = tensor<string, []>("matmul_40_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 512]> matmul_40_cast_fp16 = matmul(transpose_x = matmul_40_transpose_x_0, transpose_y = matmul_40_transpose_y_0, x = reshape_120_cast_fp16, y = reshape_121_cast_fp16)[name = tensor<string, []>("matmul_40_cast_fp16")];
            tensor<int32, [4]> concat_409 = const()[name = tensor<string, []>("concat_409"), val = tensor<int32, [4]>([4, 1, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> reshape_122_cast_fp16 = reshape(shape = concat_409, x = matmul_40_cast_fp16)[name = tensor<string, []>("reshape_122_cast_fp16")];
            tensor<int32, [4]> attn_scores_61_perm_0 = const()[name = tensor<string, []>("attn_scores_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3632 = const()[name = tensor<string, []>("op_3632"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 512, 1]> var_3633_cast_fp16 = transpose(perm = var_3632, x = pt_21_cast_fp16)[name = tensor<string, []>("transpose_197")];
            tensor<fp16, [4, 32, 512, 512]> var_3634_cast_fp16 = mul(x = var_3633_cast_fp16, y = core_pos_basis_512_to_fp16)[name = tensor<string, []>("op_3634_cast_fp16")];
            tensor<int32, [1]> pos_scores_21_axes_0 = const()[name = tensor<string, []>("pos_scores_21_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_21_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_21_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 512, 512]> pos_scores_21_cast_fp16 = reduce_sum(axes = pos_scores_21_axes_0, keep_dims = pos_scores_21_keep_dims_0, x = var_3634_cast_fp16)[name = tensor<string, []>("pos_scores_21_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = tensor<string, []>("op_3637"), val = tensor<int32, [4]>([4, 512, 1, 512])];
            tensor<fp16, [4, 512, 1, 512]> var_3638_cast_fp16 = reshape(shape = var_3637, x = pos_scores_21_cast_fp16)[name = tensor<string, []>("op_3638_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_61_cast_fp16 = transpose(perm = attn_scores_61_perm_0, x = reshape_122_cast_fp16)[name = tensor<string, []>("transpose_198")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_63_cast_fp16 = add(x = attn_scores_61_cast_fp16, y = var_3638_cast_fp16)[name = tensor<string, []>("attn_scores_63_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_65_cast_fp16 = add(x = attn_scores_63_cast_fp16, y = attn_bias_3_cast_fp16)[name = tensor<string, []>("attn_scores_65_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_weights_21_cast_fp16 = softmax(axis = var_41, x = attn_scores_65_cast_fp16)[name = tensor<string, []>("attn_weights_21_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_329_cast_fp16 = add(x = input_327_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("input_329_cast_fp16")];
            tensor<string, []> x_457_pad_type_0 = const()[name = tensor<string, []>("x_457_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_457_strides_0 = const()[name = tensor<string, []>("x_457_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_457_pad_0 = const()[name = tensor<string, []>("x_457_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_457_dilations_0 = const()[name = tensor<string, []>("x_457_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_457_groups_0 = const()[name = tensor<string, []>("x_457_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_3_stack_layers_2_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238044544)))];
            tensor<fp16, [1152]> core_stacks_3_stack_layers_2_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239224256)))];
            tensor<fp16, [1, 1152, 1, 512]> x_457_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_feed_forward1_in_proj_bias_to_fp16, dilations = x_457_dilations_0, groups = x_457_groups_0, pad = x_457_pad_0, pad_type = x_457_pad_type_0, strides = x_457_strides_0, weight = core_stacks_3_stack_layers_2_feed_forward1_in_proj_weight_to_fp16, x = input_329_cast_fp16)[name = tensor<string, []>("x_457_cast_fp16")];
            tensor<fp16, []> var_3652_to_fp16 = const()[name = tensor<string, []>("op_3652_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 512]> x_459_cast_fp16 = sub(x = x_457_cast_fp16, y = var_3652_to_fp16)[name = tensor<string, []>("x_459_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3654_cast_fp16 = relu(x = x_459_cast_fp16)[name = tensor<string, []>("op_3654_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3655_cast_fp16 = abs(x = x_459_cast_fp16)[name = tensor<string, []>("op_3655_cast_fp16")];
            tensor<fp16, []> const_79_promoted_to_fp16 = const()[name = tensor<string, []>("const_79_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_3656_cast_fp16 = mul(x = var_3655_cast_fp16, y = const_79_promoted_to_fp16)[name = tensor<string, []>("op_3656_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3657_cast_fp16 = exp(x = var_3656_cast_fp16)[name = tensor<string, []>("op_3657_cast_fp16")];
            tensor<fp32, []> var_3658_epsilon_0 = const()[name = tensor<string, []>("op_3658_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_3658_cast_fp16 = log(epsilon = var_3658_epsilon_0, x = var_3657_cast_fp16)[name = tensor<string, []>("op_3658_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3659_cast_fp16 = add(x = var_3654_cast_fp16, y = var_3658_cast_fp16)[name = tensor<string, []>("op_3659_cast_fp16")];
            tensor<fp16, []> var_3660_to_fp16 = const()[name = tensor<string, []>("op_3660_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 512]> var_3661_cast_fp16 = mul(x = x_457_cast_fp16, y = var_3660_to_fp16)[name = tensor<string, []>("op_3661_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3662_cast_fp16 = sub(x = var_3659_cast_fp16, y = var_3661_cast_fp16)[name = tensor<string, []>("op_3662_cast_fp16")];
            tensor<fp16, []> var_3663_to_fp16 = const()[name = tensor<string, []>("op_3663_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 512]> input_331_cast_fp16 = sub(x = var_3662_cast_fp16, y = var_3663_to_fp16)[name = tensor<string, []>("input_331_cast_fp16")];
            tensor<string, []> var_3671_pad_type_0 = const()[name = tensor<string, []>("op_3671_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3671_strides_0 = const()[name = tensor<string, []>("op_3671_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3671_pad_0 = const()[name = tensor<string, []>("op_3671_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3671_dilations_0 = const()[name = tensor<string, []>("op_3671_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3671_groups_0 = const()[name = tensor<string, []>("op_3671_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_3_stack_layers_2_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239226624)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_2_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240406336)))];
            tensor<fp16, [1, 512, 1, 512]> var_3671_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_feed_forward1_out_proj_bias_to_fp16, dilations = var_3671_dilations_0, groups = var_3671_groups_0, pad = var_3671_pad_0, pad_type = var_3671_pad_type_0, strides = var_3671_strides_0, weight = core_stacks_3_stack_layers_2_feed_forward1_out_proj_weight_to_fp16, x = input_331_cast_fp16)[name = tensor<string, []>("op_3671_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_333_cast_fp16 = add(x = input_329_cast_fp16, y = var_3671_cast_fp16)[name = tensor<string, []>("input_333_cast_fp16")];
            tensor<int32, [4]> var_3673_begin_0 = const()[name = tensor<string, []>("op_3673_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3673_end_0 = const()[name = tensor<string, []>("op_3673_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_3673_end_mask_0 = const()[name = tensor<string, []>("op_3673_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3673_cast_fp16 = slice_by_index(begin = var_3673_begin_0, end = var_3673_end_0, end_mask = var_3673_end_mask_0, x = attn_weights_21_cast_fp16)[name = tensor<string, []>("op_3673_cast_fp16")];
            tensor<string, []> x_461_pad_type_0 = const()[name = tensor<string, []>("x_461_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_461_strides_0 = const()[name = tensor<string, []>("x_461_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_461_pad_0 = const()[name = tensor<string, []>("x_461_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_461_dilations_0 = const()[name = tensor<string, []>("x_461_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_461_groups_0 = const()[name = tensor<string, []>("x_461_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_3_stack_layers_2_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240407424)))];
            tensor<fp16, [1152]> core_stacks_3_stack_layers_2_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241587136)))];
            tensor<fp16, [1, 1152, 1, 512]> x_461_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_nonlin_attention_in_proj_bias_to_fp16, dilations = x_461_dilations_0, groups = x_461_groups_0, pad = x_461_pad_0, pad_type = x_461_pad_type_0, strides = x_461_strides_0, weight = core_stacks_3_stack_layers_2_nonlin_attention_in_proj_weight_to_fp16, x = input_333_cast_fp16)[name = tensor<string, []>("x_461_cast_fp16")];
            tensor<int32, [4]> var_3684_begin_0 = const()[name = tensor<string, []>("op_3684_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3684_end_0 = const()[name = tensor<string, []>("op_3684_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_3684_end_mask_0 = const()[name = tensor<string, []>("op_3684_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_3684_cast_fp16 = slice_by_index(begin = var_3684_begin_0, end = var_3684_end_0, end_mask = var_3684_end_mask_0, x = x_461_cast_fp16)[name = tensor<string, []>("op_3684_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> s_gate_21_cast_fp16 = tanh(x = var_3684_cast_fp16)[name = tensor<string, []>("s_gate_21_cast_fp16")];
            tensor<int32, [4]> var_3687_begin_0 = const()[name = tensor<string, []>("op_3687_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_3687_end_0 = const()[name = tensor<string, []>("op_3687_end_0"), val = tensor<int32, [4]>([1, 768, 1, 512])];
            tensor<bool, [4]> var_3687_end_mask_0 = const()[name = tensor<string, []>("op_3687_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> 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 = x_461_cast_fp16)[name = tensor<string, []>("op_3687_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> var_3688_cast_fp16 = mul(x = var_3687_cast_fp16, y = s_gate_21_cast_fp16)[name = tensor<string, []>("op_3688_cast_fp16")];
            tensor<int32, [4]> y_21_begin_0 = const()[name = tensor<string, []>("y_21_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_21_end_0 = const()[name = tensor<string, []>("y_21_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 512])];
            tensor<bool, [4]> y_21_end_mask_0 = const()[name = tensor<string, []>("y_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 512]> y_21_cast_fp16 = slice_by_index(begin = y_21_begin_0, end = y_21_end_0, end_mask = y_21_end_mask_0, x = x_461_cast_fp16)[name = tensor<string, []>("y_21_cast_fp16")];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = tensor<string, []>("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_414 = const()[name = tensor<string, []>("concat_414"), val = tensor<int32, [3]>([1, 512, 512])];
            tensor<fp16, [1, 1, 512, 512]> transpose_82_cast_fp16 = transpose(perm = transpose_82_perm_0, x = var_3673_cast_fp16)[name = tensor<string, []>("transpose_196")];
            tensor<fp16, [1, 512, 512]> reshape_123_cast_fp16 = reshape(shape = concat_414, x = transpose_82_cast_fp16)[name = tensor<string, []>("reshape_123_cast_fp16")];
            tensor<int32, [4]> transpose_83_perm_0 = const()[name = tensor<string, []>("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_415 = const()[name = tensor<string, []>("concat_415"), val = tensor<int32, [3]>([1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> transpose_83_cast_fp16 = transpose(perm = transpose_83_perm_0, x = var_3688_cast_fp16)[name = tensor<string, []>("transpose_195")];
            tensor<fp16, [1, 512, 384]> reshape_124_cast_fp16 = reshape(shape = concat_415, x = transpose_83_cast_fp16)[name = tensor<string, []>("reshape_124_cast_fp16")];
            tensor<bool, []> matmul_41_transpose_x_0 = const()[name = tensor<string, []>("matmul_41_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_41_transpose_y_0 = const()[name = tensor<string, []>("matmul_41_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 384]> matmul_41_cast_fp16 = matmul(transpose_x = matmul_41_transpose_x_0, transpose_y = matmul_41_transpose_y_0, x = reshape_123_cast_fp16, y = reshape_124_cast_fp16)[name = tensor<string, []>("matmul_41_cast_fp16")];
            tensor<int32, [4]> concat_419 = const()[name = tensor<string, []>("concat_419"), val = tensor<int32, [4]>([1, 1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> reshape_125_cast_fp16 = reshape(shape = concat_419, x = matmul_41_cast_fp16)[name = tensor<string, []>("reshape_125_cast_fp16")];
            tensor<int32, [4]> v_21_perm_0 = const()[name = tensor<string, []>("v_21_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 512]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = reshape_125_cast_fp16)[name = tensor<string, []>("transpose_194")];
            tensor<fp16, [1, 384, 1, 512]> input_335_cast_fp16 = mul(x = v_21_cast_fp16, y = y_21_cast_fp16)[name = tensor<string, []>("input_335_cast_fp16")];
            tensor<string, []> var_3700_pad_type_0 = const()[name = tensor<string, []>("op_3700_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3700_strides_0 = const()[name = tensor<string, []>("op_3700_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3700_pad_0 = const()[name = tensor<string, []>("op_3700_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3700_dilations_0 = const()[name = tensor<string, []>("op_3700_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3700_groups_0 = const()[name = tensor<string, []>("op_3700_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_3_stack_layers_2_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241589504)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_2_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241982784)))];
            tensor<fp16, [1, 512, 1, 512]> var_3700_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_nonlin_attention_out_proj_bias_to_fp16, dilations = var_3700_dilations_0, groups = var_3700_groups_0, pad = var_3700_pad_0, pad_type = var_3700_pad_type_0, strides = var_3700_strides_0, weight = core_stacks_3_stack_layers_2_nonlin_attention_out_proj_weight_to_fp16, x = input_335_cast_fp16)[name = tensor<string, []>("op_3700_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_463_cast_fp16 = add(x = input_333_cast_fp16, y = var_3700_cast_fp16)[name = tensor<string, []>("x_463_cast_fp16")];
            tensor<string, []> var_3711_pad_type_0 = const()[name = tensor<string, []>("op_3711_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3711_strides_0 = const()[name = tensor<string, []>("op_3711_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3711_pad_0 = const()[name = tensor<string, []>("op_3711_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3711_dilations_0 = const()[name = tensor<string, []>("op_3711_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3711_groups_0 = const()[name = tensor<string, []>("op_3711_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_3_stack_layers_2_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241983872)))];
            tensor<fp16, [48]> core_stacks_3_stack_layers_2_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242033088)))];
            tensor<fp16, [1, 48, 1, 512]> var_3711_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_self_attn1_in_proj_bias_to_fp16, dilations = var_3711_dilations_0, groups = var_3711_groups_0, pad = var_3711_pad_0, pad_type = var_3711_pad_type_0, strides = var_3711_strides_0, weight = core_stacks_3_stack_layers_2_self_attn1_in_proj_weight_to_fp16, x = x_463_cast_fp16)[name = tensor<string, []>("op_3711_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = tensor<string, []>("op_3712"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_3713_cast_fp16 = reshape(shape = var_3712, x = var_3711_cast_fp16)[name = tensor<string, []>("op_3713_cast_fp16")];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = tensor<string, []>("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_424 = const()[name = tensor<string, []>("concat_424"), val = tensor<int32, [3]>([4, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> transpose_84_cast_fp16 = transpose(perm = transpose_84_perm_0, x = attn_weights_21_cast_fp16)[name = tensor<string, []>("transpose_193")];
            tensor<fp16, [4, 512, 512]> reshape_126_cast_fp16 = reshape(shape = concat_424, x = transpose_84_cast_fp16)[name = tensor<string, []>("reshape_126_cast_fp16")];
            tensor<int32, [4]> transpose_85_perm_0 = const()[name = tensor<string, []>("transpose_85_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_425 = const()[name = tensor<string, []>("concat_425"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_85_cast_fp16 = transpose(perm = transpose_85_perm_0, x = var_3713_cast_fp16)[name = tensor<string, []>("transpose_192")];
            tensor<fp16, [4, 512, 12]> reshape_127_cast_fp16 = reshape(shape = concat_425, x = transpose_85_cast_fp16)[name = tensor<string, []>("reshape_127_cast_fp16")];
            tensor<bool, []> matmul_42_transpose_x_0 = const()[name = tensor<string, []>("matmul_42_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_42_transpose_y_0 = const()[name = tensor<string, []>("matmul_42_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_42_cast_fp16 = matmul(transpose_x = matmul_42_transpose_x_0, transpose_y = matmul_42_transpose_y_0, x = reshape_126_cast_fp16, y = reshape_127_cast_fp16)[name = tensor<string, []>("matmul_42_cast_fp16")];
            tensor<int32, [4]> concat_429 = const()[name = tensor<string, []>("concat_429"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_128_cast_fp16 = reshape(shape = concat_429, x = matmul_42_cast_fp16)[name = tensor<string, []>("reshape_128_cast_fp16")];
            tensor<int32, [4]> out_41_perm_0 = const()[name = tensor<string, []>("out_41_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_3716 = const()[name = tensor<string, []>("op_3716"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_41_cast_fp16 = transpose(perm = out_41_perm_0, x = reshape_128_cast_fp16)[name = tensor<string, []>("transpose_191")];
            tensor<fp16, [1, 48, 1, 512]> input_337_cast_fp16 = reshape(shape = var_3716, x = out_41_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
            tensor<string, []> var_3724_pad_type_0 = const()[name = tensor<string, []>("op_3724_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3724_strides_0 = const()[name = tensor<string, []>("op_3724_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3724_pad_0 = const()[name = tensor<string, []>("op_3724_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3724_dilations_0 = const()[name = tensor<string, []>("op_3724_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3724_groups_0 = const()[name = tensor<string, []>("op_3724_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_3_stack_layers_2_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242033280)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_2_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242082496)))];
            tensor<fp16, [1, 512, 1, 512]> var_3724_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_self_attn1_out_proj_bias_to_fp16, dilations = var_3724_dilations_0, groups = var_3724_groups_0, pad = var_3724_pad_0, pad_type = var_3724_pad_type_0, strides = var_3724_strides_0, weight = core_stacks_3_stack_layers_2_self_attn1_out_proj_weight_to_fp16, x = input_337_cast_fp16)[name = tensor<string, []>("op_3724_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_465_cast_fp16 = add(x = x_463_cast_fp16, y = var_3724_cast_fp16)[name = tensor<string, []>("x_465_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_467_cast_fp16 = add(x = x_465_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("x_467_cast_fp16")];
            tensor<string, []> x_469_pad_type_0 = const()[name = tensor<string, []>("x_469_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_469_strides_0 = const()[name = tensor<string, []>("x_469_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_469_pad_0 = const()[name = tensor<string, []>("x_469_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_469_dilations_0 = const()[name = tensor<string, []>("x_469_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_469_groups_0 = const()[name = tensor<string, []>("x_469_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_3_stack_layers_2_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242083584)))];
            tensor<fp16, [1024]> core_stacks_3_stack_layers_2_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243132224)))];
            tensor<fp16, [1, 1024, 1, 512]> x_469_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_conv_module1_in_proj_bias_to_fp16, dilations = x_469_dilations_0, groups = x_469_groups_0, pad = x_469_pad_0, pad_type = x_469_pad_type_0, strides = x_469_strides_0, weight = core_stacks_3_stack_layers_2_conv_module1_in_proj_weight_to_fp16, x = x_467_cast_fp16)[name = tensor<string, []>("x_469_cast_fp16")];
            tensor<int32, [4]> var_3738_begin_0 = const()[name = tensor<string, []>("op_3738_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3738_end_0 = const()[name = tensor<string, []>("op_3738_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_3738_end_mask_0 = const()[name = tensor<string, []>("op_3738_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3738_cast_fp16 = slice_by_index(begin = var_3738_begin_0, end = var_3738_end_0, end_mask = var_3738_end_mask_0, x = x_469_cast_fp16)[name = tensor<string, []>("op_3738_cast_fp16")];
            tensor<int32, [4]> var_3740_begin_0 = const()[name = tensor<string, []>("op_3740_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_3740_end_0 = const()[name = tensor<string, []>("op_3740_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_3740_end_mask_0 = const()[name = tensor<string, []>("op_3740_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3740_cast_fp16 = slice_by_index(begin = var_3740_begin_0, end = var_3740_end_0, end_mask = var_3740_end_mask_0, x = x_469_cast_fp16)[name = tensor<string, []>("op_3740_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3741_cast_fp16 = sigmoid(x = var_3740_cast_fp16)[name = tensor<string, []>("op_3741_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_471_cast_fp16 = mul(x = var_3738_cast_fp16, y = var_3741_cast_fp16)[name = tensor<string, []>("x_471_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_339_cast_fp16 = mul(x = x_471_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_339_cast_fp16")];
            tensor<string, []> x_473_pad_type_0 = const()[name = tensor<string, []>("x_473_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_473_pad_0 = const()[name = tensor<string, []>("x_473_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_473_groups_0 = const()[name = tensor<string, []>("x_473_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_473_strides_0 = const()[name = tensor<string, []>("x_473_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_473_dilations_0 = const()[name = tensor<string, []>("x_473_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_3_stack_layers_2_conv_module1_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module1_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243134336)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_2_conv_module1_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module1_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243149760)))];
            tensor<fp16, [1, 512, 1, 512]> x_473_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_conv_module1_dw_bias_to_fp16, dilations = x_473_dilations_0, groups = x_473_groups_0, pad = x_473_pad_0, pad_type = x_473_pad_type_0, strides = x_473_strides_0, weight = core_stacks_3_stack_layers_2_conv_module1_dw_weight_to_fp16, x = input_339_cast_fp16)[name = tensor<string, []>("x_473_cast_fp16")];
            tensor<fp16, []> var_3751_to_fp16 = const()[name = tensor<string, []>("op_3751_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_475_cast_fp16 = sub(x = x_473_cast_fp16, y = var_3751_to_fp16)[name = tensor<string, []>("x_475_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3753_cast_fp16 = relu(x = x_475_cast_fp16)[name = tensor<string, []>("op_3753_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3754_cast_fp16 = abs(x = x_475_cast_fp16)[name = tensor<string, []>("op_3754_cast_fp16")];
            tensor<fp16, []> const_81_promoted_to_fp16 = const()[name = tensor<string, []>("const_81_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3755_cast_fp16 = mul(x = var_3754_cast_fp16, y = const_81_promoted_to_fp16)[name = tensor<string, []>("op_3755_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3756_cast_fp16 = exp(x = var_3755_cast_fp16)[name = tensor<string, []>("op_3756_cast_fp16")];
            tensor<fp32, []> var_3757_epsilon_0 = const()[name = tensor<string, []>("op_3757_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3757_cast_fp16 = log(epsilon = var_3757_epsilon_0, x = var_3756_cast_fp16)[name = tensor<string, []>("op_3757_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3758_cast_fp16 = add(x = var_3753_cast_fp16, y = var_3757_cast_fp16)[name = tensor<string, []>("op_3758_cast_fp16")];
            tensor<fp16, []> var_3759_to_fp16 = const()[name = tensor<string, []>("op_3759_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_3760_cast_fp16 = mul(x = x_473_cast_fp16, y = var_3759_to_fp16)[name = tensor<string, []>("op_3760_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3761_cast_fp16 = sub(x = var_3758_cast_fp16, y = var_3760_cast_fp16)[name = tensor<string, []>("op_3761_cast_fp16")];
            tensor<fp16, []> var_3762_to_fp16 = const()[name = tensor<string, []>("op_3762_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_341_cast_fp16 = sub(x = var_3761_cast_fp16, y = var_3762_to_fp16)[name = tensor<string, []>("input_341_cast_fp16")];
            tensor<string, []> var_3770_pad_type_0 = const()[name = tensor<string, []>("op_3770_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3770_strides_0 = const()[name = tensor<string, []>("op_3770_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3770_pad_0 = const()[name = tensor<string, []>("op_3770_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3770_dilations_0 = const()[name = tensor<string, []>("op_3770_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3770_groups_0 = const()[name = tensor<string, []>("op_3770_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_3_stack_layers_2_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243150848)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_2_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243675200)))];
            tensor<fp16, [1, 512, 1, 512]> var_3770_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_conv_module1_out_proj_bias_to_fp16, dilations = var_3770_dilations_0, groups = var_3770_groups_0, pad = var_3770_pad_0, pad_type = var_3770_pad_type_0, strides = var_3770_strides_0, weight = core_stacks_3_stack_layers_2_conv_module1_out_proj_weight_to_fp16, x = input_341_cast_fp16)[name = tensor<string, []>("op_3770_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_343_cast_fp16 = add(x = x_467_cast_fp16, y = var_3770_cast_fp16)[name = tensor<string, []>("input_343_cast_fp16")];
            tensor<string, []> x_477_pad_type_0 = const()[name = tensor<string, []>("x_477_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_477_strides_0 = const()[name = tensor<string, []>("x_477_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_477_pad_0 = const()[name = tensor<string, []>("x_477_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_477_dilations_0 = const()[name = tensor<string, []>("x_477_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_477_groups_0 = const()[name = tensor<string, []>("x_477_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_3_stack_layers_2_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243676288)))];
            tensor<fp16, [1536]> core_stacks_3_stack_layers_2_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245249216)))];
            tensor<fp16, [1, 1536, 1, 512]> x_477_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_feed_forward2_in_proj_bias_to_fp16, dilations = x_477_dilations_0, groups = x_477_groups_0, pad = x_477_pad_0, pad_type = x_477_pad_type_0, strides = x_477_strides_0, weight = core_stacks_3_stack_layers_2_feed_forward2_in_proj_weight_to_fp16, x = input_343_cast_fp16)[name = tensor<string, []>("x_477_cast_fp16")];
            tensor<fp16, []> var_3781_to_fp16 = const()[name = tensor<string, []>("op_3781_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 512]> x_479_cast_fp16 = sub(x = x_477_cast_fp16, y = var_3781_to_fp16)[name = tensor<string, []>("x_479_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3783_cast_fp16 = relu(x = x_479_cast_fp16)[name = tensor<string, []>("op_3783_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3784_cast_fp16 = abs(x = x_479_cast_fp16)[name = tensor<string, []>("op_3784_cast_fp16")];
            tensor<fp16, []> const_82_promoted_to_fp16 = const()[name = tensor<string, []>("const_82_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_3785_cast_fp16 = mul(x = var_3784_cast_fp16, y = const_82_promoted_to_fp16)[name = tensor<string, []>("op_3785_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3786_cast_fp16 = exp(x = var_3785_cast_fp16)[name = tensor<string, []>("op_3786_cast_fp16")];
            tensor<fp32, []> var_3787_epsilon_0 = const()[name = tensor<string, []>("op_3787_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_3787_cast_fp16 = log(epsilon = var_3787_epsilon_0, x = var_3786_cast_fp16)[name = tensor<string, []>("op_3787_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3788_cast_fp16 = add(x = var_3783_cast_fp16, y = var_3787_cast_fp16)[name = tensor<string, []>("op_3788_cast_fp16")];
            tensor<fp16, []> var_3789_to_fp16 = const()[name = tensor<string, []>("op_3789_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 512]> var_3790_cast_fp16 = mul(x = x_477_cast_fp16, y = var_3789_to_fp16)[name = tensor<string, []>("op_3790_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_3791_cast_fp16 = sub(x = var_3788_cast_fp16, y = var_3790_cast_fp16)[name = tensor<string, []>("op_3791_cast_fp16")];
            tensor<fp16, []> var_3792_to_fp16 = const()[name = tensor<string, []>("op_3792_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 512]> input_345_cast_fp16 = sub(x = var_3791_cast_fp16, y = var_3792_to_fp16)[name = tensor<string, []>("input_345_cast_fp16")];
            tensor<string, []> var_3800_pad_type_0 = const()[name = tensor<string, []>("op_3800_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3800_strides_0 = const()[name = tensor<string, []>("op_3800_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3800_pad_0 = const()[name = tensor<string, []>("op_3800_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3800_dilations_0 = const()[name = tensor<string, []>("op_3800_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3800_groups_0 = const()[name = tensor<string, []>("op_3800_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_3_stack_layers_2_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(245252352)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_2_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246825280)))];
            tensor<fp16, [1, 512, 1, 512]> var_3800_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_feed_forward2_out_proj_bias_to_fp16, dilations = var_3800_dilations_0, groups = var_3800_groups_0, pad = var_3800_pad_0, pad_type = var_3800_pad_type_0, strides = var_3800_strides_0, weight = core_stacks_3_stack_layers_2_feed_forward2_out_proj_weight_to_fp16, x = input_345_cast_fp16)[name = tensor<string, []>("op_3800_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> src_45_cast_fp16 = add(x = input_343_cast_fp16, y = var_3800_cast_fp16)[name = tensor<string, []>("src_45_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3803_cast_fp16 = sub(x = src_45_cast_fp16, y = input_327_cast_fp16)[name = tensor<string, []>("op_3803_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_2_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246826368)))];
            tensor<fp16, [1, 512, 1, 512]> var_3804_cast_fp16 = mul(x = var_3803_cast_fp16, y = core_stacks_3_stack_layers_2_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_3804_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_481_cast_fp16 = add(x = input_327_cast_fp16, y = var_3804_cast_fp16)[name = tensor<string, []>("x_481_cast_fp16")];
            tensor<string, []> var_3815_pad_type_0 = const()[name = tensor<string, []>("op_3815_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3815_strides_0 = const()[name = tensor<string, []>("op_3815_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3815_pad_0 = const()[name = tensor<string, []>("op_3815_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3815_dilations_0 = const()[name = tensor<string, []>("op_3815_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3815_groups_0 = const()[name = tensor<string, []>("op_3815_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_3_stack_layers_2_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246827456)))];
            tensor<fp16, [48]> core_stacks_3_stack_layers_2_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246876672)))];
            tensor<fp16, [1, 48, 1, 512]> var_3815_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_self_attn2_in_proj_bias_to_fp16, dilations = var_3815_dilations_0, groups = var_3815_groups_0, pad = var_3815_pad_0, pad_type = var_3815_pad_type_0, strides = var_3815_strides_0, weight = core_stacks_3_stack_layers_2_self_attn2_in_proj_weight_to_fp16, x = x_481_cast_fp16)[name = tensor<string, []>("op_3815_cast_fp16")];
            tensor<int32, [4]> var_3816 = const()[name = tensor<string, []>("op_3816"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_3817_cast_fp16 = reshape(shape = var_3816, x = var_3815_cast_fp16)[name = tensor<string, []>("op_3817_cast_fp16")];
            tensor<int32, [4]> transpose_87_perm_0 = const()[name = tensor<string, []>("transpose_87_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_435 = const()[name = tensor<string, []>("concat_435"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_87_cast_fp16 = transpose(perm = transpose_87_perm_0, x = var_3817_cast_fp16)[name = tensor<string, []>("transpose_190")];
            tensor<fp16, [4, 512, 12]> reshape_130_cast_fp16 = reshape(shape = concat_435, x = transpose_87_cast_fp16)[name = tensor<string, []>("reshape_130_cast_fp16")];
            tensor<bool, []> matmul_43_transpose_x_0 = const()[name = tensor<string, []>("matmul_43_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_43_transpose_y_0 = const()[name = tensor<string, []>("matmul_43_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_43_cast_fp16 = matmul(transpose_x = matmul_43_transpose_x_0, transpose_y = matmul_43_transpose_y_0, x = reshape_126_cast_fp16, y = reshape_130_cast_fp16)[name = tensor<string, []>("matmul_43_cast_fp16")];
            tensor<int32, [4]> concat_439 = const()[name = tensor<string, []>("concat_439"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_131_cast_fp16 = reshape(shape = concat_439, x = matmul_43_cast_fp16)[name = tensor<string, []>("reshape_131_cast_fp16")];
            tensor<int32, [4]> out_43_perm_0 = const()[name = tensor<string, []>("out_43_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_3820 = const()[name = tensor<string, []>("op_3820"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_43_cast_fp16 = transpose(perm = out_43_perm_0, x = reshape_131_cast_fp16)[name = tensor<string, []>("transpose_189")];
            tensor<fp16, [1, 48, 1, 512]> input_347_cast_fp16 = reshape(shape = var_3820, x = out_43_cast_fp16)[name = tensor<string, []>("input_347_cast_fp16")];
            tensor<string, []> var_3828_pad_type_0 = const()[name = tensor<string, []>("op_3828_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3828_strides_0 = const()[name = tensor<string, []>("op_3828_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3828_pad_0 = const()[name = tensor<string, []>("op_3828_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3828_dilations_0 = const()[name = tensor<string, []>("op_3828_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3828_groups_0 = const()[name = tensor<string, []>("op_3828_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_3_stack_layers_2_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246876864)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_2_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246926080)))];
            tensor<fp16, [1, 512, 1, 512]> var_3828_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_self_attn2_out_proj_bias_to_fp16, dilations = var_3828_dilations_0, groups = var_3828_groups_0, pad = var_3828_pad_0, pad_type = var_3828_pad_type_0, strides = var_3828_strides_0, weight = core_stacks_3_stack_layers_2_self_attn2_out_proj_weight_to_fp16, x = input_347_cast_fp16)[name = tensor<string, []>("op_3828_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_483_cast_fp16 = add(x = x_481_cast_fp16, y = var_3828_cast_fp16)[name = tensor<string, []>("x_483_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_485_cast_fp16 = add(x = x_483_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("x_485_cast_fp16")];
            tensor<string, []> x_487_pad_type_0 = const()[name = tensor<string, []>("x_487_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_487_strides_0 = const()[name = tensor<string, []>("x_487_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_487_pad_0 = const()[name = tensor<string, []>("x_487_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_487_dilations_0 = const()[name = tensor<string, []>("x_487_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_487_groups_0 = const()[name = tensor<string, []>("x_487_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_3_stack_layers_2_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246927168)))];
            tensor<fp16, [1024]> core_stacks_3_stack_layers_2_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247975808)))];
            tensor<fp16, [1, 1024, 1, 512]> x_487_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_conv_module2_in_proj_bias_to_fp16, dilations = x_487_dilations_0, groups = x_487_groups_0, pad = x_487_pad_0, pad_type = x_487_pad_type_0, strides = x_487_strides_0, weight = core_stacks_3_stack_layers_2_conv_module2_in_proj_weight_to_fp16, x = x_485_cast_fp16)[name = tensor<string, []>("x_487_cast_fp16")];
            tensor<int32, [4]> var_3842_begin_0 = const()[name = tensor<string, []>("op_3842_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3842_end_0 = const()[name = tensor<string, []>("op_3842_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_3842_end_mask_0 = const()[name = tensor<string, []>("op_3842_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3842_cast_fp16 = slice_by_index(begin = var_3842_begin_0, end = var_3842_end_0, end_mask = var_3842_end_mask_0, x = x_487_cast_fp16)[name = tensor<string, []>("op_3842_cast_fp16")];
            tensor<int32, [4]> var_3844_begin_0 = const()[name = tensor<string, []>("op_3844_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_3844_end_0 = const()[name = tensor<string, []>("op_3844_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_3844_end_mask_0 = const()[name = tensor<string, []>("op_3844_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3844_cast_fp16 = slice_by_index(begin = var_3844_begin_0, end = var_3844_end_0, end_mask = var_3844_end_mask_0, x = x_487_cast_fp16)[name = tensor<string, []>("op_3844_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3845_cast_fp16 = sigmoid(x = var_3844_cast_fp16)[name = tensor<string, []>("op_3845_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_489_cast_fp16 = mul(x = var_3842_cast_fp16, y = var_3845_cast_fp16)[name = tensor<string, []>("x_489_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_349_cast_fp16 = mul(x = x_489_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_349_cast_fp16")];
            tensor<string, []> x_491_pad_type_0 = const()[name = tensor<string, []>("x_491_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_491_pad_0 = const()[name = tensor<string, []>("x_491_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_491_groups_0 = const()[name = tensor<string, []>("x_491_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_491_strides_0 = const()[name = tensor<string, []>("x_491_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_491_dilations_0 = const()[name = tensor<string, []>("x_491_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_3_stack_layers_2_conv_module2_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module2_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247977920)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_2_conv_module2_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module2_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247993344)))];
            tensor<fp16, [1, 512, 1, 512]> x_491_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_conv_module2_dw_bias_to_fp16, dilations = x_491_dilations_0, groups = x_491_groups_0, pad = x_491_pad_0, pad_type = x_491_pad_type_0, strides = x_491_strides_0, weight = core_stacks_3_stack_layers_2_conv_module2_dw_weight_to_fp16, x = input_349_cast_fp16)[name = tensor<string, []>("x_491_cast_fp16")];
            tensor<fp16, []> var_3855_to_fp16 = const()[name = tensor<string, []>("op_3855_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_493_cast_fp16 = sub(x = x_491_cast_fp16, y = var_3855_to_fp16)[name = tensor<string, []>("x_493_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3857_cast_fp16 = relu(x = x_493_cast_fp16)[name = tensor<string, []>("op_3857_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3858_cast_fp16 = abs(x = x_493_cast_fp16)[name = tensor<string, []>("op_3858_cast_fp16")];
            tensor<fp16, []> const_84_promoted_to_fp16 = const()[name = tensor<string, []>("const_84_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3859_cast_fp16 = mul(x = var_3858_cast_fp16, y = const_84_promoted_to_fp16)[name = tensor<string, []>("op_3859_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3860_cast_fp16 = exp(x = var_3859_cast_fp16)[name = tensor<string, []>("op_3860_cast_fp16")];
            tensor<fp32, []> var_3861_epsilon_0 = const()[name = tensor<string, []>("op_3861_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_3861_cast_fp16 = log(epsilon = var_3861_epsilon_0, x = var_3860_cast_fp16)[name = tensor<string, []>("op_3861_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3862_cast_fp16 = add(x = var_3857_cast_fp16, y = var_3861_cast_fp16)[name = tensor<string, []>("op_3862_cast_fp16")];
            tensor<fp16, []> var_3863_to_fp16 = const()[name = tensor<string, []>("op_3863_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_3864_cast_fp16 = mul(x = x_491_cast_fp16, y = var_3863_to_fp16)[name = tensor<string, []>("op_3864_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3865_cast_fp16 = sub(x = var_3862_cast_fp16, y = var_3864_cast_fp16)[name = tensor<string, []>("op_3865_cast_fp16")];
            tensor<fp16, []> var_3866_to_fp16 = const()[name = tensor<string, []>("op_3866_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_351_cast_fp16 = sub(x = var_3865_cast_fp16, y = var_3866_to_fp16)[name = tensor<string, []>("input_351_cast_fp16")];
            tensor<string, []> var_3874_pad_type_0 = const()[name = tensor<string, []>("op_3874_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3874_strides_0 = const()[name = tensor<string, []>("op_3874_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3874_pad_0 = const()[name = tensor<string, []>("op_3874_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3874_dilations_0 = const()[name = tensor<string, []>("op_3874_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3874_groups_0 = const()[name = tensor<string, []>("op_3874_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_3_stack_layers_2_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247994432)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_2_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248518784)))];
            tensor<fp16, [1, 512, 1, 512]> var_3874_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_conv_module2_out_proj_bias_to_fp16, dilations = var_3874_dilations_0, groups = var_3874_groups_0, pad = var_3874_pad_0, pad_type = var_3874_pad_type_0, strides = var_3874_strides_0, weight = core_stacks_3_stack_layers_2_conv_module2_out_proj_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("op_3874_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_353_cast_fp16 = add(x = x_485_cast_fp16, y = var_3874_cast_fp16)[name = tensor<string, []>("input_353_cast_fp16")];
            tensor<string, []> x_495_pad_type_0 = const()[name = tensor<string, []>("x_495_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_495_strides_0 = const()[name = tensor<string, []>("x_495_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_495_pad_0 = const()[name = tensor<string, []>("x_495_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_495_dilations_0 = const()[name = tensor<string, []>("x_495_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_495_groups_0 = const()[name = tensor<string, []>("x_495_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_3_stack_layers_2_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248519872)))];
            tensor<fp16, [1920]> core_stacks_3_stack_layers_2_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250486016)))];
            tensor<fp16, [1, 1920, 1, 512]> x_495_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_feed_forward3_in_proj_bias_to_fp16, dilations = x_495_dilations_0, groups = x_495_groups_0, pad = x_495_pad_0, pad_type = x_495_pad_type_0, strides = x_495_strides_0, weight = core_stacks_3_stack_layers_2_feed_forward3_in_proj_weight_to_fp16, x = input_353_cast_fp16)[name = tensor<string, []>("x_495_cast_fp16")];
            tensor<fp16, []> var_3885_to_fp16 = const()[name = tensor<string, []>("op_3885_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 512]> x_497_cast_fp16 = sub(x = x_495_cast_fp16, y = var_3885_to_fp16)[name = tensor<string, []>("x_497_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3887_cast_fp16 = relu(x = x_497_cast_fp16)[name = tensor<string, []>("op_3887_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3888_cast_fp16 = abs(x = x_497_cast_fp16)[name = tensor<string, []>("op_3888_cast_fp16")];
            tensor<fp16, []> const_85_promoted_to_fp16 = const()[name = tensor<string, []>("const_85_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_3889_cast_fp16 = mul(x = var_3888_cast_fp16, y = const_85_promoted_to_fp16)[name = tensor<string, []>("op_3889_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3890_cast_fp16 = exp(x = var_3889_cast_fp16)[name = tensor<string, []>("op_3890_cast_fp16")];
            tensor<fp32, []> var_3891_epsilon_0 = const()[name = tensor<string, []>("op_3891_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_3891_cast_fp16 = log(epsilon = var_3891_epsilon_0, x = var_3890_cast_fp16)[name = tensor<string, []>("op_3891_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3892_cast_fp16 = add(x = var_3887_cast_fp16, y = var_3891_cast_fp16)[name = tensor<string, []>("op_3892_cast_fp16")];
            tensor<fp16, []> var_3893_to_fp16 = const()[name = tensor<string, []>("op_3893_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 512]> var_3894_cast_fp16 = mul(x = x_495_cast_fp16, y = var_3893_to_fp16)[name = tensor<string, []>("op_3894_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_3895_cast_fp16 = sub(x = var_3892_cast_fp16, y = var_3894_cast_fp16)[name = tensor<string, []>("op_3895_cast_fp16")];
            tensor<fp16, []> var_3896_to_fp16 = const()[name = tensor<string, []>("op_3896_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 512]> input_355_cast_fp16 = sub(x = var_3895_cast_fp16, y = var_3896_to_fp16)[name = tensor<string, []>("input_355_cast_fp16")];
            tensor<string, []> var_3904_pad_type_0 = const()[name = tensor<string, []>("op_3904_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3904_strides_0 = const()[name = tensor<string, []>("op_3904_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3904_pad_0 = const()[name = tensor<string, []>("op_3904_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3904_dilations_0 = const()[name = tensor<string, []>("op_3904_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3904_groups_0 = const()[name = tensor<string, []>("op_3904_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_3_stack_layers_2_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250489920)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_2_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252456064)))];
            tensor<fp16, [1, 512, 1, 512]> var_3904_cast_fp16 = conv(bias = core_stacks_3_stack_layers_2_feed_forward3_out_proj_bias_to_fp16, dilations = var_3904_dilations_0, groups = var_3904_groups_0, pad = var_3904_pad_0, pad_type = var_3904_pad_type_0, strides = var_3904_strides_0, weight = core_stacks_3_stack_layers_2_feed_forward3_out_proj_weight_to_fp16, x = input_355_cast_fp16)[name = tensor<string, []>("op_3904_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_499_cast_fp16 = add(x = input_353_cast_fp16, y = var_3904_cast_fp16)[name = tensor<string, []>("x_499_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_2_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252457152)))];
            tensor<fp16, [1, 512, 1, 512]> d_21_cast_fp16 = sub(x = x_499_cast_fp16, y = core_stacks_3_stack_layers_2_norm_bias_to_fp16)[name = tensor<string, []>("d_21_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3909_cast_fp16 = mul(x = d_21_cast_fp16, y = d_21_cast_fp16)[name = tensor<string, []>("op_3909_cast_fp16")];
            tensor<int32, [1]> var_3911_axes_0 = const()[name = tensor<string, []>("op_3911_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_3911_keep_dims_0 = const()[name = tensor<string, []>("op_3911_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 512]> var_3911_cast_fp16 = reduce_mean(axes = var_3911_axes_0, keep_dims = var_3911_keep_dims_0, x = var_3909_cast_fp16)[name = tensor<string, []>("op_3911_cast_fp16")];
            tensor<fp32, []> scales_21_epsilon_0 = const()[name = tensor<string, []>("scales_21_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 512]> scales_21_cast_fp16 = rsqrt(epsilon = scales_21_epsilon_0, x = var_3911_cast_fp16)[name = tensor<string, []>("scales_21_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3913_cast_fp16 = mul(x = x_499_cast_fp16, y = scales_21_cast_fp16)[name = tensor<string, []>("op_3913_cast_fp16")];
            tensor<fp16, []> core_stacks_3_stack_layers_2_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.01p-2)];
            tensor<fp16, [1, 512, 1, 512]> src_47_cast_fp16 = mul(x = var_3913_cast_fp16, y = core_stacks_3_stack_layers_2_norm_scale_to_fp16)[name = tensor<string, []>("src_47_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_3916_cast_fp16 = sub(x = src_47_cast_fp16, y = input_327_cast_fp16)[name = tensor<string, []>("op_3916_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_2_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_2_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252458240)))];
            tensor<fp16, [1, 512, 1, 512]> var_3917_cast_fp16 = mul(x = var_3916_cast_fp16, y = core_stacks_3_stack_layers_2_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_3917_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_357_cast_fp16 = add(x = input_327_cast_fp16, y = var_3917_cast_fp16)[name = tensor<string, []>("input_357_cast_fp16")];
            tensor<string, []> proj_23_pad_type_0 = const()[name = tensor<string, []>("proj_23_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_23_strides_0 = const()[name = tensor<string, []>("proj_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_23_pad_0 = const()[name = tensor<string, []>("proj_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_23_dilations_0 = const()[name = tensor<string, []>("proj_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_23_groups_0 = const()[name = tensor<string, []>("proj_23_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_3_stack_layers_3_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252459328)))];
            tensor<fp16, [384]> core_stacks_3_stack_layers_3_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252852608)))];
            tensor<fp16, [1, 384, 1, 512]> proj_23_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_23_dilations_0, groups = proj_23_groups_0, pad = proj_23_pad_0, pad_type = proj_23_pad_type_0, strides = proj_23_strides_0, weight = core_stacks_3_stack_layers_3_self_attn_weights_in_proj_weight_to_fp16, x = input_357_cast_fp16)[name = tensor<string, []>("proj_23_cast_fp16")];
            tensor<int32, [4]> var_3940_begin_0 = const()[name = tensor<string, []>("op_3940_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3940_end_0 = const()[name = tensor<string, []>("op_3940_end_0"), val = tensor<int32, [4]>([1, 128, 1, 512])];
            tensor<bool, [4]> var_3940_end_mask_0 = const()[name = tensor<string, []>("op_3940_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_3940_cast_fp16 = slice_by_index(begin = var_3940_begin_0, end = var_3940_end_0, end_mask = var_3940_end_mask_0, x = proj_23_cast_fp16)[name = tensor<string, []>("op_3940_cast_fp16")];
            tensor<int32, [4]> var_3941 = const()[name = tensor<string, []>("op_3941"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_3942_cast_fp16 = reshape(shape = var_3941, x = var_3940_cast_fp16)[name = tensor<string, []>("op_3942_cast_fp16")];
            tensor<int32, [4]> var_3944_begin_0 = const()[name = tensor<string, []>("op_3944_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_3944_end_0 = const()[name = tensor<string, []>("op_3944_end_0"), val = tensor<int32, [4]>([1, 256, 1, 512])];
            tensor<bool, [4]> var_3944_end_mask_0 = const()[name = tensor<string, []>("op_3944_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_3944_cast_fp16 = slice_by_index(begin = var_3944_begin_0, end = var_3944_end_0, end_mask = var_3944_end_mask_0, x = proj_23_cast_fp16)[name = tensor<string, []>("op_3944_cast_fp16")];
            tensor<int32, [4]> var_3945 = const()[name = tensor<string, []>("op_3945"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> var_3946_cast_fp16 = reshape(shape = var_3945, x = var_3944_cast_fp16)[name = tensor<string, []>("op_3946_cast_fp16")];
            tensor<int32, [4]> var_3948_begin_0 = const()[name = tensor<string, []>("op_3948_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_3948_end_0 = const()[name = tensor<string, []>("op_3948_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_3948_end_mask_0 = const()[name = tensor<string, []>("op_3948_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 512]> var_3948_cast_fp16 = slice_by_index(begin = var_3948_begin_0, end = var_3948_end_0, end_mask = var_3948_end_mask_0, x = proj_23_cast_fp16)[name = tensor<string, []>("op_3948_cast_fp16")];
            tensor<int32, [4]> var_3949 = const()[name = tensor<string, []>("op_3949"), val = tensor<int32, [4]>([4, 32, 1, 512])];
            tensor<fp16, [4, 32, 1, 512]> pt_23_cast_fp16 = reshape(shape = var_3949, x = var_3948_cast_fp16)[name = tensor<string, []>("pt_23_cast_fp16")];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = tensor<string, []>("transpose_88_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_444 = const()[name = tensor<string, []>("concat_444"), val = tensor<int32, [3]>([4, 512, 32])];
            tensor<fp16, [4, 1, 512, 32]> transpose_88_cast_fp16 = transpose(perm = transpose_88_perm_0, x = var_3942_cast_fp16)[name = tensor<string, []>("transpose_188")];
            tensor<fp16, [4, 512, 32]> reshape_132_cast_fp16 = reshape(shape = concat_444, x = transpose_88_cast_fp16)[name = tensor<string, []>("reshape_132_cast_fp16")];
            tensor<int32, [4]> transpose_89_perm_0 = const()[name = tensor<string, []>("transpose_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_445 = const()[name = tensor<string, []>("concat_445"), val = tensor<int32, [3]>([4, 32, 512])];
            tensor<fp16, [4, 1, 32, 512]> transpose_89_cast_fp16 = transpose(perm = transpose_89_perm_0, x = var_3946_cast_fp16)[name = tensor<string, []>("transpose_187")];
            tensor<fp16, [4, 32, 512]> reshape_133_cast_fp16 = reshape(shape = concat_445, x = transpose_89_cast_fp16)[name = tensor<string, []>("reshape_133_cast_fp16")];
            tensor<bool, []> matmul_44_transpose_x_0 = const()[name = tensor<string, []>("matmul_44_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_44_transpose_y_0 = const()[name = tensor<string, []>("matmul_44_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 512]> matmul_44_cast_fp16 = matmul(transpose_x = matmul_44_transpose_x_0, transpose_y = matmul_44_transpose_y_0, x = reshape_132_cast_fp16, y = reshape_133_cast_fp16)[name = tensor<string, []>("matmul_44_cast_fp16")];
            tensor<int32, [4]> concat_449 = const()[name = tensor<string, []>("concat_449"), val = tensor<int32, [4]>([4, 1, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> reshape_134_cast_fp16 = reshape(shape = concat_449, x = matmul_44_cast_fp16)[name = tensor<string, []>("reshape_134_cast_fp16")];
            tensor<int32, [4]> attn_scores_67_perm_0 = const()[name = tensor<string, []>("attn_scores_67_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_3953 = const()[name = tensor<string, []>("op_3953"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 512, 1]> var_3954_cast_fp16 = transpose(perm = var_3953, x = pt_23_cast_fp16)[name = tensor<string, []>("transpose_185")];
            tensor<fp16, [4, 32, 512, 512]> var_3955_cast_fp16 = mul(x = var_3954_cast_fp16, y = core_pos_basis_512_to_fp16)[name = tensor<string, []>("op_3955_cast_fp16")];
            tensor<int32, [1]> pos_scores_23_axes_0 = const()[name = tensor<string, []>("pos_scores_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_23_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_23_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 512, 512]> pos_scores_23_cast_fp16 = reduce_sum(axes = pos_scores_23_axes_0, keep_dims = pos_scores_23_keep_dims_0, x = var_3955_cast_fp16)[name = tensor<string, []>("pos_scores_23_cast_fp16")];
            tensor<int32, [4]> var_3958 = const()[name = tensor<string, []>("op_3958"), val = tensor<int32, [4]>([4, 512, 1, 512])];
            tensor<fp16, [4, 512, 1, 512]> var_3959_cast_fp16 = reshape(shape = var_3958, x = pos_scores_23_cast_fp16)[name = tensor<string, []>("op_3959_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_67_cast_fp16 = transpose(perm = attn_scores_67_perm_0, x = reshape_134_cast_fp16)[name = tensor<string, []>("transpose_186")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_69_cast_fp16 = add(x = attn_scores_67_cast_fp16, y = var_3959_cast_fp16)[name = tensor<string, []>("attn_scores_69_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_scores_71_cast_fp16 = add(x = attn_scores_69_cast_fp16, y = attn_bias_3_cast_fp16)[name = tensor<string, []>("attn_scores_71_cast_fp16")];
            tensor<fp16, [4, 512, 1, 512]> attn_weights_23_cast_fp16 = softmax(axis = var_41, x = attn_scores_71_cast_fp16)[name = tensor<string, []>("attn_weights_23_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_359_cast_fp16 = add(x = input_357_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
            tensor<string, []> x_501_pad_type_0 = const()[name = tensor<string, []>("x_501_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_501_strides_0 = const()[name = tensor<string, []>("x_501_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_501_pad_0 = const()[name = tensor<string, []>("x_501_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_501_dilations_0 = const()[name = tensor<string, []>("x_501_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_501_groups_0 = const()[name = tensor<string, []>("x_501_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_3_stack_layers_3_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(252853440)))];
            tensor<fp16, [1152]> core_stacks_3_stack_layers_3_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254033152)))];
            tensor<fp16, [1, 1152, 1, 512]> x_501_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_feed_forward1_in_proj_bias_to_fp16, dilations = x_501_dilations_0, groups = x_501_groups_0, pad = x_501_pad_0, pad_type = x_501_pad_type_0, strides = x_501_strides_0, weight = core_stacks_3_stack_layers_3_feed_forward1_in_proj_weight_to_fp16, x = input_359_cast_fp16)[name = tensor<string, []>("x_501_cast_fp16")];
            tensor<fp16, []> var_3973_to_fp16 = const()[name = tensor<string, []>("op_3973_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 512]> x_503_cast_fp16 = sub(x = x_501_cast_fp16, y = var_3973_to_fp16)[name = tensor<string, []>("x_503_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3975_cast_fp16 = relu(x = x_503_cast_fp16)[name = tensor<string, []>("op_3975_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3976_cast_fp16 = abs(x = x_503_cast_fp16)[name = tensor<string, []>("op_3976_cast_fp16")];
            tensor<fp16, []> const_86_promoted_to_fp16 = const()[name = tensor<string, []>("const_86_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_3977_cast_fp16 = mul(x = var_3976_cast_fp16, y = const_86_promoted_to_fp16)[name = tensor<string, []>("op_3977_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3978_cast_fp16 = exp(x = var_3977_cast_fp16)[name = tensor<string, []>("op_3978_cast_fp16")];
            tensor<fp32, []> var_3979_epsilon_0 = const()[name = tensor<string, []>("op_3979_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 512]> var_3979_cast_fp16 = log(epsilon = var_3979_epsilon_0, x = var_3978_cast_fp16)[name = tensor<string, []>("op_3979_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3980_cast_fp16 = add(x = var_3975_cast_fp16, y = var_3979_cast_fp16)[name = tensor<string, []>("op_3980_cast_fp16")];
            tensor<fp16, []> var_3981_to_fp16 = const()[name = tensor<string, []>("op_3981_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 512]> var_3982_cast_fp16 = mul(x = x_501_cast_fp16, y = var_3981_to_fp16)[name = tensor<string, []>("op_3982_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 512]> var_3983_cast_fp16 = sub(x = var_3980_cast_fp16, y = var_3982_cast_fp16)[name = tensor<string, []>("op_3983_cast_fp16")];
            tensor<fp16, []> var_3984_to_fp16 = const()[name = tensor<string, []>("op_3984_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 512]> input_361_cast_fp16 = sub(x = var_3983_cast_fp16, y = var_3984_to_fp16)[name = tensor<string, []>("input_361_cast_fp16")];
            tensor<string, []> var_3992_pad_type_0 = const()[name = tensor<string, []>("op_3992_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_3992_strides_0 = const()[name = tensor<string, []>("op_3992_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3992_pad_0 = const()[name = tensor<string, []>("op_3992_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3992_dilations_0 = const()[name = tensor<string, []>("op_3992_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_3992_groups_0 = const()[name = tensor<string, []>("op_3992_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_3_stack_layers_3_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254035520)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_3_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255215232)))];
            tensor<fp16, [1, 512, 1, 512]> var_3992_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_feed_forward1_out_proj_bias_to_fp16, dilations = var_3992_dilations_0, groups = var_3992_groups_0, pad = var_3992_pad_0, pad_type = var_3992_pad_type_0, strides = var_3992_strides_0, weight = core_stacks_3_stack_layers_3_feed_forward1_out_proj_weight_to_fp16, x = input_361_cast_fp16)[name = tensor<string, []>("op_3992_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_363_cast_fp16 = add(x = input_359_cast_fp16, y = var_3992_cast_fp16)[name = tensor<string, []>("input_363_cast_fp16")];
            tensor<int32, [4]> var_3994_begin_0 = const()[name = tensor<string, []>("op_3994_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3994_end_0 = const()[name = tensor<string, []>("op_3994_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_3994_end_mask_0 = const()[name = tensor<string, []>("op_3994_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_3994_cast_fp16 = slice_by_index(begin = var_3994_begin_0, end = var_3994_end_0, end_mask = var_3994_end_mask_0, x = attn_weights_23_cast_fp16)[name = tensor<string, []>("op_3994_cast_fp16")];
            tensor<string, []> x_505_pad_type_0 = const()[name = tensor<string, []>("x_505_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_505_strides_0 = const()[name = tensor<string, []>("x_505_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_505_pad_0 = const()[name = tensor<string, []>("x_505_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_505_dilations_0 = const()[name = tensor<string, []>("x_505_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_505_groups_0 = const()[name = tensor<string, []>("x_505_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_3_stack_layers_3_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255216320)))];
            tensor<fp16, [1152]> core_stacks_3_stack_layers_3_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256396032)))];
            tensor<fp16, [1, 1152, 1, 512]> x_505_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_nonlin_attention_in_proj_bias_to_fp16, dilations = x_505_dilations_0, groups = x_505_groups_0, pad = x_505_pad_0, pad_type = x_505_pad_type_0, strides = x_505_strides_0, weight = core_stacks_3_stack_layers_3_nonlin_attention_in_proj_weight_to_fp16, x = input_363_cast_fp16)[name = tensor<string, []>("x_505_cast_fp16")];
            tensor<int32, [4]> var_4005_begin_0 = const()[name = tensor<string, []>("op_4005_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4005_end_0 = const()[name = tensor<string, []>("op_4005_end_0"), val = tensor<int32, [4]>([1, 384, 1, 512])];
            tensor<bool, [4]> var_4005_end_mask_0 = const()[name = tensor<string, []>("op_4005_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_4005_cast_fp16 = slice_by_index(begin = var_4005_begin_0, end = var_4005_end_0, end_mask = var_4005_end_mask_0, x = x_505_cast_fp16)[name = tensor<string, []>("op_4005_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> s_gate_23_cast_fp16 = tanh(x = var_4005_cast_fp16)[name = tensor<string, []>("s_gate_23_cast_fp16")];
            tensor<int32, [4]> var_4008_begin_0 = const()[name = tensor<string, []>("op_4008_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_4008_end_0 = const()[name = tensor<string, []>("op_4008_end_0"), val = tensor<int32, [4]>([1, 768, 1, 512])];
            tensor<bool, [4]> var_4008_end_mask_0 = const()[name = tensor<string, []>("op_4008_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 512]> var_4008_cast_fp16 = slice_by_index(begin = var_4008_begin_0, end = var_4008_end_0, end_mask = var_4008_end_mask_0, x = x_505_cast_fp16)[name = tensor<string, []>("op_4008_cast_fp16")];
            tensor<fp16, [1, 384, 1, 512]> var_4009_cast_fp16 = mul(x = var_4008_cast_fp16, y = s_gate_23_cast_fp16)[name = tensor<string, []>("op_4009_cast_fp16")];
            tensor<int32, [4]> y_23_begin_0 = const()[name = tensor<string, []>("y_23_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_23_end_0 = const()[name = tensor<string, []>("y_23_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 512])];
            tensor<bool, [4]> y_23_end_mask_0 = const()[name = tensor<string, []>("y_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 512]> y_23_cast_fp16 = slice_by_index(begin = y_23_begin_0, end = y_23_end_0, end_mask = y_23_end_mask_0, x = x_505_cast_fp16)[name = tensor<string, []>("y_23_cast_fp16")];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = tensor<string, []>("transpose_90_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_454 = const()[name = tensor<string, []>("concat_454"), val = tensor<int32, [3]>([1, 512, 512])];
            tensor<fp16, [1, 1, 512, 512]> transpose_90_cast_fp16 = transpose(perm = transpose_90_perm_0, x = var_3994_cast_fp16)[name = tensor<string, []>("transpose_184")];
            tensor<fp16, [1, 512, 512]> reshape_135_cast_fp16 = reshape(shape = concat_454, x = transpose_90_cast_fp16)[name = tensor<string, []>("reshape_135_cast_fp16")];
            tensor<int32, [4]> transpose_91_perm_0 = const()[name = tensor<string, []>("transpose_91_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_455 = const()[name = tensor<string, []>("concat_455"), val = tensor<int32, [3]>([1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> transpose_91_cast_fp16 = transpose(perm = transpose_91_perm_0, x = var_4009_cast_fp16)[name = tensor<string, []>("transpose_183")];
            tensor<fp16, [1, 512, 384]> reshape_136_cast_fp16 = reshape(shape = concat_455, x = transpose_91_cast_fp16)[name = tensor<string, []>("reshape_136_cast_fp16")];
            tensor<bool, []> matmul_45_transpose_x_0 = const()[name = tensor<string, []>("matmul_45_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_45_transpose_y_0 = const()[name = tensor<string, []>("matmul_45_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 512, 384]> matmul_45_cast_fp16 = matmul(transpose_x = matmul_45_transpose_x_0, transpose_y = matmul_45_transpose_y_0, x = reshape_135_cast_fp16, y = reshape_136_cast_fp16)[name = tensor<string, []>("matmul_45_cast_fp16")];
            tensor<int32, [4]> concat_459 = const()[name = tensor<string, []>("concat_459"), val = tensor<int32, [4]>([1, 1, 512, 384])];
            tensor<fp16, [1, 1, 512, 384]> reshape_137_cast_fp16 = reshape(shape = concat_459, x = matmul_45_cast_fp16)[name = tensor<string, []>("reshape_137_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = tensor<string, []>("v_23_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 512]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = reshape_137_cast_fp16)[name = tensor<string, []>("transpose_182")];
            tensor<fp16, [1, 384, 1, 512]> input_365_cast_fp16 = mul(x = v_23_cast_fp16, y = y_23_cast_fp16)[name = tensor<string, []>("input_365_cast_fp16")];
            tensor<string, []> var_4021_pad_type_0 = const()[name = tensor<string, []>("op_4021_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4021_strides_0 = const()[name = tensor<string, []>("op_4021_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4021_pad_0 = const()[name = tensor<string, []>("op_4021_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4021_dilations_0 = const()[name = tensor<string, []>("op_4021_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4021_groups_0 = const()[name = tensor<string, []>("op_4021_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_3_stack_layers_3_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256398400)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_3_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256791680)))];
            tensor<fp16, [1, 512, 1, 512]> var_4021_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_nonlin_attention_out_proj_bias_to_fp16, dilations = var_4021_dilations_0, groups = var_4021_groups_0, pad = var_4021_pad_0, pad_type = var_4021_pad_type_0, strides = var_4021_strides_0, weight = core_stacks_3_stack_layers_3_nonlin_attention_out_proj_weight_to_fp16, x = input_365_cast_fp16)[name = tensor<string, []>("op_4021_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_507_cast_fp16 = add(x = input_363_cast_fp16, y = var_4021_cast_fp16)[name = tensor<string, []>("x_507_cast_fp16")];
            tensor<string, []> var_4032_pad_type_0 = const()[name = tensor<string, []>("op_4032_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4032_strides_0 = const()[name = tensor<string, []>("op_4032_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4032_pad_0 = const()[name = tensor<string, []>("op_4032_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4032_dilations_0 = const()[name = tensor<string, []>("op_4032_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4032_groups_0 = const()[name = tensor<string, []>("op_4032_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_3_stack_layers_3_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256792768)))];
            tensor<fp16, [48]> core_stacks_3_stack_layers_3_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256841984)))];
            tensor<fp16, [1, 48, 1, 512]> var_4032_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_self_attn1_in_proj_bias_to_fp16, dilations = var_4032_dilations_0, groups = var_4032_groups_0, pad = var_4032_pad_0, pad_type = var_4032_pad_type_0, strides = var_4032_strides_0, weight = core_stacks_3_stack_layers_3_self_attn1_in_proj_weight_to_fp16, x = x_507_cast_fp16)[name = tensor<string, []>("op_4032_cast_fp16")];
            tensor<int32, [4]> var_4033 = const()[name = tensor<string, []>("op_4033"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_4034_cast_fp16 = reshape(shape = var_4033, x = var_4032_cast_fp16)[name = tensor<string, []>("op_4034_cast_fp16")];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = tensor<string, []>("transpose_92_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_464 = const()[name = tensor<string, []>("concat_464"), val = tensor<int32, [3]>([4, 512, 512])];
            tensor<fp16, [4, 1, 512, 512]> transpose_92_cast_fp16 = transpose(perm = transpose_92_perm_0, x = attn_weights_23_cast_fp16)[name = tensor<string, []>("transpose_181")];
            tensor<fp16, [4, 512, 512]> reshape_138_cast_fp16 = reshape(shape = concat_464, x = transpose_92_cast_fp16)[name = tensor<string, []>("reshape_138_cast_fp16")];
            tensor<int32, [4]> transpose_93_perm_0 = const()[name = tensor<string, []>("transpose_93_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_465 = const()[name = tensor<string, []>("concat_465"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_93_cast_fp16 = transpose(perm = transpose_93_perm_0, x = var_4034_cast_fp16)[name = tensor<string, []>("transpose_180")];
            tensor<fp16, [4, 512, 12]> reshape_139_cast_fp16 = reshape(shape = concat_465, x = transpose_93_cast_fp16)[name = tensor<string, []>("reshape_139_cast_fp16")];
            tensor<bool, []> matmul_46_transpose_x_0 = const()[name = tensor<string, []>("matmul_46_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_46_transpose_y_0 = const()[name = tensor<string, []>("matmul_46_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_46_cast_fp16 = matmul(transpose_x = matmul_46_transpose_x_0, transpose_y = matmul_46_transpose_y_0, x = reshape_138_cast_fp16, y = reshape_139_cast_fp16)[name = tensor<string, []>("matmul_46_cast_fp16")];
            tensor<int32, [4]> concat_469 = const()[name = tensor<string, []>("concat_469"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_140_cast_fp16 = reshape(shape = concat_469, x = matmul_46_cast_fp16)[name = tensor<string, []>("reshape_140_cast_fp16")];
            tensor<int32, [4]> out_45_perm_0 = const()[name = tensor<string, []>("out_45_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_4037 = const()[name = tensor<string, []>("op_4037"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_45_cast_fp16 = transpose(perm = out_45_perm_0, x = reshape_140_cast_fp16)[name = tensor<string, []>("transpose_179")];
            tensor<fp16, [1, 48, 1, 512]> input_367_cast_fp16 = reshape(shape = var_4037, x = out_45_cast_fp16)[name = tensor<string, []>("input_367_cast_fp16")];
            tensor<string, []> var_4045_pad_type_0 = const()[name = tensor<string, []>("op_4045_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4045_strides_0 = const()[name = tensor<string, []>("op_4045_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4045_pad_0 = const()[name = tensor<string, []>("op_4045_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4045_dilations_0 = const()[name = tensor<string, []>("op_4045_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4045_groups_0 = const()[name = tensor<string, []>("op_4045_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_3_stack_layers_3_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256842176)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_3_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256891392)))];
            tensor<fp16, [1, 512, 1, 512]> var_4045_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_self_attn1_out_proj_bias_to_fp16, dilations = var_4045_dilations_0, groups = var_4045_groups_0, pad = var_4045_pad_0, pad_type = var_4045_pad_type_0, strides = var_4045_strides_0, weight = core_stacks_3_stack_layers_3_self_attn1_out_proj_weight_to_fp16, x = input_367_cast_fp16)[name = tensor<string, []>("op_4045_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_509_cast_fp16 = add(x = x_507_cast_fp16, y = var_4045_cast_fp16)[name = tensor<string, []>("x_509_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_511_cast_fp16 = add(x = x_509_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("x_511_cast_fp16")];
            tensor<string, []> x_513_pad_type_0 = const()[name = tensor<string, []>("x_513_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_513_strides_0 = const()[name = tensor<string, []>("x_513_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_513_pad_0 = const()[name = tensor<string, []>("x_513_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_513_dilations_0 = const()[name = tensor<string, []>("x_513_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_513_groups_0 = const()[name = tensor<string, []>("x_513_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_3_stack_layers_3_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256892480)))];
            tensor<fp16, [1024]> core_stacks_3_stack_layers_3_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257941120)))];
            tensor<fp16, [1, 1024, 1, 512]> x_513_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_conv_module1_in_proj_bias_to_fp16, dilations = x_513_dilations_0, groups = x_513_groups_0, pad = x_513_pad_0, pad_type = x_513_pad_type_0, strides = x_513_strides_0, weight = core_stacks_3_stack_layers_3_conv_module1_in_proj_weight_to_fp16, x = x_511_cast_fp16)[name = tensor<string, []>("x_513_cast_fp16")];
            tensor<int32, [4]> var_4059_begin_0 = const()[name = tensor<string, []>("op_4059_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4059_end_0 = const()[name = tensor<string, []>("op_4059_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_4059_end_mask_0 = const()[name = tensor<string, []>("op_4059_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> 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 = x_513_cast_fp16)[name = tensor<string, []>("op_4059_cast_fp16")];
            tensor<int32, [4]> var_4061_begin_0 = const()[name = tensor<string, []>("op_4061_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_4061_end_0 = const()[name = tensor<string, []>("op_4061_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_4061_end_mask_0 = const()[name = tensor<string, []>("op_4061_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_4061_cast_fp16 = slice_by_index(begin = var_4061_begin_0, end = var_4061_end_0, end_mask = var_4061_end_mask_0, x = x_513_cast_fp16)[name = tensor<string, []>("op_4061_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4062_cast_fp16 = sigmoid(x = var_4061_cast_fp16)[name = tensor<string, []>("op_4062_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_515_cast_fp16 = mul(x = var_4059_cast_fp16, y = var_4062_cast_fp16)[name = tensor<string, []>("x_515_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_369_cast_fp16 = mul(x = x_515_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_369_cast_fp16")];
            tensor<string, []> x_517_pad_type_0 = const()[name = tensor<string, []>("x_517_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_517_pad_0 = const()[name = tensor<string, []>("x_517_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_517_groups_0 = const()[name = tensor<string, []>("x_517_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_517_strides_0 = const()[name = tensor<string, []>("x_517_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_517_dilations_0 = const()[name = tensor<string, []>("x_517_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_3_stack_layers_3_conv_module1_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module1_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257943232)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_3_conv_module1_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module1_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257958656)))];
            tensor<fp16, [1, 512, 1, 512]> x_517_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_conv_module1_dw_bias_to_fp16, dilations = x_517_dilations_0, groups = x_517_groups_0, pad = x_517_pad_0, pad_type = x_517_pad_type_0, strides = x_517_strides_0, weight = core_stacks_3_stack_layers_3_conv_module1_dw_weight_to_fp16, x = input_369_cast_fp16)[name = tensor<string, []>("x_517_cast_fp16")];
            tensor<fp16, []> var_4072_to_fp16 = const()[name = tensor<string, []>("op_4072_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_519_cast_fp16 = sub(x = x_517_cast_fp16, y = var_4072_to_fp16)[name = tensor<string, []>("x_519_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4074_cast_fp16 = relu(x = x_519_cast_fp16)[name = tensor<string, []>("op_4074_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4075_cast_fp16 = abs(x = x_519_cast_fp16)[name = tensor<string, []>("op_4075_cast_fp16")];
            tensor<fp16, []> const_88_promoted_to_fp16 = const()[name = tensor<string, []>("const_88_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_4076_cast_fp16 = mul(x = var_4075_cast_fp16, y = const_88_promoted_to_fp16)[name = tensor<string, []>("op_4076_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4077_cast_fp16 = exp(x = var_4076_cast_fp16)[name = tensor<string, []>("op_4077_cast_fp16")];
            tensor<fp32, []> var_4078_epsilon_0 = const()[name = tensor<string, []>("op_4078_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_4078_cast_fp16 = log(epsilon = var_4078_epsilon_0, x = var_4077_cast_fp16)[name = tensor<string, []>("op_4078_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4079_cast_fp16 = add(x = var_4074_cast_fp16, y = var_4078_cast_fp16)[name = tensor<string, []>("op_4079_cast_fp16")];
            tensor<fp16, []> var_4080_to_fp16 = const()[name = tensor<string, []>("op_4080_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_4081_cast_fp16 = mul(x = x_517_cast_fp16, y = var_4080_to_fp16)[name = tensor<string, []>("op_4081_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4082_cast_fp16 = sub(x = var_4079_cast_fp16, y = var_4081_cast_fp16)[name = tensor<string, []>("op_4082_cast_fp16")];
            tensor<fp16, []> var_4083_to_fp16 = const()[name = tensor<string, []>("op_4083_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_371_cast_fp16 = sub(x = var_4082_cast_fp16, y = var_4083_to_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
            tensor<string, []> var_4091_pad_type_0 = const()[name = tensor<string, []>("op_4091_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4091_strides_0 = const()[name = tensor<string, []>("op_4091_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4091_pad_0 = const()[name = tensor<string, []>("op_4091_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4091_dilations_0 = const()[name = tensor<string, []>("op_4091_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4091_groups_0 = const()[name = tensor<string, []>("op_4091_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_3_stack_layers_3_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257959744)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_3_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258484096)))];
            tensor<fp16, [1, 512, 1, 512]> var_4091_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_conv_module1_out_proj_bias_to_fp16, dilations = var_4091_dilations_0, groups = var_4091_groups_0, pad = var_4091_pad_0, pad_type = var_4091_pad_type_0, strides = var_4091_strides_0, weight = core_stacks_3_stack_layers_3_conv_module1_out_proj_weight_to_fp16, x = input_371_cast_fp16)[name = tensor<string, []>("op_4091_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_373_cast_fp16 = add(x = x_511_cast_fp16, y = var_4091_cast_fp16)[name = tensor<string, []>("input_373_cast_fp16")];
            tensor<string, []> x_521_pad_type_0 = const()[name = tensor<string, []>("x_521_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_521_strides_0 = const()[name = tensor<string, []>("x_521_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_521_pad_0 = const()[name = tensor<string, []>("x_521_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_521_dilations_0 = const()[name = tensor<string, []>("x_521_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_521_groups_0 = const()[name = tensor<string, []>("x_521_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_3_stack_layers_3_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258485184)))];
            tensor<fp16, [1536]> core_stacks_3_stack_layers_3_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260058112)))];
            tensor<fp16, [1, 1536, 1, 512]> x_521_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_feed_forward2_in_proj_bias_to_fp16, dilations = x_521_dilations_0, groups = x_521_groups_0, pad = x_521_pad_0, pad_type = x_521_pad_type_0, strides = x_521_strides_0, weight = core_stacks_3_stack_layers_3_feed_forward2_in_proj_weight_to_fp16, x = input_373_cast_fp16)[name = tensor<string, []>("x_521_cast_fp16")];
            tensor<fp16, []> var_4102_to_fp16 = const()[name = tensor<string, []>("op_4102_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 512]> x_523_cast_fp16 = sub(x = x_521_cast_fp16, y = var_4102_to_fp16)[name = tensor<string, []>("x_523_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_4104_cast_fp16 = relu(x = x_523_cast_fp16)[name = tensor<string, []>("op_4104_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_4105_cast_fp16 = abs(x = x_523_cast_fp16)[name = tensor<string, []>("op_4105_cast_fp16")];
            tensor<fp16, []> const_89_promoted_to_fp16 = const()[name = tensor<string, []>("const_89_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_4106_cast_fp16 = mul(x = var_4105_cast_fp16, y = const_89_promoted_to_fp16)[name = tensor<string, []>("op_4106_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_4107_cast_fp16 = exp(x = var_4106_cast_fp16)[name = tensor<string, []>("op_4107_cast_fp16")];
            tensor<fp32, []> var_4108_epsilon_0 = const()[name = tensor<string, []>("op_4108_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 512]> var_4108_cast_fp16 = log(epsilon = var_4108_epsilon_0, x = var_4107_cast_fp16)[name = tensor<string, []>("op_4108_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_4109_cast_fp16 = add(x = var_4104_cast_fp16, y = var_4108_cast_fp16)[name = tensor<string, []>("op_4109_cast_fp16")];
            tensor<fp16, []> var_4110_to_fp16 = const()[name = tensor<string, []>("op_4110_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 512]> var_4111_cast_fp16 = mul(x = x_521_cast_fp16, y = var_4110_to_fp16)[name = tensor<string, []>("op_4111_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 512]> var_4112_cast_fp16 = sub(x = var_4109_cast_fp16, y = var_4111_cast_fp16)[name = tensor<string, []>("op_4112_cast_fp16")];
            tensor<fp16, []> var_4113_to_fp16 = const()[name = tensor<string, []>("op_4113_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 512]> input_375_cast_fp16 = sub(x = var_4112_cast_fp16, y = var_4113_to_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
            tensor<string, []> var_4121_pad_type_0 = const()[name = tensor<string, []>("op_4121_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4121_strides_0 = const()[name = tensor<string, []>("op_4121_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4121_pad_0 = const()[name = tensor<string, []>("op_4121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4121_dilations_0 = const()[name = tensor<string, []>("op_4121_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4121_groups_0 = const()[name = tensor<string, []>("op_4121_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_3_stack_layers_3_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260061248)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_3_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261634176)))];
            tensor<fp16, [1, 512, 1, 512]> var_4121_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_feed_forward2_out_proj_bias_to_fp16, dilations = var_4121_dilations_0, groups = var_4121_groups_0, pad = var_4121_pad_0, pad_type = var_4121_pad_type_0, strides = var_4121_strides_0, weight = core_stacks_3_stack_layers_3_feed_forward2_out_proj_weight_to_fp16, x = input_375_cast_fp16)[name = tensor<string, []>("op_4121_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> src_49_cast_fp16 = add(x = input_373_cast_fp16, y = var_4121_cast_fp16)[name = tensor<string, []>("src_49_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4124_cast_fp16 = sub(x = src_49_cast_fp16, y = input_357_cast_fp16)[name = tensor<string, []>("op_4124_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_3_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261635264)))];
            tensor<fp16, [1, 512, 1, 512]> var_4125_cast_fp16 = mul(x = var_4124_cast_fp16, y = core_stacks_3_stack_layers_3_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_4125_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_525_cast_fp16 = add(x = input_357_cast_fp16, y = var_4125_cast_fp16)[name = tensor<string, []>("x_525_cast_fp16")];
            tensor<string, []> var_4136_pad_type_0 = const()[name = tensor<string, []>("op_4136_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4136_strides_0 = const()[name = tensor<string, []>("op_4136_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4136_pad_0 = const()[name = tensor<string, []>("op_4136_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4136_dilations_0 = const()[name = tensor<string, []>("op_4136_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4136_groups_0 = const()[name = tensor<string, []>("op_4136_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_3_stack_layers_3_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261636352)))];
            tensor<fp16, [48]> core_stacks_3_stack_layers_3_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261685568)))];
            tensor<fp16, [1, 48, 1, 512]> var_4136_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_self_attn2_in_proj_bias_to_fp16, dilations = var_4136_dilations_0, groups = var_4136_groups_0, pad = var_4136_pad_0, pad_type = var_4136_pad_type_0, strides = var_4136_strides_0, weight = core_stacks_3_stack_layers_3_self_attn2_in_proj_weight_to_fp16, x = x_525_cast_fp16)[name = tensor<string, []>("op_4136_cast_fp16")];
            tensor<int32, [4]> var_4137 = const()[name = tensor<string, []>("op_4137"), val = tensor<int32, [4]>([4, 12, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> var_4138_cast_fp16 = reshape(shape = var_4137, x = var_4136_cast_fp16)[name = tensor<string, []>("op_4138_cast_fp16")];
            tensor<int32, [4]> transpose_95_perm_0 = const()[name = tensor<string, []>("transpose_95_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_475 = const()[name = tensor<string, []>("concat_475"), val = tensor<int32, [3]>([4, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> transpose_95_cast_fp16 = transpose(perm = transpose_95_perm_0, x = var_4138_cast_fp16)[name = tensor<string, []>("transpose_178")];
            tensor<fp16, [4, 512, 12]> reshape_142_cast_fp16 = reshape(shape = concat_475, x = transpose_95_cast_fp16)[name = tensor<string, []>("reshape_142_cast_fp16")];
            tensor<bool, []> matmul_47_transpose_x_0 = const()[name = tensor<string, []>("matmul_47_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_47_transpose_y_0 = const()[name = tensor<string, []>("matmul_47_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 512, 12]> matmul_47_cast_fp16 = matmul(transpose_x = matmul_47_transpose_x_0, transpose_y = matmul_47_transpose_y_0, x = reshape_138_cast_fp16, y = reshape_142_cast_fp16)[name = tensor<string, []>("matmul_47_cast_fp16")];
            tensor<int32, [4]> concat_479 = const()[name = tensor<string, []>("concat_479"), val = tensor<int32, [4]>([4, 1, 512, 12])];
            tensor<fp16, [4, 1, 512, 12]> reshape_143_cast_fp16 = reshape(shape = concat_479, x = matmul_47_cast_fp16)[name = tensor<string, []>("reshape_143_cast_fp16")];
            tensor<int32, [4]> out_47_perm_0 = const()[name = tensor<string, []>("out_47_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_4141 = const()[name = tensor<string, []>("op_4141"), val = tensor<int32, [4]>([1, 48, 1, 512])];
            tensor<fp16, [4, 12, 1, 512]> out_47_cast_fp16 = transpose(perm = out_47_perm_0, x = reshape_143_cast_fp16)[name = tensor<string, []>("transpose_177")];
            tensor<fp16, [1, 48, 1, 512]> input_377_cast_fp16 = reshape(shape = var_4141, x = out_47_cast_fp16)[name = tensor<string, []>("input_377_cast_fp16")];
            tensor<string, []> var_4149_pad_type_0 = const()[name = tensor<string, []>("op_4149_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4149_strides_0 = const()[name = tensor<string, []>("op_4149_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4149_pad_0 = const()[name = tensor<string, []>("op_4149_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4149_dilations_0 = const()[name = tensor<string, []>("op_4149_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4149_groups_0 = const()[name = tensor<string, []>("op_4149_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_3_stack_layers_3_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261685760)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_3_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261734976)))];
            tensor<fp16, [1, 512, 1, 512]> var_4149_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_self_attn2_out_proj_bias_to_fp16, dilations = var_4149_dilations_0, groups = var_4149_groups_0, pad = var_4149_pad_0, pad_type = var_4149_pad_type_0, strides = var_4149_strides_0, weight = core_stacks_3_stack_layers_3_self_attn2_out_proj_weight_to_fp16, x = input_377_cast_fp16)[name = tensor<string, []>("op_4149_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_527_cast_fp16 = add(x = x_525_cast_fp16, y = var_4149_cast_fp16)[name = tensor<string, []>("x_527_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_529_cast_fp16 = add(x = x_527_cast_fp16, y = time_emb_7_cast_fp16)[name = tensor<string, []>("x_529_cast_fp16")];
            tensor<string, []> x_531_pad_type_0 = const()[name = tensor<string, []>("x_531_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_531_strides_0 = const()[name = tensor<string, []>("x_531_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_531_pad_0 = const()[name = tensor<string, []>("x_531_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_531_dilations_0 = const()[name = tensor<string, []>("x_531_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_531_groups_0 = const()[name = tensor<string, []>("x_531_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_3_stack_layers_3_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261736064)))];
            tensor<fp16, [1024]> core_stacks_3_stack_layers_3_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262784704)))];
            tensor<fp16, [1, 1024, 1, 512]> x_531_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_conv_module2_in_proj_bias_to_fp16, dilations = x_531_dilations_0, groups = x_531_groups_0, pad = x_531_pad_0, pad_type = x_531_pad_type_0, strides = x_531_strides_0, weight = core_stacks_3_stack_layers_3_conv_module2_in_proj_weight_to_fp16, x = x_529_cast_fp16)[name = tensor<string, []>("x_531_cast_fp16")];
            tensor<int32, [4]> var_4163_begin_0 = const()[name = tensor<string, []>("op_4163_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4163_end_0 = const()[name = tensor<string, []>("op_4163_end_0"), val = tensor<int32, [4]>([1, 512, 1, 512])];
            tensor<bool, [4]> var_4163_end_mask_0 = const()[name = tensor<string, []>("op_4163_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_4163_cast_fp16 = slice_by_index(begin = var_4163_begin_0, end = var_4163_end_0, end_mask = var_4163_end_mask_0, x = x_531_cast_fp16)[name = tensor<string, []>("op_4163_cast_fp16")];
            tensor<int32, [4]> var_4165_begin_0 = const()[name = tensor<string, []>("op_4165_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_4165_end_0 = const()[name = tensor<string, []>("op_4165_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 512])];
            tensor<bool, [4]> var_4165_end_mask_0 = const()[name = tensor<string, []>("op_4165_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 512]> var_4165_cast_fp16 = slice_by_index(begin = var_4165_begin_0, end = var_4165_end_0, end_mask = var_4165_end_mask_0, x = x_531_cast_fp16)[name = tensor<string, []>("op_4165_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4166_cast_fp16 = sigmoid(x = var_4165_cast_fp16)[name = tensor<string, []>("op_4166_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_533_cast_fp16 = mul(x = var_4163_cast_fp16, y = var_4166_cast_fp16)[name = tensor<string, []>("x_533_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_379_cast_fp16 = mul(x = x_533_cast_fp16, y = conv_keep_3_cast_fp16)[name = tensor<string, []>("input_379_cast_fp16")];
            tensor<string, []> x_535_pad_type_0 = const()[name = tensor<string, []>("x_535_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_535_pad_0 = const()[name = tensor<string, []>("x_535_pad_0"), val = tensor<int32, [4]>([0, 0, 7, 7])];
            tensor<int32, []> x_535_groups_0 = const()[name = tensor<string, []>("x_535_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> x_535_strides_0 = const()[name = tensor<string, []>("x_535_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_535_dilations_0 = const()[name = tensor<string, []>("x_535_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_3_stack_layers_3_conv_module2_dw_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module2_dw_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262786816)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_3_conv_module2_dw_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module2_dw_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262802240)))];
            tensor<fp16, [1, 512, 1, 512]> x_535_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_conv_module2_dw_bias_to_fp16, dilations = x_535_dilations_0, groups = x_535_groups_0, pad = x_535_pad_0, pad_type = x_535_pad_type_0, strides = x_535_strides_0, weight = core_stacks_3_stack_layers_3_conv_module2_dw_weight_to_fp16, x = input_379_cast_fp16)[name = tensor<string, []>("x_535_cast_fp16")];
            tensor<fp16, []> var_4176_to_fp16 = const()[name = tensor<string, []>("op_4176_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> x_537_cast_fp16 = sub(x = x_535_cast_fp16, y = var_4176_to_fp16)[name = tensor<string, []>("x_537_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4178_cast_fp16 = relu(x = x_537_cast_fp16)[name = tensor<string, []>("op_4178_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4179_cast_fp16 = abs(x = x_537_cast_fp16)[name = tensor<string, []>("op_4179_cast_fp16")];
            tensor<fp16, []> const_91_promoted_to_fp16 = const()[name = tensor<string, []>("const_91_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_4180_cast_fp16 = mul(x = var_4179_cast_fp16, y = const_91_promoted_to_fp16)[name = tensor<string, []>("op_4180_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4181_cast_fp16 = exp(x = var_4180_cast_fp16)[name = tensor<string, []>("op_4181_cast_fp16")];
            tensor<fp32, []> var_4182_epsilon_0 = const()[name = tensor<string, []>("op_4182_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 512]> var_4182_cast_fp16 = log(epsilon = var_4182_epsilon_0, x = var_4181_cast_fp16)[name = tensor<string, []>("op_4182_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4183_cast_fp16 = add(x = var_4178_cast_fp16, y = var_4182_cast_fp16)[name = tensor<string, []>("op_4183_cast_fp16")];
            tensor<fp16, []> var_4184_to_fp16 = const()[name = tensor<string, []>("op_4184_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 512]> var_4185_cast_fp16 = mul(x = x_535_cast_fp16, y = var_4184_to_fp16)[name = tensor<string, []>("op_4185_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4186_cast_fp16 = sub(x = var_4183_cast_fp16, y = var_4185_cast_fp16)[name = tensor<string, []>("op_4186_cast_fp16")];
            tensor<fp16, []> var_4187_to_fp16 = const()[name = tensor<string, []>("op_4187_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 512]> input_381_cast_fp16 = sub(x = var_4186_cast_fp16, y = var_4187_to_fp16)[name = tensor<string, []>("input_381_cast_fp16")];
            tensor<string, []> var_4195_pad_type_0 = const()[name = tensor<string, []>("op_4195_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4195_strides_0 = const()[name = tensor<string, []>("op_4195_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4195_pad_0 = const()[name = tensor<string, []>("op_4195_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4195_dilations_0 = const()[name = tensor<string, []>("op_4195_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4195_groups_0 = const()[name = tensor<string, []>("op_4195_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_3_stack_layers_3_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262803328)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_3_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263327680)))];
            tensor<fp16, [1, 512, 1, 512]> var_4195_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_conv_module2_out_proj_bias_to_fp16, dilations = var_4195_dilations_0, groups = var_4195_groups_0, pad = var_4195_pad_0, pad_type = var_4195_pad_type_0, strides = var_4195_strides_0, weight = core_stacks_3_stack_layers_3_conv_module2_out_proj_weight_to_fp16, x = input_381_cast_fp16)[name = tensor<string, []>("op_4195_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_383_cast_fp16 = add(x = x_529_cast_fp16, y = var_4195_cast_fp16)[name = tensor<string, []>("input_383_cast_fp16")];
            tensor<string, []> x_539_pad_type_0 = const()[name = tensor<string, []>("x_539_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_539_strides_0 = const()[name = tensor<string, []>("x_539_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_539_pad_0 = const()[name = tensor<string, []>("x_539_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_539_dilations_0 = const()[name = tensor<string, []>("x_539_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_539_groups_0 = const()[name = tensor<string, []>("x_539_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_3_stack_layers_3_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263328768)))];
            tensor<fp16, [1920]> core_stacks_3_stack_layers_3_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265294912)))];
            tensor<fp16, [1, 1920, 1, 512]> x_539_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_feed_forward3_in_proj_bias_to_fp16, dilations = x_539_dilations_0, groups = x_539_groups_0, pad = x_539_pad_0, pad_type = x_539_pad_type_0, strides = x_539_strides_0, weight = core_stacks_3_stack_layers_3_feed_forward3_in_proj_weight_to_fp16, x = input_383_cast_fp16)[name = tensor<string, []>("x_539_cast_fp16")];
            tensor<fp16, []> var_4206_to_fp16 = const()[name = tensor<string, []>("op_4206_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 512]> x_541_cast_fp16 = sub(x = x_539_cast_fp16, y = var_4206_to_fp16)[name = tensor<string, []>("x_541_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_4208_cast_fp16 = relu(x = x_541_cast_fp16)[name = tensor<string, []>("op_4208_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_4209_cast_fp16 = abs(x = x_541_cast_fp16)[name = tensor<string, []>("op_4209_cast_fp16")];
            tensor<fp16, []> const_92_promoted_to_fp16 = const()[name = tensor<string, []>("const_92_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_4210_cast_fp16 = mul(x = var_4209_cast_fp16, y = const_92_promoted_to_fp16)[name = tensor<string, []>("op_4210_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_4211_cast_fp16 = exp(x = var_4210_cast_fp16)[name = tensor<string, []>("op_4211_cast_fp16")];
            tensor<fp32, []> var_4212_epsilon_0 = const()[name = tensor<string, []>("op_4212_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 512]> var_4212_cast_fp16 = log(epsilon = var_4212_epsilon_0, x = var_4211_cast_fp16)[name = tensor<string, []>("op_4212_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_4213_cast_fp16 = add(x = var_4208_cast_fp16, y = var_4212_cast_fp16)[name = tensor<string, []>("op_4213_cast_fp16")];
            tensor<fp16, []> var_4214_to_fp16 = const()[name = tensor<string, []>("op_4214_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 512]> var_4215_cast_fp16 = mul(x = x_539_cast_fp16, y = var_4214_to_fp16)[name = tensor<string, []>("op_4215_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 512]> var_4216_cast_fp16 = sub(x = var_4213_cast_fp16, y = var_4215_cast_fp16)[name = tensor<string, []>("op_4216_cast_fp16")];
            tensor<fp16, []> var_4217_to_fp16 = const()[name = tensor<string, []>("op_4217_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 512]> input_385_cast_fp16 = sub(x = var_4216_cast_fp16, y = var_4217_to_fp16)[name = tensor<string, []>("input_385_cast_fp16")];
            tensor<string, []> var_4225_pad_type_0 = const()[name = tensor<string, []>("op_4225_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4225_strides_0 = const()[name = tensor<string, []>("op_4225_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4225_pad_0 = const()[name = tensor<string, []>("op_4225_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4225_dilations_0 = const()[name = tensor<string, []>("op_4225_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4225_groups_0 = const()[name = tensor<string, []>("op_4225_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_3_stack_layers_3_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265298816)))];
            tensor<fp16, [512]> core_stacks_3_stack_layers_3_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267264960)))];
            tensor<fp16, [1, 512, 1, 512]> var_4225_cast_fp16 = conv(bias = core_stacks_3_stack_layers_3_feed_forward3_out_proj_bias_to_fp16, dilations = var_4225_dilations_0, groups = var_4225_groups_0, pad = var_4225_pad_0, pad_type = var_4225_pad_type_0, strides = var_4225_strides_0, weight = core_stacks_3_stack_layers_3_feed_forward3_out_proj_weight_to_fp16, x = input_385_cast_fp16)[name = tensor<string, []>("op_4225_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> x_543_cast_fp16 = add(x = input_383_cast_fp16, y = var_4225_cast_fp16)[name = tensor<string, []>("x_543_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_3_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267266048)))];
            tensor<fp16, [1, 512, 1, 512]> d_23_cast_fp16 = sub(x = x_543_cast_fp16, y = core_stacks_3_stack_layers_3_norm_bias_to_fp16)[name = tensor<string, []>("d_23_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4230_cast_fp16 = mul(x = d_23_cast_fp16, y = d_23_cast_fp16)[name = tensor<string, []>("op_4230_cast_fp16")];
            tensor<int32, [1]> var_4232_axes_0 = const()[name = tensor<string, []>("op_4232_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_4232_keep_dims_0 = const()[name = tensor<string, []>("op_4232_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 512]> var_4232_cast_fp16 = reduce_mean(axes = var_4232_axes_0, keep_dims = var_4232_keep_dims_0, x = var_4230_cast_fp16)[name = tensor<string, []>("op_4232_cast_fp16")];
            tensor<fp32, []> scales_23_epsilon_0 = const()[name = tensor<string, []>("scales_23_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 512]> scales_23_cast_fp16 = rsqrt(epsilon = scales_23_epsilon_0, x = var_4232_cast_fp16)[name = tensor<string, []>("scales_23_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4234_cast_fp16 = mul(x = x_543_cast_fp16, y = scales_23_cast_fp16)[name = tensor<string, []>("op_4234_cast_fp16")];
            tensor<fp16, []> core_stacks_3_stack_layers_3_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.aacp-2)];
            tensor<fp16, [1, 512, 1, 512]> src_51_cast_fp16 = mul(x = var_4234_cast_fp16, y = core_stacks_3_stack_layers_3_norm_scale_to_fp16)[name = tensor<string, []>("src_51_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> var_4237_cast_fp16 = sub(x = src_51_cast_fp16, y = input_357_cast_fp16)[name = tensor<string, []>("op_4237_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_stack_layers_3_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_stack_layers_3_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267267136)))];
            tensor<fp16, [1, 512, 1, 512]> var_4238_cast_fp16 = mul(x = var_4237_cast_fp16, y = core_stacks_3_stack_layers_3_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_4238_cast_fp16")];
            tensor<fp16, [1, 512, 1, 512]> input_387_cast_fp16 = add(x = input_357_cast_fp16, y = var_4238_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
            tensor<fp32, []> src_53_scale_factor_height_0 = const()[name = tensor<string, []>("src_53_scale_factor_height_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, []> src_53_scale_factor_width_0 = const()[name = tensor<string, []>("src_53_scale_factor_width_0"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp16, [1, 512, 1, 1024]> src_53_cast_fp16 = upsample_nearest_neighbor(scale_factor_height = src_53_scale_factor_height_0, scale_factor_width = src_53_scale_factor_width_0, x = input_387_cast_fp16)[name = tensor<string, []>("src_53_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4243_cast_fp16 = sub(x = src_53_cast_fp16, y = input_263_cast_fp16)[name = tensor<string, []>("op_4243_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_3_out_combiner_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_3_out_combiner_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267268224)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4244_cast_fp16 = mul(x = var_4243_cast_fp16, y = core_stacks_3_out_combiner_bypass_scale_to_fp16)[name = tensor<string, []>("op_4244_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_391_cast_fp16 = add(x = input_263_cast_fp16, y = var_4244_cast_fp16)[name = tensor<string, []>("input_391_cast_fp16")];
            tensor<string, []> time_emb_pad_type_0 = const()[name = tensor<string, []>("time_emb_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> time_emb_strides_0 = const()[name = tensor<string, []>("time_emb_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> time_emb_pad_0 = const()[name = tensor<string, []>("time_emb_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> time_emb_dilations_0 = const()[name = tensor<string, []>("time_emb_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> time_emb_groups_0 = const()[name = tensor<string, []>("time_emb_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 192, 1, 1]> core_stacks_4_time_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_time_proj_weight_to_fp16"), val = tensor<fp16, [512, 192, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267269312)))];
            tensor<fp16, [512]> core_stacks_4_time_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_time_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267465984)))];
            tensor<fp16, [1, 512, 1, 1]> time_emb_cast_fp16 = conv(bias = core_stacks_4_time_proj_bias_to_fp16, dilations = time_emb_dilations_0, groups = time_emb_groups_0, pad = time_emb_pad_0, pad_type = time_emb_pad_type_0, strides = time_emb_strides_0, weight = core_stacks_4_time_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("time_emb_cast_fp16")];
            tensor<string, []> proj_25_pad_type_0 = const()[name = tensor<string, []>("proj_25_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_25_strides_0 = const()[name = tensor<string, []>("proj_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_25_pad_0 = const()[name = tensor<string, []>("proj_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_25_dilations_0 = const()[name = tensor<string, []>("proj_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_25_groups_0 = const()[name = tensor<string, []>("proj_25_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_4_layers_0_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267467072)))];
            tensor<fp16, [384]> core_stacks_4_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267860352)))];
            tensor<fp16, [1, 384, 1, 1024]> proj_25_cast_fp16 = conv(bias = core_stacks_4_layers_0_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_25_dilations_0, groups = proj_25_groups_0, pad = proj_25_pad_0, pad_type = proj_25_pad_type_0, strides = proj_25_strides_0, weight = core_stacks_4_layers_0_self_attn_weights_in_proj_weight_to_fp16, x = input_391_cast_fp16)[name = tensor<string, []>("proj_25_cast_fp16")];
            tensor<int32, [4]> var_4296_begin_0 = const()[name = tensor<string, []>("op_4296_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4296_end_0 = const()[name = tensor<string, []>("op_4296_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1024])];
            tensor<bool, [4]> var_4296_end_mask_0 = const()[name = tensor<string, []>("op_4296_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_4296_cast_fp16 = slice_by_index(begin = var_4296_begin_0, end = var_4296_end_0, end_mask = var_4296_end_mask_0, x = proj_25_cast_fp16)[name = tensor<string, []>("op_4296_cast_fp16")];
            tensor<int32, [4]> var_4297 = const()[name = tensor<string, []>("op_4297"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_4298_cast_fp16 = reshape(shape = var_4297, x = var_4296_cast_fp16)[name = tensor<string, []>("op_4298_cast_fp16")];
            tensor<int32, [4]> var_4300_begin_0 = const()[name = tensor<string, []>("op_4300_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_4300_end_0 = const()[name = tensor<string, []>("op_4300_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1024])];
            tensor<bool, [4]> var_4300_end_mask_0 = const()[name = tensor<string, []>("op_4300_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_4300_cast_fp16 = slice_by_index(begin = var_4300_begin_0, end = var_4300_end_0, end_mask = var_4300_end_mask_0, x = proj_25_cast_fp16)[name = tensor<string, []>("op_4300_cast_fp16")];
            tensor<int32, [4]> var_4301 = const()[name = tensor<string, []>("op_4301"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_4302_cast_fp16 = reshape(shape = var_4301, x = var_4300_cast_fp16)[name = tensor<string, []>("op_4302_cast_fp16")];
            tensor<int32, [4]> var_4304_begin_0 = const()[name = tensor<string, []>("op_4304_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_4304_end_0 = const()[name = tensor<string, []>("op_4304_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_4304_end_mask_0 = const()[name = tensor<string, []>("op_4304_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_4304_cast_fp16 = slice_by_index(begin = var_4304_begin_0, end = var_4304_end_0, end_mask = var_4304_end_mask_0, x = proj_25_cast_fp16)[name = tensor<string, []>("op_4304_cast_fp16")];
            tensor<int32, [4]> var_4305 = const()[name = tensor<string, []>("op_4305"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> pt_25_cast_fp16 = reshape(shape = var_4305, x = var_4304_cast_fp16)[name = tensor<string, []>("pt_25_cast_fp16")];
            tensor<int32, [4]> transpose_96_perm_0 = const()[name = tensor<string, []>("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_484 = const()[name = tensor<string, []>("concat_484"), val = tensor<int32, [3]>([4, 1024, 32])];
            tensor<fp16, [4, 1, 1024, 32]> transpose_96_cast_fp16 = transpose(perm = transpose_96_perm_0, x = var_4298_cast_fp16)[name = tensor<string, []>("transpose_176")];
            tensor<fp16, [4, 1024, 32]> reshape_144_cast_fp16 = reshape(shape = concat_484, x = transpose_96_cast_fp16)[name = tensor<string, []>("reshape_144_cast_fp16")];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = tensor<string, []>("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_485 = const()[name = tensor<string, []>("concat_485"), val = tensor<int32, [3]>([4, 32, 1024])];
            tensor<fp16, [4, 1, 32, 1024]> transpose_97_cast_fp16 = transpose(perm = transpose_97_perm_0, x = var_4302_cast_fp16)[name = tensor<string, []>("transpose_175")];
            tensor<fp16, [4, 32, 1024]> reshape_145_cast_fp16 = reshape(shape = concat_485, x = transpose_97_cast_fp16)[name = tensor<string, []>("reshape_145_cast_fp16")];
            tensor<bool, []> matmul_48_transpose_x_0 = const()[name = tensor<string, []>("matmul_48_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_48_transpose_y_0 = const()[name = tensor<string, []>("matmul_48_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 1024]> matmul_48_cast_fp16 = matmul(transpose_x = matmul_48_transpose_x_0, transpose_y = matmul_48_transpose_y_0, x = reshape_144_cast_fp16, y = reshape_145_cast_fp16)[name = tensor<string, []>("matmul_48_cast_fp16")];
            tensor<int32, [4]> concat_489 = const()[name = tensor<string, []>("concat_489"), val = tensor<int32, [4]>([4, 1, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> reshape_146_cast_fp16 = reshape(shape = concat_489, x = matmul_48_cast_fp16)[name = tensor<string, []>("reshape_146_cast_fp16")];
            tensor<int32, [4]> attn_scores_73_perm_0 = const()[name = tensor<string, []>("attn_scores_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4309 = const()[name = tensor<string, []>("op_4309"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 1024, 1]> var_4310_cast_fp16 = transpose(perm = var_4309, x = pt_25_cast_fp16)[name = tensor<string, []>("transpose_173")];
            tensor<fp16, [4, 32, 1024, 1024]> var_4311_cast_fp16 = mul(x = var_4310_cast_fp16, y = core_pos_basis_1024_to_fp16)[name = tensor<string, []>("op_4311_cast_fp16")];
            tensor<int32, [1]> pos_scores_25_axes_0 = const()[name = tensor<string, []>("pos_scores_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_25_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_25_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 1024, 1024]> pos_scores_25_cast_fp16 = reduce_sum(axes = pos_scores_25_axes_0, keep_dims = pos_scores_25_keep_dims_0, x = var_4311_cast_fp16)[name = tensor<string, []>("pos_scores_25_cast_fp16")];
            tensor<int32, [4]> var_4314 = const()[name = tensor<string, []>("op_4314"), val = tensor<int32, [4]>([4, 1024, 1, 1024])];
            tensor<fp16, [4, 1024, 1, 1024]> var_4315_cast_fp16 = reshape(shape = var_4314, x = pos_scores_25_cast_fp16)[name = tensor<string, []>("op_4315_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_73_cast_fp16 = transpose(perm = attn_scores_73_perm_0, x = reshape_146_cast_fp16)[name = tensor<string, []>("transpose_174")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_75_cast_fp16 = add(x = attn_scores_73_cast_fp16, y = var_4315_cast_fp16)[name = tensor<string, []>("attn_scores_75_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_77_cast_fp16 = add(x = attn_scores_75_cast_fp16, y = attn_bias_1_cast_fp16)[name = tensor<string, []>("attn_scores_77_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_weights_25_cast_fp16 = softmax(axis = var_41, x = attn_scores_77_cast_fp16)[name = tensor<string, []>("attn_weights_25_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_393_cast_fp16 = add(x = input_391_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("input_393_cast_fp16")];
            tensor<string, []> x_547_pad_type_0 = const()[name = tensor<string, []>("x_547_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_547_strides_0 = const()[name = tensor<string, []>("x_547_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_547_pad_0 = const()[name = tensor<string, []>("x_547_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_547_dilations_0 = const()[name = tensor<string, []>("x_547_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_547_groups_0 = const()[name = tensor<string, []>("x_547_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_4_layers_0_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267861184)))];
            tensor<fp16, [1152]> core_stacks_4_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269040896)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_547_cast_fp16 = conv(bias = core_stacks_4_layers_0_feed_forward1_in_proj_bias_to_fp16, dilations = x_547_dilations_0, groups = x_547_groups_0, pad = x_547_pad_0, pad_type = x_547_pad_type_0, strides = x_547_strides_0, weight = core_stacks_4_layers_0_feed_forward1_in_proj_weight_to_fp16, x = input_393_cast_fp16)[name = tensor<string, []>("x_547_cast_fp16")];
            tensor<fp16, []> var_4329_to_fp16 = const()[name = tensor<string, []>("op_4329_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 1024]> x_549_cast_fp16 = sub(x = x_547_cast_fp16, y = var_4329_to_fp16)[name = tensor<string, []>("x_549_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_4331_cast_fp16 = relu(x = x_549_cast_fp16)[name = tensor<string, []>("op_4331_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_4332_cast_fp16 = abs(x = x_549_cast_fp16)[name = tensor<string, []>("op_4332_cast_fp16")];
            tensor<fp16, []> const_94_promoted_to_fp16 = const()[name = tensor<string, []>("const_94_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_4333_cast_fp16 = mul(x = var_4332_cast_fp16, y = const_94_promoted_to_fp16)[name = tensor<string, []>("op_4333_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_4334_cast_fp16 = exp(x = var_4333_cast_fp16)[name = tensor<string, []>("op_4334_cast_fp16")];
            tensor<fp32, []> var_4335_epsilon_0 = const()[name = tensor<string, []>("op_4335_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_4335_cast_fp16 = log(epsilon = var_4335_epsilon_0, x = var_4334_cast_fp16)[name = tensor<string, []>("op_4335_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_4336_cast_fp16 = add(x = var_4331_cast_fp16, y = var_4335_cast_fp16)[name = tensor<string, []>("op_4336_cast_fp16")];
            tensor<fp16, []> var_4337_to_fp16 = const()[name = tensor<string, []>("op_4337_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 1024]> var_4338_cast_fp16 = mul(x = x_547_cast_fp16, y = var_4337_to_fp16)[name = tensor<string, []>("op_4338_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_4339_cast_fp16 = sub(x = var_4336_cast_fp16, y = var_4338_cast_fp16)[name = tensor<string, []>("op_4339_cast_fp16")];
            tensor<fp16, []> var_4340_to_fp16 = const()[name = tensor<string, []>("op_4340_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 1024]> input_395_cast_fp16 = sub(x = var_4339_cast_fp16, y = var_4340_to_fp16)[name = tensor<string, []>("input_395_cast_fp16")];
            tensor<string, []> var_4348_pad_type_0 = const()[name = tensor<string, []>("op_4348_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4348_strides_0 = const()[name = tensor<string, []>("op_4348_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4348_pad_0 = const()[name = tensor<string, []>("op_4348_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4348_dilations_0 = const()[name = tensor<string, []>("op_4348_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4348_groups_0 = const()[name = tensor<string, []>("op_4348_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_4_layers_0_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269043264)))];
            tensor<fp16, [512]> core_stacks_4_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270222976)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4348_cast_fp16 = conv(bias = core_stacks_4_layers_0_feed_forward1_out_proj_bias_to_fp16, dilations = var_4348_dilations_0, groups = var_4348_groups_0, pad = var_4348_pad_0, pad_type = var_4348_pad_type_0, strides = var_4348_strides_0, weight = core_stacks_4_layers_0_feed_forward1_out_proj_weight_to_fp16, x = input_395_cast_fp16)[name = tensor<string, []>("op_4348_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_397_cast_fp16 = add(x = input_393_cast_fp16, y = var_4348_cast_fp16)[name = tensor<string, []>("input_397_cast_fp16")];
            tensor<int32, [4]> var_4350_begin_0 = const()[name = tensor<string, []>("op_4350_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4350_end_0 = const()[name = tensor<string, []>("op_4350_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_4350_end_mask_0 = const()[name = tensor<string, []>("op_4350_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> var_4350_cast_fp16 = slice_by_index(begin = var_4350_begin_0, end = var_4350_end_0, end_mask = var_4350_end_mask_0, x = attn_weights_25_cast_fp16)[name = tensor<string, []>("op_4350_cast_fp16")];
            tensor<string, []> x_551_pad_type_0 = const()[name = tensor<string, []>("x_551_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_551_strides_0 = const()[name = tensor<string, []>("x_551_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_551_pad_0 = const()[name = tensor<string, []>("x_551_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_551_dilations_0 = const()[name = tensor<string, []>("x_551_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_551_groups_0 = const()[name = tensor<string, []>("x_551_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_4_layers_0_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(270224064)))];
            tensor<fp16, [1152]> core_stacks_4_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271403776)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_551_cast_fp16 = conv(bias = core_stacks_4_layers_0_nonlin_attention_in_proj_bias_to_fp16, dilations = x_551_dilations_0, groups = x_551_groups_0, pad = x_551_pad_0, pad_type = x_551_pad_type_0, strides = x_551_strides_0, weight = core_stacks_4_layers_0_nonlin_attention_in_proj_weight_to_fp16, x = input_397_cast_fp16)[name = tensor<string, []>("x_551_cast_fp16")];
            tensor<int32, [4]> var_4361_begin_0 = const()[name = tensor<string, []>("op_4361_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4361_end_0 = const()[name = tensor<string, []>("op_4361_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_4361_end_mask_0 = const()[name = tensor<string, []>("op_4361_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_4361_cast_fp16 = slice_by_index(begin = var_4361_begin_0, end = var_4361_end_0, end_mask = var_4361_end_mask_0, x = x_551_cast_fp16)[name = tensor<string, []>("op_4361_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> s_gate_25_cast_fp16 = tanh(x = var_4361_cast_fp16)[name = tensor<string, []>("s_gate_25_cast_fp16")];
            tensor<int32, [4]> var_4364_begin_0 = const()[name = tensor<string, []>("op_4364_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_4364_end_0 = const()[name = tensor<string, []>("op_4364_end_0"), val = tensor<int32, [4]>([1, 768, 1, 1024])];
            tensor<bool, [4]> var_4364_end_mask_0 = const()[name = tensor<string, []>("op_4364_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_4364_cast_fp16 = slice_by_index(begin = var_4364_begin_0, end = var_4364_end_0, end_mask = var_4364_end_mask_0, x = x_551_cast_fp16)[name = tensor<string, []>("op_4364_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> var_4365_cast_fp16 = mul(x = var_4364_cast_fp16, y = s_gate_25_cast_fp16)[name = tensor<string, []>("op_4365_cast_fp16")];
            tensor<int32, [4]> y_25_begin_0 = const()[name = tensor<string, []>("y_25_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_25_end_0 = const()[name = tensor<string, []>("y_25_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 1024])];
            tensor<bool, [4]> y_25_end_mask_0 = const()[name = tensor<string, []>("y_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> y_25_cast_fp16 = slice_by_index(begin = y_25_begin_0, end = y_25_end_0, end_mask = y_25_end_mask_0, x = x_551_cast_fp16)[name = tensor<string, []>("y_25_cast_fp16")];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = tensor<string, []>("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_494 = const()[name = tensor<string, []>("concat_494"), val = tensor<int32, [3]>([1, 1024, 1024])];
            tensor<fp16, [1, 1, 1024, 1024]> transpose_98_cast_fp16 = transpose(perm = transpose_98_perm_0, x = var_4350_cast_fp16)[name = tensor<string, []>("transpose_172")];
            tensor<fp16, [1, 1024, 1024]> reshape_147_cast_fp16 = reshape(shape = concat_494, x = transpose_98_cast_fp16)[name = tensor<string, []>("reshape_147_cast_fp16")];
            tensor<int32, [4]> transpose_99_perm_0 = const()[name = tensor<string, []>("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_495 = const()[name = tensor<string, []>("concat_495"), val = tensor<int32, [3]>([1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> transpose_99_cast_fp16 = transpose(perm = transpose_99_perm_0, x = var_4365_cast_fp16)[name = tensor<string, []>("transpose_171")];
            tensor<fp16, [1, 1024, 384]> reshape_148_cast_fp16 = reshape(shape = concat_495, x = transpose_99_cast_fp16)[name = tensor<string, []>("reshape_148_cast_fp16")];
            tensor<bool, []> matmul_49_transpose_x_0 = const()[name = tensor<string, []>("matmul_49_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_49_transpose_y_0 = const()[name = tensor<string, []>("matmul_49_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1024, 384]> matmul_49_cast_fp16 = matmul(transpose_x = matmul_49_transpose_x_0, transpose_y = matmul_49_transpose_y_0, x = reshape_147_cast_fp16, y = reshape_148_cast_fp16)[name = tensor<string, []>("matmul_49_cast_fp16")];
            tensor<int32, [4]> concat_499 = const()[name = tensor<string, []>("concat_499"), val = tensor<int32, [4]>([1, 1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> reshape_149_cast_fp16 = reshape(shape = concat_499, x = matmul_49_cast_fp16)[name = tensor<string, []>("reshape_149_cast_fp16")];
            tensor<int32, [4]> v_25_perm_0 = const()[name = tensor<string, []>("v_25_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 1024]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = reshape_149_cast_fp16)[name = tensor<string, []>("transpose_170")];
            tensor<fp16, [1, 384, 1, 1024]> input_399_cast_fp16 = mul(x = v_25_cast_fp16, y = y_25_cast_fp16)[name = tensor<string, []>("input_399_cast_fp16")];
            tensor<string, []> var_4377_pad_type_0 = const()[name = tensor<string, []>("op_4377_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4377_strides_0 = const()[name = tensor<string, []>("op_4377_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4377_pad_0 = const()[name = tensor<string, []>("op_4377_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4377_dilations_0 = const()[name = tensor<string, []>("op_4377_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4377_groups_0 = const()[name = tensor<string, []>("op_4377_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_4_layers_0_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271406144)))];
            tensor<fp16, [512]> core_stacks_4_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271799424)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4377_cast_fp16 = conv(bias = core_stacks_4_layers_0_nonlin_attention_out_proj_bias_to_fp16, dilations = var_4377_dilations_0, groups = var_4377_groups_0, pad = var_4377_pad_0, pad_type = var_4377_pad_type_0, strides = var_4377_strides_0, weight = core_stacks_4_layers_0_nonlin_attention_out_proj_weight_to_fp16, x = input_399_cast_fp16)[name = tensor<string, []>("op_4377_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_553_cast_fp16 = add(x = input_397_cast_fp16, y = var_4377_cast_fp16)[name = tensor<string, []>("x_553_cast_fp16")];
            tensor<string, []> var_4388_pad_type_0 = const()[name = tensor<string, []>("op_4388_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4388_strides_0 = const()[name = tensor<string, []>("op_4388_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4388_pad_0 = const()[name = tensor<string, []>("op_4388_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4388_dilations_0 = const()[name = tensor<string, []>("op_4388_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4388_groups_0 = const()[name = tensor<string, []>("op_4388_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_4_layers_0_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271800512)))];
            tensor<fp16, [48]> core_stacks_4_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271849728)))];
            tensor<fp16, [1, 48, 1, 1024]> var_4388_cast_fp16 = conv(bias = core_stacks_4_layers_0_self_attn1_in_proj_bias_to_fp16, dilations = var_4388_dilations_0, groups = var_4388_groups_0, pad = var_4388_pad_0, pad_type = var_4388_pad_type_0, strides = var_4388_strides_0, weight = core_stacks_4_layers_0_self_attn1_in_proj_weight_to_fp16, x = x_553_cast_fp16)[name = tensor<string, []>("op_4388_cast_fp16")];
            tensor<int32, [4]> var_4389 = const()[name = tensor<string, []>("op_4389"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_4390_cast_fp16 = reshape(shape = var_4389, x = var_4388_cast_fp16)[name = tensor<string, []>("op_4390_cast_fp16")];
            tensor<int32, [4]> transpose_100_perm_0 = const()[name = tensor<string, []>("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_504 = const()[name = tensor<string, []>("concat_504"), val = tensor<int32, [3]>([4, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> transpose_100_cast_fp16 = transpose(perm = transpose_100_perm_0, x = attn_weights_25_cast_fp16)[name = tensor<string, []>("transpose_169")];
            tensor<fp16, [4, 1024, 1024]> reshape_150_cast_fp16 = reshape(shape = concat_504, x = transpose_100_cast_fp16)[name = tensor<string, []>("reshape_150_cast_fp16")];
            tensor<int32, [4]> transpose_101_perm_0 = const()[name = tensor<string, []>("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_505 = const()[name = tensor<string, []>("concat_505"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_101_cast_fp16 = transpose(perm = transpose_101_perm_0, x = var_4390_cast_fp16)[name = tensor<string, []>("transpose_168")];
            tensor<fp16, [4, 1024, 12]> reshape_151_cast_fp16 = reshape(shape = concat_505, x = transpose_101_cast_fp16)[name = tensor<string, []>("reshape_151_cast_fp16")];
            tensor<bool, []> matmul_50_transpose_x_0 = const()[name = tensor<string, []>("matmul_50_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_50_transpose_y_0 = const()[name = tensor<string, []>("matmul_50_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_50_cast_fp16 = matmul(transpose_x = matmul_50_transpose_x_0, transpose_y = matmul_50_transpose_y_0, x = reshape_150_cast_fp16, y = reshape_151_cast_fp16)[name = tensor<string, []>("matmul_50_cast_fp16")];
            tensor<int32, [4]> concat_509 = const()[name = tensor<string, []>("concat_509"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_152_cast_fp16 = reshape(shape = concat_509, x = matmul_50_cast_fp16)[name = tensor<string, []>("reshape_152_cast_fp16")];
            tensor<int32, [4]> out_49_perm_0 = const()[name = tensor<string, []>("out_49_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_4393 = const()[name = tensor<string, []>("op_4393"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_49_cast_fp16 = transpose(perm = out_49_perm_0, x = reshape_152_cast_fp16)[name = tensor<string, []>("transpose_167")];
            tensor<fp16, [1, 48, 1, 1024]> input_401_cast_fp16 = reshape(shape = var_4393, x = out_49_cast_fp16)[name = tensor<string, []>("input_401_cast_fp16")];
            tensor<string, []> var_4401_pad_type_0 = const()[name = tensor<string, []>("op_4401_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4401_strides_0 = const()[name = tensor<string, []>("op_4401_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4401_pad_0 = const()[name = tensor<string, []>("op_4401_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4401_dilations_0 = const()[name = tensor<string, []>("op_4401_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4401_groups_0 = const()[name = tensor<string, []>("op_4401_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_4_layers_0_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271849920)))];
            tensor<fp16, [512]> core_stacks_4_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271899136)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4401_cast_fp16 = conv(bias = core_stacks_4_layers_0_self_attn1_out_proj_bias_to_fp16, dilations = var_4401_dilations_0, groups = var_4401_groups_0, pad = var_4401_pad_0, pad_type = var_4401_pad_type_0, strides = var_4401_strides_0, weight = core_stacks_4_layers_0_self_attn1_out_proj_weight_to_fp16, x = input_401_cast_fp16)[name = tensor<string, []>("op_4401_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_555_cast_fp16 = add(x = x_553_cast_fp16, y = var_4401_cast_fp16)[name = tensor<string, []>("x_555_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_557_cast_fp16 = add(x = x_555_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("x_557_cast_fp16")];
            tensor<string, []> x_559_pad_type_0 = const()[name = tensor<string, []>("x_559_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_559_strides_0 = const()[name = tensor<string, []>("x_559_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_559_pad_0 = const()[name = tensor<string, []>("x_559_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_559_dilations_0 = const()[name = tensor<string, []>("x_559_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_559_groups_0 = const()[name = tensor<string, []>("x_559_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_4_layers_0_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271900224)))];
            tensor<fp16, [1024]> core_stacks_4_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272948864)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_559_cast_fp16 = conv(bias = core_stacks_4_layers_0_conv_module1_in_proj_bias_to_fp16, dilations = x_559_dilations_0, groups = x_559_groups_0, pad = x_559_pad_0, pad_type = x_559_pad_type_0, strides = x_559_strides_0, weight = core_stacks_4_layers_0_conv_module1_in_proj_weight_to_fp16, x = x_557_cast_fp16)[name = tensor<string, []>("x_559_cast_fp16")];
            tensor<int32, [4]> var_4419_begin_0 = const()[name = tensor<string, []>("op_4419_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4419_end_0 = const()[name = tensor<string, []>("op_4419_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_4419_end_mask_0 = const()[name = tensor<string, []>("op_4419_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4419_cast_fp16 = slice_by_index(begin = var_4419_begin_0, end = var_4419_end_0, end_mask = var_4419_end_mask_0, x = x_559_cast_fp16)[name = tensor<string, []>("op_4419_cast_fp16")];
            tensor<int32, [4]> var_4421_begin_0 = const()[name = tensor<string, []>("op_4421_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_4421_end_0 = const()[name = tensor<string, []>("op_4421_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_4421_end_mask_0 = const()[name = tensor<string, []>("op_4421_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4421_cast_fp16 = slice_by_index(begin = var_4421_begin_0, end = var_4421_end_0, end_mask = var_4421_end_mask_0, x = x_559_cast_fp16)[name = tensor<string, []>("op_4421_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4422_cast_fp16 = sigmoid(x = var_4421_cast_fp16)[name = tensor<string, []>("op_4422_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_561_cast_fp16 = mul(x = var_4419_cast_fp16, y = var_4422_cast_fp16)[name = tensor<string, []>("x_561_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_403_cast_fp16 = mul(x = x_561_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_403_cast_fp16")];
            tensor<string, []> var_4431_pad_type_0 = const()[name = tensor<string, []>("op_4431_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4431_pad_0 = const()[name = tensor<string, []>("op_4431_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4431_groups_0 = const()[name = tensor<string, []>("op_4431_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4431_strides_0 = const()[name = tensor<string, []>("op_4431_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4431_dilations_0 = const()[name = tensor<string, []>("op_4431_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_0_conv_module1_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module1_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272950976)))];
            tensor<fp16, [512]> core_stacks_4_layers_0_conv_module1_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module1_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272966400)))];
            tensor<fp16, [1, 512, 1, 1040]> var_4431_cast_fp16 = conv(bias = core_stacks_4_layers_0_conv_module1_dw_a_bias_to_fp16, dilations = var_4431_dilations_0, groups = var_4431_groups_0, pad = var_4431_pad_0, pad_type = var_4431_pad_type_0, strides = var_4431_strides_0, weight = core_stacks_4_layers_0_conv_module1_dw_a_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("op_4431_cast_fp16")];
            tensor<int32, [4]> var_4432_begin_0 = const()[name = tensor<string, []>("op_4432_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4432_end_0 = const()[name = tensor<string, []>("op_4432_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_4432_end_mask_0 = const()[name = tensor<string, []>("op_4432_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_4432_cast_fp16 = slice_by_index(begin = var_4432_begin_0, end = var_4432_end_0, end_mask = var_4432_end_mask_0, x = var_4431_cast_fp16)[name = tensor<string, []>("op_4432_cast_fp16")];
            tensor<string, []> var_4438_pad_type_0 = const()[name = tensor<string, []>("op_4438_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4438_pad_0 = const()[name = tensor<string, []>("op_4438_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4438_groups_0 = const()[name = tensor<string, []>("op_4438_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4438_strides_0 = const()[name = tensor<string, []>("op_4438_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4438_dilations_0 = const()[name = tensor<string, []>("op_4438_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_0_conv_module1_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module1_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272967488)))];
            tensor<fp16, [1, 512, 1, 1040]> var_4438_cast_fp16 = conv(dilations = var_4438_dilations_0, groups = var_4438_groups_0, pad = var_4438_pad_0, pad_type = var_4438_pad_type_0, strides = var_4438_strides_0, weight = core_stacks_4_layers_0_conv_module1_dw_b_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("op_4438_cast_fp16")];
            tensor<int32, [4]> var_4442_begin_0 = const()[name = tensor<string, []>("op_4442_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_4442_end_0 = const()[name = tensor<string, []>("op_4442_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_4442_end_mask_0 = const()[name = tensor<string, []>("op_4442_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_4442_cast_fp16 = slice_by_index(begin = var_4442_begin_0, end = var_4442_end_0, end_mask = var_4442_end_mask_0, x = var_4438_cast_fp16)[name = tensor<string, []>("op_4442_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4443_cast_fp16 = add(x = var_4432_cast_fp16, y = var_4442_cast_fp16)[name = tensor<string, []>("op_4443_cast_fp16")];
            tensor<string, []> var_4449_pad_type_0 = const()[name = tensor<string, []>("op_4449_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4449_pad_0 = const()[name = tensor<string, []>("op_4449_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4449_groups_0 = const()[name = tensor<string, []>("op_4449_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4449_strides_0 = const()[name = tensor<string, []>("op_4449_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4449_dilations_0 = const()[name = tensor<string, []>("op_4449_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_4_layers_0_conv_module1_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module1_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272982912)))];
            tensor<fp16, [1, 512, 1, 1054]> var_4449_cast_fp16 = conv(dilations = var_4449_dilations_0, groups = var_4449_groups_0, pad = var_4449_pad_0, pad_type = var_4449_pad_type_0, strides = var_4449_strides_0, weight = core_stacks_4_layers_0_conv_module1_dw_c_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("op_4449_cast_fp16")];
            tensor<int32, [4]> var_4453_begin_0 = const()[name = tensor<string, []>("op_4453_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_4453_end_0 = const()[name = tensor<string, []>("op_4453_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_4453_end_mask_0 = const()[name = tensor<string, []>("op_4453_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4453_cast_fp16 = slice_by_index(begin = var_4453_begin_0, end = var_4453_end_0, end_mask = var_4453_end_mask_0, x = var_4449_cast_fp16)[name = tensor<string, []>("op_4453_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_563_cast_fp16 = add(x = var_4443_cast_fp16, y = var_4453_cast_fp16)[name = tensor<string, []>("x_563_cast_fp16")];
            tensor<fp16, []> var_4455_to_fp16 = const()[name = tensor<string, []>("op_4455_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_565_cast_fp16 = sub(x = x_563_cast_fp16, y = var_4455_to_fp16)[name = tensor<string, []>("x_565_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4457_cast_fp16 = relu(x = x_565_cast_fp16)[name = tensor<string, []>("op_4457_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4458_cast_fp16 = abs(x = x_565_cast_fp16)[name = tensor<string, []>("op_4458_cast_fp16")];
            tensor<fp16, []> const_97_promoted_to_fp16 = const()[name = tensor<string, []>("const_97_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_4459_cast_fp16 = mul(x = var_4458_cast_fp16, y = const_97_promoted_to_fp16)[name = tensor<string, []>("op_4459_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4460_cast_fp16 = exp(x = var_4459_cast_fp16)[name = tensor<string, []>("op_4460_cast_fp16")];
            tensor<fp32, []> var_4461_epsilon_0 = const()[name = tensor<string, []>("op_4461_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_4461_cast_fp16 = log(epsilon = var_4461_epsilon_0, x = var_4460_cast_fp16)[name = tensor<string, []>("op_4461_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4462_cast_fp16 = add(x = var_4457_cast_fp16, y = var_4461_cast_fp16)[name = tensor<string, []>("op_4462_cast_fp16")];
            tensor<fp16, []> var_4463_to_fp16 = const()[name = tensor<string, []>("op_4463_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_4464_cast_fp16 = mul(x = x_563_cast_fp16, y = var_4463_to_fp16)[name = tensor<string, []>("op_4464_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4465_cast_fp16 = sub(x = var_4462_cast_fp16, y = var_4464_cast_fp16)[name = tensor<string, []>("op_4465_cast_fp16")];
            tensor<fp16, []> var_4466_to_fp16 = const()[name = tensor<string, []>("op_4466_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_405_cast_fp16 = sub(x = var_4465_cast_fp16, y = var_4466_to_fp16)[name = tensor<string, []>("input_405_cast_fp16")];
            tensor<string, []> var_4474_pad_type_0 = const()[name = tensor<string, []>("op_4474_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4474_strides_0 = const()[name = tensor<string, []>("op_4474_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4474_pad_0 = const()[name = tensor<string, []>("op_4474_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4474_dilations_0 = const()[name = tensor<string, []>("op_4474_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4474_groups_0 = const()[name = tensor<string, []>("op_4474_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_4_layers_0_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272984000)))];
            tensor<fp16, [512]> core_stacks_4_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273508352)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4474_cast_fp16 = conv(bias = core_stacks_4_layers_0_conv_module1_out_proj_bias_to_fp16, dilations = var_4474_dilations_0, groups = var_4474_groups_0, pad = var_4474_pad_0, pad_type = var_4474_pad_type_0, strides = var_4474_strides_0, weight = core_stacks_4_layers_0_conv_module1_out_proj_weight_to_fp16, x = input_405_cast_fp16)[name = tensor<string, []>("op_4474_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_407_cast_fp16 = add(x = x_557_cast_fp16, y = var_4474_cast_fp16)[name = tensor<string, []>("input_407_cast_fp16")];
            tensor<string, []> x_567_pad_type_0 = const()[name = tensor<string, []>("x_567_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_567_strides_0 = const()[name = tensor<string, []>("x_567_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_567_pad_0 = const()[name = tensor<string, []>("x_567_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_567_dilations_0 = const()[name = tensor<string, []>("x_567_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_567_groups_0 = const()[name = tensor<string, []>("x_567_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_4_layers_0_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273509440)))];
            tensor<fp16, [1536]> core_stacks_4_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275082368)))];
            tensor<fp16, [1, 1536, 1, 1024]> x_567_cast_fp16 = conv(bias = core_stacks_4_layers_0_feed_forward2_in_proj_bias_to_fp16, dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = core_stacks_4_layers_0_feed_forward2_in_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor<string, []>("x_567_cast_fp16")];
            tensor<fp16, []> var_4485_to_fp16 = const()[name = tensor<string, []>("op_4485_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 1024]> x_569_cast_fp16 = sub(x = x_567_cast_fp16, y = var_4485_to_fp16)[name = tensor<string, []>("x_569_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_4487_cast_fp16 = relu(x = x_569_cast_fp16)[name = tensor<string, []>("op_4487_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_4488_cast_fp16 = abs(x = x_569_cast_fp16)[name = tensor<string, []>("op_4488_cast_fp16")];
            tensor<fp16, []> const_98_promoted_to_fp16 = const()[name = tensor<string, []>("const_98_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_4489_cast_fp16 = mul(x = var_4488_cast_fp16, y = const_98_promoted_to_fp16)[name = tensor<string, []>("op_4489_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_4490_cast_fp16 = exp(x = var_4489_cast_fp16)[name = tensor<string, []>("op_4490_cast_fp16")];
            tensor<fp32, []> var_4491_epsilon_0 = const()[name = tensor<string, []>("op_4491_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_4491_cast_fp16 = log(epsilon = var_4491_epsilon_0, x = var_4490_cast_fp16)[name = tensor<string, []>("op_4491_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_4492_cast_fp16 = add(x = var_4487_cast_fp16, y = var_4491_cast_fp16)[name = tensor<string, []>("op_4492_cast_fp16")];
            tensor<fp16, []> var_4493_to_fp16 = const()[name = tensor<string, []>("op_4493_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 1024]> var_4494_cast_fp16 = mul(x = x_567_cast_fp16, y = var_4493_to_fp16)[name = tensor<string, []>("op_4494_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_4495_cast_fp16 = sub(x = var_4492_cast_fp16, y = var_4494_cast_fp16)[name = tensor<string, []>("op_4495_cast_fp16")];
            tensor<fp16, []> var_4496_to_fp16 = const()[name = tensor<string, []>("op_4496_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 1024]> input_409_cast_fp16 = sub(x = var_4495_cast_fp16, y = var_4496_to_fp16)[name = tensor<string, []>("input_409_cast_fp16")];
            tensor<string, []> var_4504_pad_type_0 = const()[name = tensor<string, []>("op_4504_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4504_strides_0 = const()[name = tensor<string, []>("op_4504_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4504_pad_0 = const()[name = tensor<string, []>("op_4504_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4504_dilations_0 = const()[name = tensor<string, []>("op_4504_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4504_groups_0 = const()[name = tensor<string, []>("op_4504_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_4_layers_0_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275085504)))];
            tensor<fp16, [512]> core_stacks_4_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276658432)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4504_cast_fp16 = conv(bias = core_stacks_4_layers_0_feed_forward2_out_proj_bias_to_fp16, dilations = var_4504_dilations_0, groups = var_4504_groups_0, pad = var_4504_pad_0, pad_type = var_4504_pad_type_0, strides = var_4504_strides_0, weight = core_stacks_4_layers_0_feed_forward2_out_proj_weight_to_fp16, x = input_409_cast_fp16)[name = tensor<string, []>("op_4504_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> src_55_cast_fp16 = add(x = input_407_cast_fp16, y = var_4504_cast_fp16)[name = tensor<string, []>("src_55_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4507_cast_fp16 = sub(x = src_55_cast_fp16, y = input_391_cast_fp16)[name = tensor<string, []>("op_4507_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276659520)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4508_cast_fp16 = mul(x = var_4507_cast_fp16, y = core_stacks_4_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_4508_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_571_cast_fp16 = add(x = input_391_cast_fp16, y = var_4508_cast_fp16)[name = tensor<string, []>("x_571_cast_fp16")];
            tensor<string, []> var_4519_pad_type_0 = const()[name = tensor<string, []>("op_4519_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4519_strides_0 = const()[name = tensor<string, []>("op_4519_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4519_pad_0 = const()[name = tensor<string, []>("op_4519_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4519_dilations_0 = const()[name = tensor<string, []>("op_4519_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4519_groups_0 = const()[name = tensor<string, []>("op_4519_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_4_layers_0_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276660608)))];
            tensor<fp16, [48]> core_stacks_4_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276709824)))];
            tensor<fp16, [1, 48, 1, 1024]> var_4519_cast_fp16 = conv(bias = core_stacks_4_layers_0_self_attn2_in_proj_bias_to_fp16, dilations = var_4519_dilations_0, groups = var_4519_groups_0, pad = var_4519_pad_0, pad_type = var_4519_pad_type_0, strides = var_4519_strides_0, weight = core_stacks_4_layers_0_self_attn2_in_proj_weight_to_fp16, x = x_571_cast_fp16)[name = tensor<string, []>("op_4519_cast_fp16")];
            tensor<int32, [4]> var_4520 = const()[name = tensor<string, []>("op_4520"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_4521_cast_fp16 = reshape(shape = var_4520, x = var_4519_cast_fp16)[name = tensor<string, []>("op_4521_cast_fp16")];
            tensor<int32, [4]> transpose_103_perm_0 = const()[name = tensor<string, []>("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_515 = const()[name = tensor<string, []>("concat_515"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_103_cast_fp16 = transpose(perm = transpose_103_perm_0, x = var_4521_cast_fp16)[name = tensor<string, []>("transpose_166")];
            tensor<fp16, [4, 1024, 12]> reshape_154_cast_fp16 = reshape(shape = concat_515, x = transpose_103_cast_fp16)[name = tensor<string, []>("reshape_154_cast_fp16")];
            tensor<bool, []> matmul_51_transpose_x_0 = const()[name = tensor<string, []>("matmul_51_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_51_transpose_y_0 = const()[name = tensor<string, []>("matmul_51_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_51_cast_fp16 = matmul(transpose_x = matmul_51_transpose_x_0, transpose_y = matmul_51_transpose_y_0, x = reshape_150_cast_fp16, y = reshape_154_cast_fp16)[name = tensor<string, []>("matmul_51_cast_fp16")];
            tensor<int32, [4]> concat_519 = const()[name = tensor<string, []>("concat_519"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_155_cast_fp16 = reshape(shape = concat_519, x = matmul_51_cast_fp16)[name = tensor<string, []>("reshape_155_cast_fp16")];
            tensor<int32, [4]> out_51_perm_0 = const()[name = tensor<string, []>("out_51_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_4524 = const()[name = tensor<string, []>("op_4524"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_51_cast_fp16 = transpose(perm = out_51_perm_0, x = reshape_155_cast_fp16)[name = tensor<string, []>("transpose_165")];
            tensor<fp16, [1, 48, 1, 1024]> input_411_cast_fp16 = reshape(shape = var_4524, x = out_51_cast_fp16)[name = tensor<string, []>("input_411_cast_fp16")];
            tensor<string, []> var_4532_pad_type_0 = const()[name = tensor<string, []>("op_4532_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4532_strides_0 = const()[name = tensor<string, []>("op_4532_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4532_pad_0 = const()[name = tensor<string, []>("op_4532_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4532_dilations_0 = const()[name = tensor<string, []>("op_4532_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4532_groups_0 = const()[name = tensor<string, []>("op_4532_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_4_layers_0_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276710016)))];
            tensor<fp16, [512]> core_stacks_4_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276759232)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4532_cast_fp16 = conv(bias = core_stacks_4_layers_0_self_attn2_out_proj_bias_to_fp16, dilations = var_4532_dilations_0, groups = var_4532_groups_0, pad = var_4532_pad_0, pad_type = var_4532_pad_type_0, strides = var_4532_strides_0, weight = core_stacks_4_layers_0_self_attn2_out_proj_weight_to_fp16, x = input_411_cast_fp16)[name = tensor<string, []>("op_4532_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_573_cast_fp16 = add(x = x_571_cast_fp16, y = var_4532_cast_fp16)[name = tensor<string, []>("x_573_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_575_cast_fp16 = add(x = x_573_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("x_575_cast_fp16")];
            tensor<string, []> x_577_pad_type_0 = const()[name = tensor<string, []>("x_577_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_577_strides_0 = const()[name = tensor<string, []>("x_577_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_577_pad_0 = const()[name = tensor<string, []>("x_577_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_577_dilations_0 = const()[name = tensor<string, []>("x_577_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_577_groups_0 = const()[name = tensor<string, []>("x_577_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_4_layers_0_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(276760320)))];
            tensor<fp16, [1024]> core_stacks_4_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277808960)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_577_cast_fp16 = conv(bias = core_stacks_4_layers_0_conv_module2_in_proj_bias_to_fp16, dilations = x_577_dilations_0, groups = x_577_groups_0, pad = x_577_pad_0, pad_type = x_577_pad_type_0, strides = x_577_strides_0, weight = core_stacks_4_layers_0_conv_module2_in_proj_weight_to_fp16, x = x_575_cast_fp16)[name = tensor<string, []>("x_577_cast_fp16")];
            tensor<int32, [4]> var_4550_begin_0 = const()[name = tensor<string, []>("op_4550_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4550_end_0 = const()[name = tensor<string, []>("op_4550_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_4550_end_mask_0 = const()[name = tensor<string, []>("op_4550_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4550_cast_fp16 = slice_by_index(begin = var_4550_begin_0, end = var_4550_end_0, end_mask = var_4550_end_mask_0, x = x_577_cast_fp16)[name = tensor<string, []>("op_4550_cast_fp16")];
            tensor<int32, [4]> var_4552_begin_0 = const()[name = tensor<string, []>("op_4552_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_4552_end_0 = const()[name = tensor<string, []>("op_4552_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_4552_end_mask_0 = const()[name = tensor<string, []>("op_4552_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4552_cast_fp16 = slice_by_index(begin = var_4552_begin_0, end = var_4552_end_0, end_mask = var_4552_end_mask_0, x = x_577_cast_fp16)[name = tensor<string, []>("op_4552_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4553_cast_fp16 = sigmoid(x = var_4552_cast_fp16)[name = tensor<string, []>("op_4553_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_579_cast_fp16 = mul(x = var_4550_cast_fp16, y = var_4553_cast_fp16)[name = tensor<string, []>("x_579_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_413_cast_fp16 = mul(x = x_579_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_413_cast_fp16")];
            tensor<string, []> var_4562_pad_type_0 = const()[name = tensor<string, []>("op_4562_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4562_pad_0 = const()[name = tensor<string, []>("op_4562_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4562_groups_0 = const()[name = tensor<string, []>("op_4562_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4562_strides_0 = const()[name = tensor<string, []>("op_4562_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4562_dilations_0 = const()[name = tensor<string, []>("op_4562_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_0_conv_module2_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module2_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277811072)))];
            tensor<fp16, [512]> core_stacks_4_layers_0_conv_module2_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module2_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277826496)))];
            tensor<fp16, [1, 512, 1, 1040]> var_4562_cast_fp16 = conv(bias = core_stacks_4_layers_0_conv_module2_dw_a_bias_to_fp16, dilations = var_4562_dilations_0, groups = var_4562_groups_0, pad = var_4562_pad_0, pad_type = var_4562_pad_type_0, strides = var_4562_strides_0, weight = core_stacks_4_layers_0_conv_module2_dw_a_weight_to_fp16, x = input_413_cast_fp16)[name = tensor<string, []>("op_4562_cast_fp16")];
            tensor<int32, [4]> var_4563_begin_0 = const()[name = tensor<string, []>("op_4563_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4563_end_0 = const()[name = tensor<string, []>("op_4563_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_4563_end_mask_0 = const()[name = tensor<string, []>("op_4563_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_4563_cast_fp16 = slice_by_index(begin = var_4563_begin_0, end = var_4563_end_0, end_mask = var_4563_end_mask_0, x = var_4562_cast_fp16)[name = tensor<string, []>("op_4563_cast_fp16")];
            tensor<string, []> var_4569_pad_type_0 = const()[name = tensor<string, []>("op_4569_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4569_pad_0 = const()[name = tensor<string, []>("op_4569_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4569_groups_0 = const()[name = tensor<string, []>("op_4569_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4569_strides_0 = const()[name = tensor<string, []>("op_4569_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4569_dilations_0 = const()[name = tensor<string, []>("op_4569_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_0_conv_module2_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module2_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277827584)))];
            tensor<fp16, [1, 512, 1, 1040]> var_4569_cast_fp16 = conv(dilations = var_4569_dilations_0, groups = var_4569_groups_0, pad = var_4569_pad_0, pad_type = var_4569_pad_type_0, strides = var_4569_strides_0, weight = core_stacks_4_layers_0_conv_module2_dw_b_weight_to_fp16, x = input_413_cast_fp16)[name = tensor<string, []>("op_4569_cast_fp16")];
            tensor<int32, [4]> var_4573_begin_0 = const()[name = tensor<string, []>("op_4573_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_4573_end_0 = const()[name = tensor<string, []>("op_4573_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_4573_end_mask_0 = const()[name = tensor<string, []>("op_4573_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_4573_cast_fp16 = slice_by_index(begin = var_4573_begin_0, end = var_4573_end_0, end_mask = var_4573_end_mask_0, x = var_4569_cast_fp16)[name = tensor<string, []>("op_4573_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4574_cast_fp16 = add(x = var_4563_cast_fp16, y = var_4573_cast_fp16)[name = tensor<string, []>("op_4574_cast_fp16")];
            tensor<string, []> var_4580_pad_type_0 = const()[name = tensor<string, []>("op_4580_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4580_pad_0 = const()[name = tensor<string, []>("op_4580_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4580_groups_0 = const()[name = tensor<string, []>("op_4580_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4580_strides_0 = const()[name = tensor<string, []>("op_4580_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4580_dilations_0 = const()[name = tensor<string, []>("op_4580_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_4_layers_0_conv_module2_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module2_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277843008)))];
            tensor<fp16, [1, 512, 1, 1054]> var_4580_cast_fp16 = conv(dilations = var_4580_dilations_0, groups = var_4580_groups_0, pad = var_4580_pad_0, pad_type = var_4580_pad_type_0, strides = var_4580_strides_0, weight = core_stacks_4_layers_0_conv_module2_dw_c_weight_to_fp16, x = input_413_cast_fp16)[name = tensor<string, []>("op_4580_cast_fp16")];
            tensor<int32, [4]> var_4584_begin_0 = const()[name = tensor<string, []>("op_4584_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_4584_end_0 = const()[name = tensor<string, []>("op_4584_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_4584_end_mask_0 = const()[name = tensor<string, []>("op_4584_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4584_cast_fp16 = slice_by_index(begin = var_4584_begin_0, end = var_4584_end_0, end_mask = var_4584_end_mask_0, x = var_4580_cast_fp16)[name = tensor<string, []>("op_4584_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_581_cast_fp16 = add(x = var_4574_cast_fp16, y = var_4584_cast_fp16)[name = tensor<string, []>("x_581_cast_fp16")];
            tensor<fp16, []> var_4586_to_fp16 = const()[name = tensor<string, []>("op_4586_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_583_cast_fp16 = sub(x = x_581_cast_fp16, y = var_4586_to_fp16)[name = tensor<string, []>("x_583_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4588_cast_fp16 = relu(x = x_583_cast_fp16)[name = tensor<string, []>("op_4588_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4589_cast_fp16 = abs(x = x_583_cast_fp16)[name = tensor<string, []>("op_4589_cast_fp16")];
            tensor<fp16, []> const_101_promoted_to_fp16 = const()[name = tensor<string, []>("const_101_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_4590_cast_fp16 = mul(x = var_4589_cast_fp16, y = const_101_promoted_to_fp16)[name = tensor<string, []>("op_4590_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4591_cast_fp16 = exp(x = var_4590_cast_fp16)[name = tensor<string, []>("op_4591_cast_fp16")];
            tensor<fp32, []> var_4592_epsilon_0 = const()[name = tensor<string, []>("op_4592_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_4592_cast_fp16 = log(epsilon = var_4592_epsilon_0, x = var_4591_cast_fp16)[name = tensor<string, []>("op_4592_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4593_cast_fp16 = add(x = var_4588_cast_fp16, y = var_4592_cast_fp16)[name = tensor<string, []>("op_4593_cast_fp16")];
            tensor<fp16, []> var_4594_to_fp16 = const()[name = tensor<string, []>("op_4594_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_4595_cast_fp16 = mul(x = x_581_cast_fp16, y = var_4594_to_fp16)[name = tensor<string, []>("op_4595_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4596_cast_fp16 = sub(x = var_4593_cast_fp16, y = var_4595_cast_fp16)[name = tensor<string, []>("op_4596_cast_fp16")];
            tensor<fp16, []> var_4597_to_fp16 = const()[name = tensor<string, []>("op_4597_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_415_cast_fp16 = sub(x = var_4596_cast_fp16, y = var_4597_to_fp16)[name = tensor<string, []>("input_415_cast_fp16")];
            tensor<string, []> var_4605_pad_type_0 = const()[name = tensor<string, []>("op_4605_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4605_strides_0 = const()[name = tensor<string, []>("op_4605_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4605_pad_0 = const()[name = tensor<string, []>("op_4605_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4605_dilations_0 = const()[name = tensor<string, []>("op_4605_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4605_groups_0 = const()[name = tensor<string, []>("op_4605_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_4_layers_0_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277844096)))];
            tensor<fp16, [512]> core_stacks_4_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278368448)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4605_cast_fp16 = conv(bias = core_stacks_4_layers_0_conv_module2_out_proj_bias_to_fp16, dilations = var_4605_dilations_0, groups = var_4605_groups_0, pad = var_4605_pad_0, pad_type = var_4605_pad_type_0, strides = var_4605_strides_0, weight = core_stacks_4_layers_0_conv_module2_out_proj_weight_to_fp16, x = input_415_cast_fp16)[name = tensor<string, []>("op_4605_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_417_cast_fp16 = add(x = x_575_cast_fp16, y = var_4605_cast_fp16)[name = tensor<string, []>("input_417_cast_fp16")];
            tensor<string, []> x_585_pad_type_0 = const()[name = tensor<string, []>("x_585_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_585_strides_0 = const()[name = tensor<string, []>("x_585_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_585_pad_0 = const()[name = tensor<string, []>("x_585_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_585_dilations_0 = const()[name = tensor<string, []>("x_585_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_585_groups_0 = const()[name = tensor<string, []>("x_585_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_4_layers_0_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278369536)))];
            tensor<fp16, [1920]> core_stacks_4_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280335680)))];
            tensor<fp16, [1, 1920, 1, 1024]> x_585_cast_fp16 = conv(bias = core_stacks_4_layers_0_feed_forward3_in_proj_bias_to_fp16, dilations = x_585_dilations_0, groups = x_585_groups_0, pad = x_585_pad_0, pad_type = x_585_pad_type_0, strides = x_585_strides_0, weight = core_stacks_4_layers_0_feed_forward3_in_proj_weight_to_fp16, x = input_417_cast_fp16)[name = tensor<string, []>("x_585_cast_fp16")];
            tensor<fp16, []> var_4616_to_fp16 = const()[name = tensor<string, []>("op_4616_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 1024]> x_587_cast_fp16 = sub(x = x_585_cast_fp16, y = var_4616_to_fp16)[name = tensor<string, []>("x_587_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_4618_cast_fp16 = relu(x = x_587_cast_fp16)[name = tensor<string, []>("op_4618_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_4619_cast_fp16 = abs(x = x_587_cast_fp16)[name = tensor<string, []>("op_4619_cast_fp16")];
            tensor<fp16, []> const_102_promoted_to_fp16 = const()[name = tensor<string, []>("const_102_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_4620_cast_fp16 = mul(x = var_4619_cast_fp16, y = const_102_promoted_to_fp16)[name = tensor<string, []>("op_4620_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_4621_cast_fp16 = exp(x = var_4620_cast_fp16)[name = tensor<string, []>("op_4621_cast_fp16")];
            tensor<fp32, []> var_4622_epsilon_0 = const()[name = tensor<string, []>("op_4622_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_4622_cast_fp16 = log(epsilon = var_4622_epsilon_0, x = var_4621_cast_fp16)[name = tensor<string, []>("op_4622_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_4623_cast_fp16 = add(x = var_4618_cast_fp16, y = var_4622_cast_fp16)[name = tensor<string, []>("op_4623_cast_fp16")];
            tensor<fp16, []> var_4624_to_fp16 = const()[name = tensor<string, []>("op_4624_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 1024]> var_4625_cast_fp16 = mul(x = x_585_cast_fp16, y = var_4624_to_fp16)[name = tensor<string, []>("op_4625_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_4626_cast_fp16 = sub(x = var_4623_cast_fp16, y = var_4625_cast_fp16)[name = tensor<string, []>("op_4626_cast_fp16")];
            tensor<fp16, []> var_4627_to_fp16 = const()[name = tensor<string, []>("op_4627_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 1024]> input_419_cast_fp16 = sub(x = var_4626_cast_fp16, y = var_4627_to_fp16)[name = tensor<string, []>("input_419_cast_fp16")];
            tensor<string, []> var_4635_pad_type_0 = const()[name = tensor<string, []>("op_4635_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4635_strides_0 = const()[name = tensor<string, []>("op_4635_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4635_pad_0 = const()[name = tensor<string, []>("op_4635_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4635_dilations_0 = const()[name = tensor<string, []>("op_4635_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4635_groups_0 = const()[name = tensor<string, []>("op_4635_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_4_layers_0_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280339584)))];
            tensor<fp16, [512]> core_stacks_4_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282305728)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4635_cast_fp16 = conv(bias = core_stacks_4_layers_0_feed_forward3_out_proj_bias_to_fp16, dilations = var_4635_dilations_0, groups = var_4635_groups_0, pad = var_4635_pad_0, pad_type = var_4635_pad_type_0, strides = var_4635_strides_0, weight = core_stacks_4_layers_0_feed_forward3_out_proj_weight_to_fp16, x = input_419_cast_fp16)[name = tensor<string, []>("op_4635_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_589_cast_fp16 = add(x = input_417_cast_fp16, y = var_4635_cast_fp16)[name = tensor<string, []>("x_589_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282306816)))];
            tensor<fp16, [1, 512, 1, 1024]> d_25_cast_fp16 = sub(x = x_589_cast_fp16, y = core_stacks_4_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("d_25_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4640_cast_fp16 = mul(x = d_25_cast_fp16, y = d_25_cast_fp16)[name = tensor<string, []>("op_4640_cast_fp16")];
            tensor<int32, [1]> var_4642_axes_0 = const()[name = tensor<string, []>("op_4642_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_4642_keep_dims_0 = const()[name = tensor<string, []>("op_4642_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 1024]> var_4642_cast_fp16 = reduce_mean(axes = var_4642_axes_0, keep_dims = var_4642_keep_dims_0, x = var_4640_cast_fp16)[name = tensor<string, []>("op_4642_cast_fp16")];
            tensor<fp32, []> scales_25_epsilon_0 = const()[name = tensor<string, []>("scales_25_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1024]> scales_25_cast_fp16 = rsqrt(epsilon = scales_25_epsilon_0, x = var_4642_cast_fp16)[name = tensor<string, []>("scales_25_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4644_cast_fp16 = mul(x = x_589_cast_fp16, y = scales_25_cast_fp16)[name = tensor<string, []>("op_4644_cast_fp16")];
            tensor<fp16, []> core_stacks_4_layers_0_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.b54p-1)];
            tensor<fp16, [1, 512, 1, 1024]> src_57_cast_fp16 = mul(x = var_4644_cast_fp16, y = core_stacks_4_layers_0_norm_scale_to_fp16)[name = tensor<string, []>("src_57_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4647_cast_fp16 = sub(x = src_57_cast_fp16, y = input_391_cast_fp16)[name = tensor<string, []>("op_4647_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282307904)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4648_cast_fp16 = mul(x = var_4647_cast_fp16, y = core_stacks_4_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_4648_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_421_cast_fp16 = add(x = input_391_cast_fp16, y = var_4648_cast_fp16)[name = tensor<string, []>("input_421_cast_fp16")];
            tensor<string, []> proj_27_pad_type_0 = const()[name = tensor<string, []>("proj_27_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_27_strides_0 = const()[name = tensor<string, []>("proj_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_27_pad_0 = const()[name = tensor<string, []>("proj_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_27_dilations_0 = const()[name = tensor<string, []>("proj_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_27_groups_0 = const()[name = tensor<string, []>("proj_27_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_4_layers_1_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282308992)))];
            tensor<fp16, [384]> core_stacks_4_layers_1_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282702272)))];
            tensor<fp16, [1, 384, 1, 1024]> proj_27_cast_fp16 = conv(bias = core_stacks_4_layers_1_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_27_dilations_0, groups = proj_27_groups_0, pad = proj_27_pad_0, pad_type = proj_27_pad_type_0, strides = proj_27_strides_0, weight = core_stacks_4_layers_1_self_attn_weights_in_proj_weight_to_fp16, x = input_421_cast_fp16)[name = tensor<string, []>("proj_27_cast_fp16")];
            tensor<int32, [4]> var_4671_begin_0 = const()[name = tensor<string, []>("op_4671_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4671_end_0 = const()[name = tensor<string, []>("op_4671_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1024])];
            tensor<bool, [4]> var_4671_end_mask_0 = const()[name = tensor<string, []>("op_4671_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_4671_cast_fp16 = slice_by_index(begin = var_4671_begin_0, end = var_4671_end_0, end_mask = var_4671_end_mask_0, x = proj_27_cast_fp16)[name = tensor<string, []>("op_4671_cast_fp16")];
            tensor<int32, [4]> var_4672 = const()[name = tensor<string, []>("op_4672"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_4673_cast_fp16 = reshape(shape = var_4672, x = var_4671_cast_fp16)[name = tensor<string, []>("op_4673_cast_fp16")];
            tensor<int32, [4]> var_4675_begin_0 = const()[name = tensor<string, []>("op_4675_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_4675_end_0 = const()[name = tensor<string, []>("op_4675_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1024])];
            tensor<bool, [4]> var_4675_end_mask_0 = const()[name = tensor<string, []>("op_4675_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_4675_cast_fp16 = slice_by_index(begin = var_4675_begin_0, end = var_4675_end_0, end_mask = var_4675_end_mask_0, x = proj_27_cast_fp16)[name = tensor<string, []>("op_4675_cast_fp16")];
            tensor<int32, [4]> var_4676 = const()[name = tensor<string, []>("op_4676"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_4677_cast_fp16 = reshape(shape = var_4676, x = var_4675_cast_fp16)[name = tensor<string, []>("op_4677_cast_fp16")];
            tensor<int32, [4]> var_4679_begin_0 = const()[name = tensor<string, []>("op_4679_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_4679_end_0 = const()[name = tensor<string, []>("op_4679_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_4679_end_mask_0 = const()[name = tensor<string, []>("op_4679_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> 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 = proj_27_cast_fp16)[name = tensor<string, []>("op_4679_cast_fp16")];
            tensor<int32, [4]> var_4680 = const()[name = tensor<string, []>("op_4680"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> pt_27_cast_fp16 = reshape(shape = var_4680, x = var_4679_cast_fp16)[name = tensor<string, []>("pt_27_cast_fp16")];
            tensor<int32, [4]> transpose_104_perm_0 = const()[name = tensor<string, []>("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_524 = const()[name = tensor<string, []>("concat_524"), val = tensor<int32, [3]>([4, 1024, 32])];
            tensor<fp16, [4, 1, 1024, 32]> transpose_104_cast_fp16 = transpose(perm = transpose_104_perm_0, x = var_4673_cast_fp16)[name = tensor<string, []>("transpose_164")];
            tensor<fp16, [4, 1024, 32]> reshape_156_cast_fp16 = reshape(shape = concat_524, x = transpose_104_cast_fp16)[name = tensor<string, []>("reshape_156_cast_fp16")];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = tensor<string, []>("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_525 = const()[name = tensor<string, []>("concat_525"), val = tensor<int32, [3]>([4, 32, 1024])];
            tensor<fp16, [4, 1, 32, 1024]> transpose_105_cast_fp16 = transpose(perm = transpose_105_perm_0, x = var_4677_cast_fp16)[name = tensor<string, []>("transpose_163")];
            tensor<fp16, [4, 32, 1024]> reshape_157_cast_fp16 = reshape(shape = concat_525, x = transpose_105_cast_fp16)[name = tensor<string, []>("reshape_157_cast_fp16")];
            tensor<bool, []> matmul_52_transpose_x_0 = const()[name = tensor<string, []>("matmul_52_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_52_transpose_y_0 = const()[name = tensor<string, []>("matmul_52_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 1024]> matmul_52_cast_fp16 = matmul(transpose_x = matmul_52_transpose_x_0, transpose_y = matmul_52_transpose_y_0, x = reshape_156_cast_fp16, y = reshape_157_cast_fp16)[name = tensor<string, []>("matmul_52_cast_fp16")];
            tensor<int32, [4]> concat_529 = const()[name = tensor<string, []>("concat_529"), val = tensor<int32, [4]>([4, 1, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> reshape_158_cast_fp16 = reshape(shape = concat_529, x = matmul_52_cast_fp16)[name = tensor<string, []>("reshape_158_cast_fp16")];
            tensor<int32, [4]> attn_scores_79_perm_0 = const()[name = tensor<string, []>("attn_scores_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_4684 = const()[name = tensor<string, []>("op_4684"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 1024, 1]> var_4685_cast_fp16 = transpose(perm = var_4684, x = pt_27_cast_fp16)[name = tensor<string, []>("transpose_161")];
            tensor<fp16, [4, 32, 1024, 1024]> var_4686_cast_fp16 = mul(x = var_4685_cast_fp16, y = core_pos_basis_1024_to_fp16)[name = tensor<string, []>("op_4686_cast_fp16")];
            tensor<int32, [1]> pos_scores_27_axes_0 = const()[name = tensor<string, []>("pos_scores_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_27_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_27_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 1024, 1024]> pos_scores_27_cast_fp16 = reduce_sum(axes = pos_scores_27_axes_0, keep_dims = pos_scores_27_keep_dims_0, x = var_4686_cast_fp16)[name = tensor<string, []>("pos_scores_27_cast_fp16")];
            tensor<int32, [4]> var_4689 = const()[name = tensor<string, []>("op_4689"), val = tensor<int32, [4]>([4, 1024, 1, 1024])];
            tensor<fp16, [4, 1024, 1, 1024]> var_4690_cast_fp16 = reshape(shape = var_4689, x = pos_scores_27_cast_fp16)[name = tensor<string, []>("op_4690_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_79_cast_fp16 = transpose(perm = attn_scores_79_perm_0, x = reshape_158_cast_fp16)[name = tensor<string, []>("transpose_162")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_81_cast_fp16 = add(x = attn_scores_79_cast_fp16, y = var_4690_cast_fp16)[name = tensor<string, []>("attn_scores_81_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_83_cast_fp16 = add(x = attn_scores_81_cast_fp16, y = attn_bias_1_cast_fp16)[name = tensor<string, []>("attn_scores_83_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_weights_27_cast_fp16 = softmax(axis = var_41, x = attn_scores_83_cast_fp16)[name = tensor<string, []>("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_423_cast_fp16 = add(x = input_421_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
            tensor<string, []> x_591_pad_type_0 = const()[name = tensor<string, []>("x_591_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_591_strides_0 = const()[name = tensor<string, []>("x_591_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_591_pad_0 = const()[name = tensor<string, []>("x_591_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_591_dilations_0 = const()[name = tensor<string, []>("x_591_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_591_groups_0 = const()[name = tensor<string, []>("x_591_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_4_layers_1_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282703104)))];
            tensor<fp16, [1152]> core_stacks_4_layers_1_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283882816)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_591_cast_fp16 = conv(bias = core_stacks_4_layers_1_feed_forward1_in_proj_bias_to_fp16, dilations = x_591_dilations_0, groups = x_591_groups_0, pad = x_591_pad_0, pad_type = x_591_pad_type_0, strides = x_591_strides_0, weight = core_stacks_4_layers_1_feed_forward1_in_proj_weight_to_fp16, x = input_423_cast_fp16)[name = tensor<string, []>("x_591_cast_fp16")];
            tensor<fp16, []> var_4704_to_fp16 = const()[name = tensor<string, []>("op_4704_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 1024]> x_593_cast_fp16 = sub(x = x_591_cast_fp16, y = var_4704_to_fp16)[name = tensor<string, []>("x_593_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_4706_cast_fp16 = relu(x = x_593_cast_fp16)[name = tensor<string, []>("op_4706_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_4707_cast_fp16 = abs(x = x_593_cast_fp16)[name = tensor<string, []>("op_4707_cast_fp16")];
            tensor<fp16, []> const_103_promoted_to_fp16 = const()[name = tensor<string, []>("const_103_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_4708_cast_fp16 = mul(x = var_4707_cast_fp16, y = const_103_promoted_to_fp16)[name = tensor<string, []>("op_4708_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_4709_cast_fp16 = exp(x = var_4708_cast_fp16)[name = tensor<string, []>("op_4709_cast_fp16")];
            tensor<fp32, []> var_4710_epsilon_0 = const()[name = tensor<string, []>("op_4710_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_4710_cast_fp16 = log(epsilon = var_4710_epsilon_0, x = var_4709_cast_fp16)[name = tensor<string, []>("op_4710_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_4711_cast_fp16 = add(x = var_4706_cast_fp16, y = var_4710_cast_fp16)[name = tensor<string, []>("op_4711_cast_fp16")];
            tensor<fp16, []> var_4712_to_fp16 = const()[name = tensor<string, []>("op_4712_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 1024]> var_4713_cast_fp16 = mul(x = x_591_cast_fp16, y = var_4712_to_fp16)[name = tensor<string, []>("op_4713_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_4714_cast_fp16 = sub(x = var_4711_cast_fp16, y = var_4713_cast_fp16)[name = tensor<string, []>("op_4714_cast_fp16")];
            tensor<fp16, []> var_4715_to_fp16 = const()[name = tensor<string, []>("op_4715_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 1024]> input_425_cast_fp16 = sub(x = var_4714_cast_fp16, y = var_4715_to_fp16)[name = tensor<string, []>("input_425_cast_fp16")];
            tensor<string, []> var_4723_pad_type_0 = const()[name = tensor<string, []>("op_4723_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4723_strides_0 = const()[name = tensor<string, []>("op_4723_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4723_pad_0 = const()[name = tensor<string, []>("op_4723_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4723_dilations_0 = const()[name = tensor<string, []>("op_4723_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4723_groups_0 = const()[name = tensor<string, []>("op_4723_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_4_layers_1_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283885184)))];
            tensor<fp16, [512]> core_stacks_4_layers_1_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285064896)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4723_cast_fp16 = conv(bias = core_stacks_4_layers_1_feed_forward1_out_proj_bias_to_fp16, dilations = var_4723_dilations_0, groups = var_4723_groups_0, pad = var_4723_pad_0, pad_type = var_4723_pad_type_0, strides = var_4723_strides_0, weight = core_stacks_4_layers_1_feed_forward1_out_proj_weight_to_fp16, x = input_425_cast_fp16)[name = tensor<string, []>("op_4723_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_427_cast_fp16 = add(x = input_423_cast_fp16, y = var_4723_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
            tensor<int32, [4]> var_4725_begin_0 = const()[name = tensor<string, []>("op_4725_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4725_end_0 = const()[name = tensor<string, []>("op_4725_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_4725_end_mask_0 = const()[name = tensor<string, []>("op_4725_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> var_4725_cast_fp16 = slice_by_index(begin = var_4725_begin_0, end = var_4725_end_0, end_mask = var_4725_end_mask_0, x = attn_weights_27_cast_fp16)[name = tensor<string, []>("op_4725_cast_fp16")];
            tensor<string, []> x_595_pad_type_0 = const()[name = tensor<string, []>("x_595_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_595_strides_0 = const()[name = tensor<string, []>("x_595_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_595_pad_0 = const()[name = tensor<string, []>("x_595_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_595_dilations_0 = const()[name = tensor<string, []>("x_595_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_595_groups_0 = const()[name = tensor<string, []>("x_595_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_4_layers_1_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285065984)))];
            tensor<fp16, [1152]> core_stacks_4_layers_1_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286245696)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_595_cast_fp16 = conv(bias = core_stacks_4_layers_1_nonlin_attention_in_proj_bias_to_fp16, dilations = x_595_dilations_0, groups = x_595_groups_0, pad = x_595_pad_0, pad_type = x_595_pad_type_0, strides = x_595_strides_0, weight = core_stacks_4_layers_1_nonlin_attention_in_proj_weight_to_fp16, x = input_427_cast_fp16)[name = tensor<string, []>("x_595_cast_fp16")];
            tensor<int32, [4]> var_4736_begin_0 = const()[name = tensor<string, []>("op_4736_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4736_end_0 = const()[name = tensor<string, []>("op_4736_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_4736_end_mask_0 = const()[name = tensor<string, []>("op_4736_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_4736_cast_fp16 = slice_by_index(begin = var_4736_begin_0, end = var_4736_end_0, end_mask = var_4736_end_mask_0, x = x_595_cast_fp16)[name = tensor<string, []>("op_4736_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> s_gate_27_cast_fp16 = tanh(x = var_4736_cast_fp16)[name = tensor<string, []>("s_gate_27_cast_fp16")];
            tensor<int32, [4]> var_4739_begin_0 = const()[name = tensor<string, []>("op_4739_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_4739_end_0 = const()[name = tensor<string, []>("op_4739_end_0"), val = tensor<int32, [4]>([1, 768, 1, 1024])];
            tensor<bool, [4]> var_4739_end_mask_0 = const()[name = tensor<string, []>("op_4739_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_4739_cast_fp16 = slice_by_index(begin = var_4739_begin_0, end = var_4739_end_0, end_mask = var_4739_end_mask_0, x = x_595_cast_fp16)[name = tensor<string, []>("op_4739_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> var_4740_cast_fp16 = mul(x = var_4739_cast_fp16, y = s_gate_27_cast_fp16)[name = tensor<string, []>("op_4740_cast_fp16")];
            tensor<int32, [4]> y_27_begin_0 = const()[name = tensor<string, []>("y_27_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_27_end_0 = const()[name = tensor<string, []>("y_27_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 1024])];
            tensor<bool, [4]> y_27_end_mask_0 = const()[name = tensor<string, []>("y_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> y_27_cast_fp16 = slice_by_index(begin = y_27_begin_0, end = y_27_end_0, end_mask = y_27_end_mask_0, x = x_595_cast_fp16)[name = tensor<string, []>("y_27_cast_fp16")];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = tensor<string, []>("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_534 = const()[name = tensor<string, []>("concat_534"), val = tensor<int32, [3]>([1, 1024, 1024])];
            tensor<fp16, [1, 1, 1024, 1024]> transpose_106_cast_fp16 = transpose(perm = transpose_106_perm_0, x = var_4725_cast_fp16)[name = tensor<string, []>("transpose_160")];
            tensor<fp16, [1, 1024, 1024]> reshape_159_cast_fp16 = reshape(shape = concat_534, x = transpose_106_cast_fp16)[name = tensor<string, []>("reshape_159_cast_fp16")];
            tensor<int32, [4]> transpose_107_perm_0 = const()[name = tensor<string, []>("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_535 = const()[name = tensor<string, []>("concat_535"), val = tensor<int32, [3]>([1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> transpose_107_cast_fp16 = transpose(perm = transpose_107_perm_0, x = var_4740_cast_fp16)[name = tensor<string, []>("transpose_159")];
            tensor<fp16, [1, 1024, 384]> reshape_160_cast_fp16 = reshape(shape = concat_535, x = transpose_107_cast_fp16)[name = tensor<string, []>("reshape_160_cast_fp16")];
            tensor<bool, []> matmul_53_transpose_x_0 = const()[name = tensor<string, []>("matmul_53_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_53_transpose_y_0 = const()[name = tensor<string, []>("matmul_53_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1024, 384]> matmul_53_cast_fp16 = matmul(transpose_x = matmul_53_transpose_x_0, transpose_y = matmul_53_transpose_y_0, x = reshape_159_cast_fp16, y = reshape_160_cast_fp16)[name = tensor<string, []>("matmul_53_cast_fp16")];
            tensor<int32, [4]> concat_539 = const()[name = tensor<string, []>("concat_539"), val = tensor<int32, [4]>([1, 1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> reshape_161_cast_fp16 = reshape(shape = concat_539, x = matmul_53_cast_fp16)[name = tensor<string, []>("reshape_161_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = tensor<string, []>("v_27_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 1024]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = reshape_161_cast_fp16)[name = tensor<string, []>("transpose_158")];
            tensor<fp16, [1, 384, 1, 1024]> input_429_cast_fp16 = mul(x = v_27_cast_fp16, y = y_27_cast_fp16)[name = tensor<string, []>("input_429_cast_fp16")];
            tensor<string, []> var_4752_pad_type_0 = const()[name = tensor<string, []>("op_4752_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4752_strides_0 = const()[name = tensor<string, []>("op_4752_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4752_pad_0 = const()[name = tensor<string, []>("op_4752_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4752_dilations_0 = const()[name = tensor<string, []>("op_4752_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4752_groups_0 = const()[name = tensor<string, []>("op_4752_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_4_layers_1_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286248064)))];
            tensor<fp16, [512]> core_stacks_4_layers_1_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286641344)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4752_cast_fp16 = conv(bias = core_stacks_4_layers_1_nonlin_attention_out_proj_bias_to_fp16, dilations = var_4752_dilations_0, groups = var_4752_groups_0, pad = var_4752_pad_0, pad_type = var_4752_pad_type_0, strides = var_4752_strides_0, weight = core_stacks_4_layers_1_nonlin_attention_out_proj_weight_to_fp16, x = input_429_cast_fp16)[name = tensor<string, []>("op_4752_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_597_cast_fp16 = add(x = input_427_cast_fp16, y = var_4752_cast_fp16)[name = tensor<string, []>("x_597_cast_fp16")];
            tensor<string, []> var_4763_pad_type_0 = const()[name = tensor<string, []>("op_4763_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4763_strides_0 = const()[name = tensor<string, []>("op_4763_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4763_pad_0 = const()[name = tensor<string, []>("op_4763_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4763_dilations_0 = const()[name = tensor<string, []>("op_4763_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4763_groups_0 = const()[name = tensor<string, []>("op_4763_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_4_layers_1_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286642432)))];
            tensor<fp16, [48]> core_stacks_4_layers_1_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286691648)))];
            tensor<fp16, [1, 48, 1, 1024]> var_4763_cast_fp16 = conv(bias = core_stacks_4_layers_1_self_attn1_in_proj_bias_to_fp16, dilations = var_4763_dilations_0, groups = var_4763_groups_0, pad = var_4763_pad_0, pad_type = var_4763_pad_type_0, strides = var_4763_strides_0, weight = core_stacks_4_layers_1_self_attn1_in_proj_weight_to_fp16, x = x_597_cast_fp16)[name = tensor<string, []>("op_4763_cast_fp16")];
            tensor<int32, [4]> var_4764 = const()[name = tensor<string, []>("op_4764"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_4765_cast_fp16 = reshape(shape = var_4764, x = var_4763_cast_fp16)[name = tensor<string, []>("op_4765_cast_fp16")];
            tensor<int32, [4]> transpose_108_perm_0 = const()[name = tensor<string, []>("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_544 = const()[name = tensor<string, []>("concat_544"), val = tensor<int32, [3]>([4, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> transpose_108_cast_fp16 = transpose(perm = transpose_108_perm_0, x = attn_weights_27_cast_fp16)[name = tensor<string, []>("transpose_157")];
            tensor<fp16, [4, 1024, 1024]> reshape_162_cast_fp16 = reshape(shape = concat_544, x = transpose_108_cast_fp16)[name = tensor<string, []>("reshape_162_cast_fp16")];
            tensor<int32, [4]> transpose_109_perm_0 = const()[name = tensor<string, []>("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_545 = const()[name = tensor<string, []>("concat_545"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_109_cast_fp16 = transpose(perm = transpose_109_perm_0, x = var_4765_cast_fp16)[name = tensor<string, []>("transpose_156")];
            tensor<fp16, [4, 1024, 12]> reshape_163_cast_fp16 = reshape(shape = concat_545, x = transpose_109_cast_fp16)[name = tensor<string, []>("reshape_163_cast_fp16")];
            tensor<bool, []> matmul_54_transpose_x_0 = const()[name = tensor<string, []>("matmul_54_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_54_transpose_y_0 = const()[name = tensor<string, []>("matmul_54_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_54_cast_fp16 = matmul(transpose_x = matmul_54_transpose_x_0, transpose_y = matmul_54_transpose_y_0, x = reshape_162_cast_fp16, y = reshape_163_cast_fp16)[name = tensor<string, []>("matmul_54_cast_fp16")];
            tensor<int32, [4]> concat_549 = const()[name = tensor<string, []>("concat_549"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_164_cast_fp16 = reshape(shape = concat_549, x = matmul_54_cast_fp16)[name = tensor<string, []>("reshape_164_cast_fp16")];
            tensor<int32, [4]> out_53_perm_0 = const()[name = tensor<string, []>("out_53_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_4768 = const()[name = tensor<string, []>("op_4768"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_53_cast_fp16 = transpose(perm = out_53_perm_0, x = reshape_164_cast_fp16)[name = tensor<string, []>("transpose_155")];
            tensor<fp16, [1, 48, 1, 1024]> input_431_cast_fp16 = reshape(shape = var_4768, x = out_53_cast_fp16)[name = tensor<string, []>("input_431_cast_fp16")];
            tensor<string, []> var_4776_pad_type_0 = const()[name = tensor<string, []>("op_4776_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4776_strides_0 = const()[name = tensor<string, []>("op_4776_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4776_pad_0 = const()[name = tensor<string, []>("op_4776_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4776_dilations_0 = const()[name = tensor<string, []>("op_4776_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4776_groups_0 = const()[name = tensor<string, []>("op_4776_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_4_layers_1_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286691840)))];
            tensor<fp16, [512]> core_stacks_4_layers_1_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286741056)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4776_cast_fp16 = conv(bias = core_stacks_4_layers_1_self_attn1_out_proj_bias_to_fp16, dilations = var_4776_dilations_0, groups = var_4776_groups_0, pad = var_4776_pad_0, pad_type = var_4776_pad_type_0, strides = var_4776_strides_0, weight = core_stacks_4_layers_1_self_attn1_out_proj_weight_to_fp16, x = input_431_cast_fp16)[name = tensor<string, []>("op_4776_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_599_cast_fp16 = add(x = x_597_cast_fp16, y = var_4776_cast_fp16)[name = tensor<string, []>("x_599_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_601_cast_fp16 = add(x = x_599_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("x_601_cast_fp16")];
            tensor<string, []> x_603_pad_type_0 = const()[name = tensor<string, []>("x_603_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_603_strides_0 = const()[name = tensor<string, []>("x_603_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_603_pad_0 = const()[name = tensor<string, []>("x_603_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_603_dilations_0 = const()[name = tensor<string, []>("x_603_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_603_groups_0 = const()[name = tensor<string, []>("x_603_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_4_layers_1_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286742144)))];
            tensor<fp16, [1024]> core_stacks_4_layers_1_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287790784)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_603_cast_fp16 = conv(bias = core_stacks_4_layers_1_conv_module1_in_proj_bias_to_fp16, dilations = x_603_dilations_0, groups = x_603_groups_0, pad = x_603_pad_0, pad_type = x_603_pad_type_0, strides = x_603_strides_0, weight = core_stacks_4_layers_1_conv_module1_in_proj_weight_to_fp16, x = x_601_cast_fp16)[name = tensor<string, []>("x_603_cast_fp16")];
            tensor<int32, [4]> var_4794_begin_0 = const()[name = tensor<string, []>("op_4794_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4794_end_0 = const()[name = tensor<string, []>("op_4794_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_4794_end_mask_0 = const()[name = tensor<string, []>("op_4794_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4794_cast_fp16 = slice_by_index(begin = var_4794_begin_0, end = var_4794_end_0, end_mask = var_4794_end_mask_0, x = x_603_cast_fp16)[name = tensor<string, []>("op_4794_cast_fp16")];
            tensor<int32, [4]> var_4796_begin_0 = const()[name = tensor<string, []>("op_4796_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_4796_end_0 = const()[name = tensor<string, []>("op_4796_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_4796_end_mask_0 = const()[name = tensor<string, []>("op_4796_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4796_cast_fp16 = slice_by_index(begin = var_4796_begin_0, end = var_4796_end_0, end_mask = var_4796_end_mask_0, x = x_603_cast_fp16)[name = tensor<string, []>("op_4796_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4797_cast_fp16 = sigmoid(x = var_4796_cast_fp16)[name = tensor<string, []>("op_4797_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_605_cast_fp16 = mul(x = var_4794_cast_fp16, y = var_4797_cast_fp16)[name = tensor<string, []>("x_605_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_433_cast_fp16 = mul(x = x_605_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_433_cast_fp16")];
            tensor<string, []> var_4806_pad_type_0 = const()[name = tensor<string, []>("op_4806_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4806_pad_0 = const()[name = tensor<string, []>("op_4806_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4806_groups_0 = const()[name = tensor<string, []>("op_4806_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4806_strides_0 = const()[name = tensor<string, []>("op_4806_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4806_dilations_0 = const()[name = tensor<string, []>("op_4806_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_1_conv_module1_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module1_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287792896)))];
            tensor<fp16, [512]> core_stacks_4_layers_1_conv_module1_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module1_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287808320)))];
            tensor<fp16, [1, 512, 1, 1040]> var_4806_cast_fp16 = conv(bias = core_stacks_4_layers_1_conv_module1_dw_a_bias_to_fp16, dilations = var_4806_dilations_0, groups = var_4806_groups_0, pad = var_4806_pad_0, pad_type = var_4806_pad_type_0, strides = var_4806_strides_0, weight = core_stacks_4_layers_1_conv_module1_dw_a_weight_to_fp16, x = input_433_cast_fp16)[name = tensor<string, []>("op_4806_cast_fp16")];
            tensor<int32, [4]> var_4807_begin_0 = const()[name = tensor<string, []>("op_4807_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4807_end_0 = const()[name = tensor<string, []>("op_4807_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_4807_end_mask_0 = const()[name = tensor<string, []>("op_4807_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_4807_cast_fp16 = slice_by_index(begin = var_4807_begin_0, end = var_4807_end_0, end_mask = var_4807_end_mask_0, x = var_4806_cast_fp16)[name = tensor<string, []>("op_4807_cast_fp16")];
            tensor<string, []> var_4813_pad_type_0 = const()[name = tensor<string, []>("op_4813_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4813_pad_0 = const()[name = tensor<string, []>("op_4813_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4813_groups_0 = const()[name = tensor<string, []>("op_4813_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4813_strides_0 = const()[name = tensor<string, []>("op_4813_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4813_dilations_0 = const()[name = tensor<string, []>("op_4813_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_1_conv_module1_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module1_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287809408)))];
            tensor<fp16, [1, 512, 1, 1040]> var_4813_cast_fp16 = conv(dilations = var_4813_dilations_0, groups = var_4813_groups_0, pad = var_4813_pad_0, pad_type = var_4813_pad_type_0, strides = var_4813_strides_0, weight = core_stacks_4_layers_1_conv_module1_dw_b_weight_to_fp16, x = input_433_cast_fp16)[name = tensor<string, []>("op_4813_cast_fp16")];
            tensor<int32, [4]> var_4817_begin_0 = const()[name = tensor<string, []>("op_4817_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_4817_end_0 = const()[name = tensor<string, []>("op_4817_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_4817_end_mask_0 = const()[name = tensor<string, []>("op_4817_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_4817_cast_fp16 = slice_by_index(begin = var_4817_begin_0, end = var_4817_end_0, end_mask = var_4817_end_mask_0, x = var_4813_cast_fp16)[name = tensor<string, []>("op_4817_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4818_cast_fp16 = add(x = var_4807_cast_fp16, y = var_4817_cast_fp16)[name = tensor<string, []>("op_4818_cast_fp16")];
            tensor<string, []> var_4824_pad_type_0 = const()[name = tensor<string, []>("op_4824_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4824_pad_0 = const()[name = tensor<string, []>("op_4824_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4824_groups_0 = const()[name = tensor<string, []>("op_4824_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4824_strides_0 = const()[name = tensor<string, []>("op_4824_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4824_dilations_0 = const()[name = tensor<string, []>("op_4824_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_4_layers_1_conv_module1_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module1_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287824832)))];
            tensor<fp16, [1, 512, 1, 1054]> var_4824_cast_fp16 = conv(dilations = var_4824_dilations_0, groups = var_4824_groups_0, pad = var_4824_pad_0, pad_type = var_4824_pad_type_0, strides = var_4824_strides_0, weight = core_stacks_4_layers_1_conv_module1_dw_c_weight_to_fp16, x = input_433_cast_fp16)[name = tensor<string, []>("op_4824_cast_fp16")];
            tensor<int32, [4]> var_4828_begin_0 = const()[name = tensor<string, []>("op_4828_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_4828_end_0 = const()[name = tensor<string, []>("op_4828_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_4828_end_mask_0 = const()[name = tensor<string, []>("op_4828_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4828_cast_fp16 = slice_by_index(begin = var_4828_begin_0, end = var_4828_end_0, end_mask = var_4828_end_mask_0, x = var_4824_cast_fp16)[name = tensor<string, []>("op_4828_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_607_cast_fp16 = add(x = var_4818_cast_fp16, y = var_4828_cast_fp16)[name = tensor<string, []>("x_607_cast_fp16")];
            tensor<fp16, []> var_4830_to_fp16 = const()[name = tensor<string, []>("op_4830_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_609_cast_fp16 = sub(x = x_607_cast_fp16, y = var_4830_to_fp16)[name = tensor<string, []>("x_609_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4832_cast_fp16 = relu(x = x_609_cast_fp16)[name = tensor<string, []>("op_4832_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4833_cast_fp16 = abs(x = x_609_cast_fp16)[name = tensor<string, []>("op_4833_cast_fp16")];
            tensor<fp16, []> const_106_promoted_to_fp16 = const()[name = tensor<string, []>("const_106_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_4834_cast_fp16 = mul(x = var_4833_cast_fp16, y = const_106_promoted_to_fp16)[name = tensor<string, []>("op_4834_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4835_cast_fp16 = exp(x = var_4834_cast_fp16)[name = tensor<string, []>("op_4835_cast_fp16")];
            tensor<fp32, []> var_4836_epsilon_0 = const()[name = tensor<string, []>("op_4836_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_4836_cast_fp16 = log(epsilon = var_4836_epsilon_0, x = var_4835_cast_fp16)[name = tensor<string, []>("op_4836_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4837_cast_fp16 = add(x = var_4832_cast_fp16, y = var_4836_cast_fp16)[name = tensor<string, []>("op_4837_cast_fp16")];
            tensor<fp16, []> var_4838_to_fp16 = const()[name = tensor<string, []>("op_4838_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_4839_cast_fp16 = mul(x = x_607_cast_fp16, y = var_4838_to_fp16)[name = tensor<string, []>("op_4839_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4840_cast_fp16 = sub(x = var_4837_cast_fp16, y = var_4839_cast_fp16)[name = tensor<string, []>("op_4840_cast_fp16")];
            tensor<fp16, []> var_4841_to_fp16 = const()[name = tensor<string, []>("op_4841_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_435_cast_fp16 = sub(x = var_4840_cast_fp16, y = var_4841_to_fp16)[name = tensor<string, []>("input_435_cast_fp16")];
            tensor<string, []> var_4849_pad_type_0 = const()[name = tensor<string, []>("op_4849_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4849_strides_0 = const()[name = tensor<string, []>("op_4849_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4849_pad_0 = const()[name = tensor<string, []>("op_4849_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4849_dilations_0 = const()[name = tensor<string, []>("op_4849_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4849_groups_0 = const()[name = tensor<string, []>("op_4849_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_4_layers_1_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287825920)))];
            tensor<fp16, [512]> core_stacks_4_layers_1_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288350272)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4849_cast_fp16 = conv(bias = core_stacks_4_layers_1_conv_module1_out_proj_bias_to_fp16, dilations = var_4849_dilations_0, groups = var_4849_groups_0, pad = var_4849_pad_0, pad_type = var_4849_pad_type_0, strides = var_4849_strides_0, weight = core_stacks_4_layers_1_conv_module1_out_proj_weight_to_fp16, x = input_435_cast_fp16)[name = tensor<string, []>("op_4849_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_437_cast_fp16 = add(x = x_601_cast_fp16, y = var_4849_cast_fp16)[name = tensor<string, []>("input_437_cast_fp16")];
            tensor<string, []> x_611_pad_type_0 = const()[name = tensor<string, []>("x_611_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_611_strides_0 = const()[name = tensor<string, []>("x_611_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_611_pad_0 = const()[name = tensor<string, []>("x_611_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_611_dilations_0 = const()[name = tensor<string, []>("x_611_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_611_groups_0 = const()[name = tensor<string, []>("x_611_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_4_layers_1_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288351360)))];
            tensor<fp16, [1536]> core_stacks_4_layers_1_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289924288)))];
            tensor<fp16, [1, 1536, 1, 1024]> x_611_cast_fp16 = conv(bias = core_stacks_4_layers_1_feed_forward2_in_proj_bias_to_fp16, dilations = x_611_dilations_0, groups = x_611_groups_0, pad = x_611_pad_0, pad_type = x_611_pad_type_0, strides = x_611_strides_0, weight = core_stacks_4_layers_1_feed_forward2_in_proj_weight_to_fp16, x = input_437_cast_fp16)[name = tensor<string, []>("x_611_cast_fp16")];
            tensor<fp16, []> var_4860_to_fp16 = const()[name = tensor<string, []>("op_4860_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 1024]> x_613_cast_fp16 = sub(x = x_611_cast_fp16, y = var_4860_to_fp16)[name = tensor<string, []>("x_613_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_4862_cast_fp16 = relu(x = x_613_cast_fp16)[name = tensor<string, []>("op_4862_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_4863_cast_fp16 = abs(x = x_613_cast_fp16)[name = tensor<string, []>("op_4863_cast_fp16")];
            tensor<fp16, []> const_107_promoted_to_fp16 = const()[name = tensor<string, []>("const_107_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_4864_cast_fp16 = mul(x = var_4863_cast_fp16, y = const_107_promoted_to_fp16)[name = tensor<string, []>("op_4864_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_4865_cast_fp16 = exp(x = var_4864_cast_fp16)[name = tensor<string, []>("op_4865_cast_fp16")];
            tensor<fp32, []> var_4866_epsilon_0 = const()[name = tensor<string, []>("op_4866_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_4866_cast_fp16 = log(epsilon = var_4866_epsilon_0, x = var_4865_cast_fp16)[name = tensor<string, []>("op_4866_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_4867_cast_fp16 = add(x = var_4862_cast_fp16, y = var_4866_cast_fp16)[name = tensor<string, []>("op_4867_cast_fp16")];
            tensor<fp16, []> var_4868_to_fp16 = const()[name = tensor<string, []>("op_4868_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 1024]> var_4869_cast_fp16 = mul(x = x_611_cast_fp16, y = var_4868_to_fp16)[name = tensor<string, []>("op_4869_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_4870_cast_fp16 = sub(x = var_4867_cast_fp16, y = var_4869_cast_fp16)[name = tensor<string, []>("op_4870_cast_fp16")];
            tensor<fp16, []> var_4871_to_fp16 = const()[name = tensor<string, []>("op_4871_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 1024]> input_439_cast_fp16 = sub(x = var_4870_cast_fp16, y = var_4871_to_fp16)[name = tensor<string, []>("input_439_cast_fp16")];
            tensor<string, []> var_4879_pad_type_0 = const()[name = tensor<string, []>("op_4879_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4879_strides_0 = const()[name = tensor<string, []>("op_4879_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4879_pad_0 = const()[name = tensor<string, []>("op_4879_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4879_dilations_0 = const()[name = tensor<string, []>("op_4879_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4879_groups_0 = const()[name = tensor<string, []>("op_4879_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_4_layers_1_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289927424)))];
            tensor<fp16, [512]> core_stacks_4_layers_1_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291500352)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4879_cast_fp16 = conv(bias = core_stacks_4_layers_1_feed_forward2_out_proj_bias_to_fp16, dilations = var_4879_dilations_0, groups = var_4879_groups_0, pad = var_4879_pad_0, pad_type = var_4879_pad_type_0, strides = var_4879_strides_0, weight = core_stacks_4_layers_1_feed_forward2_out_proj_weight_to_fp16, x = input_439_cast_fp16)[name = tensor<string, []>("op_4879_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> src_59_cast_fp16 = add(x = input_437_cast_fp16, y = var_4879_cast_fp16)[name = tensor<string, []>("src_59_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4882_cast_fp16 = sub(x = src_59_cast_fp16, y = input_421_cast_fp16)[name = tensor<string, []>("op_4882_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_1_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291501440)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4883_cast_fp16 = mul(x = var_4882_cast_fp16, y = core_stacks_4_layers_1_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_4883_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_615_cast_fp16 = add(x = input_421_cast_fp16, y = var_4883_cast_fp16)[name = tensor<string, []>("x_615_cast_fp16")];
            tensor<string, []> var_4894_pad_type_0 = const()[name = tensor<string, []>("op_4894_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4894_strides_0 = const()[name = tensor<string, []>("op_4894_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4894_pad_0 = const()[name = tensor<string, []>("op_4894_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4894_dilations_0 = const()[name = tensor<string, []>("op_4894_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4894_groups_0 = const()[name = tensor<string, []>("op_4894_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_4_layers_1_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291502528)))];
            tensor<fp16, [48]> core_stacks_4_layers_1_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291551744)))];
            tensor<fp16, [1, 48, 1, 1024]> var_4894_cast_fp16 = conv(bias = core_stacks_4_layers_1_self_attn2_in_proj_bias_to_fp16, dilations = var_4894_dilations_0, groups = var_4894_groups_0, pad = var_4894_pad_0, pad_type = var_4894_pad_type_0, strides = var_4894_strides_0, weight = core_stacks_4_layers_1_self_attn2_in_proj_weight_to_fp16, x = x_615_cast_fp16)[name = tensor<string, []>("op_4894_cast_fp16")];
            tensor<int32, [4]> var_4895 = const()[name = tensor<string, []>("op_4895"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_4896_cast_fp16 = reshape(shape = var_4895, x = var_4894_cast_fp16)[name = tensor<string, []>("op_4896_cast_fp16")];
            tensor<int32, [4]> transpose_111_perm_0 = const()[name = tensor<string, []>("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_555 = const()[name = tensor<string, []>("concat_555"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_111_cast_fp16 = transpose(perm = transpose_111_perm_0, x = var_4896_cast_fp16)[name = tensor<string, []>("transpose_154")];
            tensor<fp16, [4, 1024, 12]> reshape_166_cast_fp16 = reshape(shape = concat_555, x = transpose_111_cast_fp16)[name = tensor<string, []>("reshape_166_cast_fp16")];
            tensor<bool, []> matmul_55_transpose_x_0 = const()[name = tensor<string, []>("matmul_55_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_55_transpose_y_0 = const()[name = tensor<string, []>("matmul_55_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_55_cast_fp16 = matmul(transpose_x = matmul_55_transpose_x_0, transpose_y = matmul_55_transpose_y_0, x = reshape_162_cast_fp16, y = reshape_166_cast_fp16)[name = tensor<string, []>("matmul_55_cast_fp16")];
            tensor<int32, [4]> concat_559 = const()[name = tensor<string, []>("concat_559"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_167_cast_fp16 = reshape(shape = concat_559, x = matmul_55_cast_fp16)[name = tensor<string, []>("reshape_167_cast_fp16")];
            tensor<int32, [4]> out_55_perm_0 = const()[name = tensor<string, []>("out_55_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_4899 = const()[name = tensor<string, []>("op_4899"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_55_cast_fp16 = transpose(perm = out_55_perm_0, x = reshape_167_cast_fp16)[name = tensor<string, []>("transpose_153")];
            tensor<fp16, [1, 48, 1, 1024]> input_441_cast_fp16 = reshape(shape = var_4899, x = out_55_cast_fp16)[name = tensor<string, []>("input_441_cast_fp16")];
            tensor<string, []> var_4907_pad_type_0 = const()[name = tensor<string, []>("op_4907_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4907_strides_0 = const()[name = tensor<string, []>("op_4907_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4907_pad_0 = const()[name = tensor<string, []>("op_4907_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4907_dilations_0 = const()[name = tensor<string, []>("op_4907_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4907_groups_0 = const()[name = tensor<string, []>("op_4907_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_4_layers_1_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291551936)))];
            tensor<fp16, [512]> core_stacks_4_layers_1_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291601152)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4907_cast_fp16 = conv(bias = core_stacks_4_layers_1_self_attn2_out_proj_bias_to_fp16, dilations = var_4907_dilations_0, groups = var_4907_groups_0, pad = var_4907_pad_0, pad_type = var_4907_pad_type_0, strides = var_4907_strides_0, weight = core_stacks_4_layers_1_self_attn2_out_proj_weight_to_fp16, x = input_441_cast_fp16)[name = tensor<string, []>("op_4907_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_617_cast_fp16 = add(x = x_615_cast_fp16, y = var_4907_cast_fp16)[name = tensor<string, []>("x_617_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_619_cast_fp16 = add(x = x_617_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("x_619_cast_fp16")];
            tensor<string, []> x_621_pad_type_0 = const()[name = tensor<string, []>("x_621_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_621_strides_0 = const()[name = tensor<string, []>("x_621_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_621_pad_0 = const()[name = tensor<string, []>("x_621_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_621_dilations_0 = const()[name = tensor<string, []>("x_621_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_621_groups_0 = const()[name = tensor<string, []>("x_621_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_4_layers_1_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291602240)))];
            tensor<fp16, [1024]> core_stacks_4_layers_1_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292650880)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_621_cast_fp16 = conv(bias = core_stacks_4_layers_1_conv_module2_in_proj_bias_to_fp16, dilations = x_621_dilations_0, groups = x_621_groups_0, pad = x_621_pad_0, pad_type = x_621_pad_type_0, strides = x_621_strides_0, weight = core_stacks_4_layers_1_conv_module2_in_proj_weight_to_fp16, x = x_619_cast_fp16)[name = tensor<string, []>("x_621_cast_fp16")];
            tensor<int32, [4]> var_4925_begin_0 = const()[name = tensor<string, []>("op_4925_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4925_end_0 = const()[name = tensor<string, []>("op_4925_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_4925_end_mask_0 = const()[name = tensor<string, []>("op_4925_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4925_cast_fp16 = slice_by_index(begin = var_4925_begin_0, end = var_4925_end_0, end_mask = var_4925_end_mask_0, x = x_621_cast_fp16)[name = tensor<string, []>("op_4925_cast_fp16")];
            tensor<int32, [4]> var_4927_begin_0 = const()[name = tensor<string, []>("op_4927_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_4927_end_0 = const()[name = tensor<string, []>("op_4927_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_4927_end_mask_0 = const()[name = tensor<string, []>("op_4927_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> 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 = x_621_cast_fp16)[name = tensor<string, []>("op_4927_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4928_cast_fp16 = sigmoid(x = var_4927_cast_fp16)[name = tensor<string, []>("op_4928_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_623_cast_fp16 = mul(x = var_4925_cast_fp16, y = var_4928_cast_fp16)[name = tensor<string, []>("x_623_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_443_cast_fp16 = mul(x = x_623_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_443_cast_fp16")];
            tensor<string, []> var_4937_pad_type_0 = const()[name = tensor<string, []>("op_4937_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4937_pad_0 = const()[name = tensor<string, []>("op_4937_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4937_groups_0 = const()[name = tensor<string, []>("op_4937_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4937_strides_0 = const()[name = tensor<string, []>("op_4937_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4937_dilations_0 = const()[name = tensor<string, []>("op_4937_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_1_conv_module2_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module2_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292652992)))];
            tensor<fp16, [512]> core_stacks_4_layers_1_conv_module2_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module2_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292668416)))];
            tensor<fp16, [1, 512, 1, 1040]> var_4937_cast_fp16 = conv(bias = core_stacks_4_layers_1_conv_module2_dw_a_bias_to_fp16, dilations = var_4937_dilations_0, groups = var_4937_groups_0, pad = var_4937_pad_0, pad_type = var_4937_pad_type_0, strides = var_4937_strides_0, weight = core_stacks_4_layers_1_conv_module2_dw_a_weight_to_fp16, x = input_443_cast_fp16)[name = tensor<string, []>("op_4937_cast_fp16")];
            tensor<int32, [4]> var_4938_begin_0 = const()[name = tensor<string, []>("op_4938_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4938_end_0 = const()[name = tensor<string, []>("op_4938_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_4938_end_mask_0 = const()[name = tensor<string, []>("op_4938_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_4938_cast_fp16 = slice_by_index(begin = var_4938_begin_0, end = var_4938_end_0, end_mask = var_4938_end_mask_0, x = var_4937_cast_fp16)[name = tensor<string, []>("op_4938_cast_fp16")];
            tensor<string, []> var_4944_pad_type_0 = const()[name = tensor<string, []>("op_4944_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4944_pad_0 = const()[name = tensor<string, []>("op_4944_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4944_groups_0 = const()[name = tensor<string, []>("op_4944_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4944_strides_0 = const()[name = tensor<string, []>("op_4944_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4944_dilations_0 = const()[name = tensor<string, []>("op_4944_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_1_conv_module2_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module2_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292669504)))];
            tensor<fp16, [1, 512, 1, 1040]> var_4944_cast_fp16 = conv(dilations = var_4944_dilations_0, groups = var_4944_groups_0, pad = var_4944_pad_0, pad_type = var_4944_pad_type_0, strides = var_4944_strides_0, weight = core_stacks_4_layers_1_conv_module2_dw_b_weight_to_fp16, x = input_443_cast_fp16)[name = tensor<string, []>("op_4944_cast_fp16")];
            tensor<int32, [4]> var_4948_begin_0 = const()[name = tensor<string, []>("op_4948_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_4948_end_0 = const()[name = tensor<string, []>("op_4948_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_4948_end_mask_0 = const()[name = tensor<string, []>("op_4948_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_4948_cast_fp16 = slice_by_index(begin = var_4948_begin_0, end = var_4948_end_0, end_mask = var_4948_end_mask_0, x = var_4944_cast_fp16)[name = tensor<string, []>("op_4948_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4949_cast_fp16 = add(x = var_4938_cast_fp16, y = var_4948_cast_fp16)[name = tensor<string, []>("op_4949_cast_fp16")];
            tensor<string, []> var_4955_pad_type_0 = const()[name = tensor<string, []>("op_4955_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_4955_pad_0 = const()[name = tensor<string, []>("op_4955_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_4955_groups_0 = const()[name = tensor<string, []>("op_4955_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_4955_strides_0 = const()[name = tensor<string, []>("op_4955_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_4955_dilations_0 = const()[name = tensor<string, []>("op_4955_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_4_layers_1_conv_module2_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module2_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292684928)))];
            tensor<fp16, [1, 512, 1, 1054]> var_4955_cast_fp16 = conv(dilations = var_4955_dilations_0, groups = var_4955_groups_0, pad = var_4955_pad_0, pad_type = var_4955_pad_type_0, strides = var_4955_strides_0, weight = core_stacks_4_layers_1_conv_module2_dw_c_weight_to_fp16, x = input_443_cast_fp16)[name = tensor<string, []>("op_4955_cast_fp16")];
            tensor<int32, [4]> var_4959_begin_0 = const()[name = tensor<string, []>("op_4959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_4959_end_0 = const()[name = tensor<string, []>("op_4959_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_4959_end_mask_0 = const()[name = tensor<string, []>("op_4959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_4959_cast_fp16 = slice_by_index(begin = var_4959_begin_0, end = var_4959_end_0, end_mask = var_4959_end_mask_0, x = var_4955_cast_fp16)[name = tensor<string, []>("op_4959_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_625_cast_fp16 = add(x = var_4949_cast_fp16, y = var_4959_cast_fp16)[name = tensor<string, []>("x_625_cast_fp16")];
            tensor<fp16, []> var_4961_to_fp16 = const()[name = tensor<string, []>("op_4961_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_627_cast_fp16 = sub(x = x_625_cast_fp16, y = var_4961_to_fp16)[name = tensor<string, []>("x_627_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4963_cast_fp16 = relu(x = x_627_cast_fp16)[name = tensor<string, []>("op_4963_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4964_cast_fp16 = abs(x = x_627_cast_fp16)[name = tensor<string, []>("op_4964_cast_fp16")];
            tensor<fp16, []> const_110_promoted_to_fp16 = const()[name = tensor<string, []>("const_110_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_4965_cast_fp16 = mul(x = var_4964_cast_fp16, y = const_110_promoted_to_fp16)[name = tensor<string, []>("op_4965_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4966_cast_fp16 = exp(x = var_4965_cast_fp16)[name = tensor<string, []>("op_4966_cast_fp16")];
            tensor<fp32, []> var_4967_epsilon_0 = const()[name = tensor<string, []>("op_4967_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_4967_cast_fp16 = log(epsilon = var_4967_epsilon_0, x = var_4966_cast_fp16)[name = tensor<string, []>("op_4967_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4968_cast_fp16 = add(x = var_4963_cast_fp16, y = var_4967_cast_fp16)[name = tensor<string, []>("op_4968_cast_fp16")];
            tensor<fp16, []> var_4969_to_fp16 = const()[name = tensor<string, []>("op_4969_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_4970_cast_fp16 = mul(x = x_625_cast_fp16, y = var_4969_to_fp16)[name = tensor<string, []>("op_4970_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_4971_cast_fp16 = sub(x = var_4968_cast_fp16, y = var_4970_cast_fp16)[name = tensor<string, []>("op_4971_cast_fp16")];
            tensor<fp16, []> var_4972_to_fp16 = const()[name = tensor<string, []>("op_4972_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_445_cast_fp16 = sub(x = var_4971_cast_fp16, y = var_4972_to_fp16)[name = tensor<string, []>("input_445_cast_fp16")];
            tensor<string, []> var_4980_pad_type_0 = const()[name = tensor<string, []>("op_4980_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_4980_strides_0 = const()[name = tensor<string, []>("op_4980_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4980_pad_0 = const()[name = tensor<string, []>("op_4980_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4980_dilations_0 = const()[name = tensor<string, []>("op_4980_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_4980_groups_0 = const()[name = tensor<string, []>("op_4980_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_4_layers_1_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292686016)))];
            tensor<fp16, [512]> core_stacks_4_layers_1_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293210368)))];
            tensor<fp16, [1, 512, 1, 1024]> var_4980_cast_fp16 = conv(bias = core_stacks_4_layers_1_conv_module2_out_proj_bias_to_fp16, dilations = var_4980_dilations_0, groups = var_4980_groups_0, pad = var_4980_pad_0, pad_type = var_4980_pad_type_0, strides = var_4980_strides_0, weight = core_stacks_4_layers_1_conv_module2_out_proj_weight_to_fp16, x = input_445_cast_fp16)[name = tensor<string, []>("op_4980_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_447_cast_fp16 = add(x = x_619_cast_fp16, y = var_4980_cast_fp16)[name = tensor<string, []>("input_447_cast_fp16")];
            tensor<string, []> x_629_pad_type_0 = const()[name = tensor<string, []>("x_629_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_629_strides_0 = const()[name = tensor<string, []>("x_629_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_629_pad_0 = const()[name = tensor<string, []>("x_629_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_629_dilations_0 = const()[name = tensor<string, []>("x_629_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_629_groups_0 = const()[name = tensor<string, []>("x_629_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_4_layers_1_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293211456)))];
            tensor<fp16, [1920]> core_stacks_4_layers_1_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295177600)))];
            tensor<fp16, [1, 1920, 1, 1024]> x_629_cast_fp16 = conv(bias = core_stacks_4_layers_1_feed_forward3_in_proj_bias_to_fp16, dilations = x_629_dilations_0, groups = x_629_groups_0, pad = x_629_pad_0, pad_type = x_629_pad_type_0, strides = x_629_strides_0, weight = core_stacks_4_layers_1_feed_forward3_in_proj_weight_to_fp16, x = input_447_cast_fp16)[name = tensor<string, []>("x_629_cast_fp16")];
            tensor<fp16, []> var_4991_to_fp16 = const()[name = tensor<string, []>("op_4991_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 1024]> x_631_cast_fp16 = sub(x = x_629_cast_fp16, y = var_4991_to_fp16)[name = tensor<string, []>("x_631_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_4993_cast_fp16 = relu(x = x_631_cast_fp16)[name = tensor<string, []>("op_4993_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_4994_cast_fp16 = abs(x = x_631_cast_fp16)[name = tensor<string, []>("op_4994_cast_fp16")];
            tensor<fp16, []> const_111_promoted_to_fp16 = const()[name = tensor<string, []>("const_111_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_4995_cast_fp16 = mul(x = var_4994_cast_fp16, y = const_111_promoted_to_fp16)[name = tensor<string, []>("op_4995_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_4996_cast_fp16 = exp(x = var_4995_cast_fp16)[name = tensor<string, []>("op_4996_cast_fp16")];
            tensor<fp32, []> var_4997_epsilon_0 = const()[name = tensor<string, []>("op_4997_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_4997_cast_fp16 = log(epsilon = var_4997_epsilon_0, x = var_4996_cast_fp16)[name = tensor<string, []>("op_4997_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_4998_cast_fp16 = add(x = var_4993_cast_fp16, y = var_4997_cast_fp16)[name = tensor<string, []>("op_4998_cast_fp16")];
            tensor<fp16, []> var_4999_to_fp16 = const()[name = tensor<string, []>("op_4999_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 1024]> var_5000_cast_fp16 = mul(x = x_629_cast_fp16, y = var_4999_to_fp16)[name = tensor<string, []>("op_5000_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5001_cast_fp16 = sub(x = var_4998_cast_fp16, y = var_5000_cast_fp16)[name = tensor<string, []>("op_5001_cast_fp16")];
            tensor<fp16, []> var_5002_to_fp16 = const()[name = tensor<string, []>("op_5002_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 1024]> input_449_cast_fp16 = sub(x = var_5001_cast_fp16, y = var_5002_to_fp16)[name = tensor<string, []>("input_449_cast_fp16")];
            tensor<string, []> var_5010_pad_type_0 = const()[name = tensor<string, []>("op_5010_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5010_strides_0 = const()[name = tensor<string, []>("op_5010_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5010_pad_0 = const()[name = tensor<string, []>("op_5010_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5010_dilations_0 = const()[name = tensor<string, []>("op_5010_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5010_groups_0 = const()[name = tensor<string, []>("op_5010_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_4_layers_1_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295181504)))];
            tensor<fp16, [512]> core_stacks_4_layers_1_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297147648)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5010_cast_fp16 = conv(bias = core_stacks_4_layers_1_feed_forward3_out_proj_bias_to_fp16, dilations = var_5010_dilations_0, groups = var_5010_groups_0, pad = var_5010_pad_0, pad_type = var_5010_pad_type_0, strides = var_5010_strides_0, weight = core_stacks_4_layers_1_feed_forward3_out_proj_weight_to_fp16, x = input_449_cast_fp16)[name = tensor<string, []>("op_5010_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_633_cast_fp16 = add(x = input_447_cast_fp16, y = var_5010_cast_fp16)[name = tensor<string, []>("x_633_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_1_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297148736)))];
            tensor<fp16, [1, 512, 1, 1024]> d_27_cast_fp16 = sub(x = x_633_cast_fp16, y = core_stacks_4_layers_1_norm_bias_to_fp16)[name = tensor<string, []>("d_27_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5015_cast_fp16 = mul(x = d_27_cast_fp16, y = d_27_cast_fp16)[name = tensor<string, []>("op_5015_cast_fp16")];
            tensor<int32, [1]> var_5017_axes_0 = const()[name = tensor<string, []>("op_5017_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_5017_keep_dims_0 = const()[name = tensor<string, []>("op_5017_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 1024]> var_5017_cast_fp16 = reduce_mean(axes = var_5017_axes_0, keep_dims = var_5017_keep_dims_0, x = var_5015_cast_fp16)[name = tensor<string, []>("op_5017_cast_fp16")];
            tensor<fp32, []> scales_27_epsilon_0 = const()[name = tensor<string, []>("scales_27_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1024]> scales_27_cast_fp16 = rsqrt(epsilon = scales_27_epsilon_0, x = var_5017_cast_fp16)[name = tensor<string, []>("scales_27_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5019_cast_fp16 = mul(x = x_633_cast_fp16, y = scales_27_cast_fp16)[name = tensor<string, []>("op_5019_cast_fp16")];
            tensor<fp16, []> core_stacks_4_layers_1_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.42cp+0)];
            tensor<fp16, [1, 512, 1, 1024]> src_61_cast_fp16 = mul(x = var_5019_cast_fp16, y = core_stacks_4_layers_1_norm_scale_to_fp16)[name = tensor<string, []>("src_61_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5022_cast_fp16 = sub(x = src_61_cast_fp16, y = input_421_cast_fp16)[name = tensor<string, []>("op_5022_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_1_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_1_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297149824)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5023_cast_fp16 = mul(x = var_5022_cast_fp16, y = core_stacks_4_layers_1_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_5023_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_451_cast_fp16 = add(x = input_421_cast_fp16, y = var_5023_cast_fp16)[name = tensor<string, []>("input_451_cast_fp16")];
            tensor<string, []> proj_29_pad_type_0 = const()[name = tensor<string, []>("proj_29_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_29_strides_0 = const()[name = tensor<string, []>("proj_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_29_pad_0 = const()[name = tensor<string, []>("proj_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_29_dilations_0 = const()[name = tensor<string, []>("proj_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_29_groups_0 = const()[name = tensor<string, []>("proj_29_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_4_layers_2_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297150912)))];
            tensor<fp16, [384]> core_stacks_4_layers_2_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297544192)))];
            tensor<fp16, [1, 384, 1, 1024]> proj_29_cast_fp16 = conv(bias = core_stacks_4_layers_2_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_29_dilations_0, groups = proj_29_groups_0, pad = proj_29_pad_0, pad_type = proj_29_pad_type_0, strides = proj_29_strides_0, weight = core_stacks_4_layers_2_self_attn_weights_in_proj_weight_to_fp16, x = input_451_cast_fp16)[name = tensor<string, []>("proj_29_cast_fp16")];
            tensor<int32, [4]> var_5046_begin_0 = const()[name = tensor<string, []>("op_5046_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5046_end_0 = const()[name = tensor<string, []>("op_5046_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1024])];
            tensor<bool, [4]> var_5046_end_mask_0 = const()[name = tensor<string, []>("op_5046_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_5046_cast_fp16 = slice_by_index(begin = var_5046_begin_0, end = var_5046_end_0, end_mask = var_5046_end_mask_0, x = proj_29_cast_fp16)[name = tensor<string, []>("op_5046_cast_fp16")];
            tensor<int32, [4]> var_5047 = const()[name = tensor<string, []>("op_5047"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_5048_cast_fp16 = reshape(shape = var_5047, x = var_5046_cast_fp16)[name = tensor<string, []>("op_5048_cast_fp16")];
            tensor<int32, [4]> var_5050_begin_0 = const()[name = tensor<string, []>("op_5050_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_5050_end_0 = const()[name = tensor<string, []>("op_5050_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1024])];
            tensor<bool, [4]> var_5050_end_mask_0 = const()[name = tensor<string, []>("op_5050_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_5050_cast_fp16 = slice_by_index(begin = var_5050_begin_0, end = var_5050_end_0, end_mask = var_5050_end_mask_0, x = proj_29_cast_fp16)[name = tensor<string, []>("op_5050_cast_fp16")];
            tensor<int32, [4]> var_5051 = const()[name = tensor<string, []>("op_5051"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_5052_cast_fp16 = reshape(shape = var_5051, x = var_5050_cast_fp16)[name = tensor<string, []>("op_5052_cast_fp16")];
            tensor<int32, [4]> var_5054_begin_0 = const()[name = tensor<string, []>("op_5054_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_5054_end_0 = const()[name = tensor<string, []>("op_5054_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_5054_end_mask_0 = const()[name = tensor<string, []>("op_5054_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_5054_cast_fp16 = slice_by_index(begin = var_5054_begin_0, end = var_5054_end_0, end_mask = var_5054_end_mask_0, x = proj_29_cast_fp16)[name = tensor<string, []>("op_5054_cast_fp16")];
            tensor<int32, [4]> var_5055 = const()[name = tensor<string, []>("op_5055"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> pt_29_cast_fp16 = reshape(shape = var_5055, x = var_5054_cast_fp16)[name = tensor<string, []>("pt_29_cast_fp16")];
            tensor<int32, [4]> transpose_112_perm_0 = const()[name = tensor<string, []>("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_564 = const()[name = tensor<string, []>("concat_564"), val = tensor<int32, [3]>([4, 1024, 32])];
            tensor<fp16, [4, 1, 1024, 32]> transpose_112_cast_fp16 = transpose(perm = transpose_112_perm_0, x = var_5048_cast_fp16)[name = tensor<string, []>("transpose_152")];
            tensor<fp16, [4, 1024, 32]> reshape_168_cast_fp16 = reshape(shape = concat_564, x = transpose_112_cast_fp16)[name = tensor<string, []>("reshape_168_cast_fp16")];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = tensor<string, []>("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_565 = const()[name = tensor<string, []>("concat_565"), val = tensor<int32, [3]>([4, 32, 1024])];
            tensor<fp16, [4, 1, 32, 1024]> transpose_113_cast_fp16 = transpose(perm = transpose_113_perm_0, x = var_5052_cast_fp16)[name = tensor<string, []>("transpose_151")];
            tensor<fp16, [4, 32, 1024]> reshape_169_cast_fp16 = reshape(shape = concat_565, x = transpose_113_cast_fp16)[name = tensor<string, []>("reshape_169_cast_fp16")];
            tensor<bool, []> matmul_56_transpose_x_0 = const()[name = tensor<string, []>("matmul_56_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_56_transpose_y_0 = const()[name = tensor<string, []>("matmul_56_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 1024]> matmul_56_cast_fp16 = matmul(transpose_x = matmul_56_transpose_x_0, transpose_y = matmul_56_transpose_y_0, x = reshape_168_cast_fp16, y = reshape_169_cast_fp16)[name = tensor<string, []>("matmul_56_cast_fp16")];
            tensor<int32, [4]> concat_569 = const()[name = tensor<string, []>("concat_569"), val = tensor<int32, [4]>([4, 1, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> reshape_170_cast_fp16 = reshape(shape = concat_569, x = matmul_56_cast_fp16)[name = tensor<string, []>("reshape_170_cast_fp16")];
            tensor<int32, [4]> attn_scores_85_perm_0 = const()[name = tensor<string, []>("attn_scores_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_5059 = const()[name = tensor<string, []>("op_5059"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 1024, 1]> var_5060_cast_fp16 = transpose(perm = var_5059, x = pt_29_cast_fp16)[name = tensor<string, []>("transpose_149")];
            tensor<fp16, [4, 32, 1024, 1024]> var_5061_cast_fp16 = mul(x = var_5060_cast_fp16, y = core_pos_basis_1024_to_fp16)[name = tensor<string, []>("op_5061_cast_fp16")];
            tensor<int32, [1]> pos_scores_29_axes_0 = const()[name = tensor<string, []>("pos_scores_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_29_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_29_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 1024, 1024]> pos_scores_29_cast_fp16 = reduce_sum(axes = pos_scores_29_axes_0, keep_dims = pos_scores_29_keep_dims_0, x = var_5061_cast_fp16)[name = tensor<string, []>("pos_scores_29_cast_fp16")];
            tensor<int32, [4]> var_5064 = const()[name = tensor<string, []>("op_5064"), val = tensor<int32, [4]>([4, 1024, 1, 1024])];
            tensor<fp16, [4, 1024, 1, 1024]> var_5065_cast_fp16 = reshape(shape = var_5064, x = pos_scores_29_cast_fp16)[name = tensor<string, []>("op_5065_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_85_cast_fp16 = transpose(perm = attn_scores_85_perm_0, x = reshape_170_cast_fp16)[name = tensor<string, []>("transpose_150")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_87_cast_fp16 = add(x = attn_scores_85_cast_fp16, y = var_5065_cast_fp16)[name = tensor<string, []>("attn_scores_87_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_89_cast_fp16 = add(x = attn_scores_87_cast_fp16, y = attn_bias_1_cast_fp16)[name = tensor<string, []>("attn_scores_89_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_weights_29_cast_fp16 = softmax(axis = var_41, x = attn_scores_89_cast_fp16)[name = tensor<string, []>("attn_weights_29_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_453_cast_fp16 = add(x = input_451_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("input_453_cast_fp16")];
            tensor<string, []> x_635_pad_type_0 = const()[name = tensor<string, []>("x_635_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_635_strides_0 = const()[name = tensor<string, []>("x_635_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_635_pad_0 = const()[name = tensor<string, []>("x_635_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_635_dilations_0 = const()[name = tensor<string, []>("x_635_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_635_groups_0 = const()[name = tensor<string, []>("x_635_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_4_layers_2_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297545024)))];
            tensor<fp16, [1152]> core_stacks_4_layers_2_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(298724736)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_635_cast_fp16 = conv(bias = core_stacks_4_layers_2_feed_forward1_in_proj_bias_to_fp16, dilations = x_635_dilations_0, groups = x_635_groups_0, pad = x_635_pad_0, pad_type = x_635_pad_type_0, strides = x_635_strides_0, weight = core_stacks_4_layers_2_feed_forward1_in_proj_weight_to_fp16, x = input_453_cast_fp16)[name = tensor<string, []>("x_635_cast_fp16")];
            tensor<fp16, []> var_5079_to_fp16 = const()[name = tensor<string, []>("op_5079_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 1024]> x_637_cast_fp16 = sub(x = x_635_cast_fp16, y = var_5079_to_fp16)[name = tensor<string, []>("x_637_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_5081_cast_fp16 = relu(x = x_637_cast_fp16)[name = tensor<string, []>("op_5081_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_5082_cast_fp16 = abs(x = x_637_cast_fp16)[name = tensor<string, []>("op_5082_cast_fp16")];
            tensor<fp16, []> const_112_promoted_to_fp16 = const()[name = tensor<string, []>("const_112_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_5083_cast_fp16 = mul(x = var_5082_cast_fp16, y = const_112_promoted_to_fp16)[name = tensor<string, []>("op_5083_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_5084_cast_fp16 = exp(x = var_5083_cast_fp16)[name = tensor<string, []>("op_5084_cast_fp16")];
            tensor<fp32, []> var_5085_epsilon_0 = const()[name = tensor<string, []>("op_5085_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_5085_cast_fp16 = log(epsilon = var_5085_epsilon_0, x = var_5084_cast_fp16)[name = tensor<string, []>("op_5085_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_5086_cast_fp16 = add(x = var_5081_cast_fp16, y = var_5085_cast_fp16)[name = tensor<string, []>("op_5086_cast_fp16")];
            tensor<fp16, []> var_5087_to_fp16 = const()[name = tensor<string, []>("op_5087_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 1024]> var_5088_cast_fp16 = mul(x = x_635_cast_fp16, y = var_5087_to_fp16)[name = tensor<string, []>("op_5088_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_5089_cast_fp16 = sub(x = var_5086_cast_fp16, y = var_5088_cast_fp16)[name = tensor<string, []>("op_5089_cast_fp16")];
            tensor<fp16, []> var_5090_to_fp16 = const()[name = tensor<string, []>("op_5090_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 1024]> input_455_cast_fp16 = sub(x = var_5089_cast_fp16, y = var_5090_to_fp16)[name = tensor<string, []>("input_455_cast_fp16")];
            tensor<string, []> var_5098_pad_type_0 = const()[name = tensor<string, []>("op_5098_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5098_strides_0 = const()[name = tensor<string, []>("op_5098_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5098_pad_0 = const()[name = tensor<string, []>("op_5098_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5098_dilations_0 = const()[name = tensor<string, []>("op_5098_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5098_groups_0 = const()[name = tensor<string, []>("op_5098_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_4_layers_2_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(298727104)))];
            tensor<fp16, [512]> core_stacks_4_layers_2_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299906816)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5098_cast_fp16 = conv(bias = core_stacks_4_layers_2_feed_forward1_out_proj_bias_to_fp16, dilations = var_5098_dilations_0, groups = var_5098_groups_0, pad = var_5098_pad_0, pad_type = var_5098_pad_type_0, strides = var_5098_strides_0, weight = core_stacks_4_layers_2_feed_forward1_out_proj_weight_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("op_5098_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_457_cast_fp16 = add(x = input_453_cast_fp16, y = var_5098_cast_fp16)[name = tensor<string, []>("input_457_cast_fp16")];
            tensor<int32, [4]> var_5100_begin_0 = const()[name = tensor<string, []>("op_5100_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5100_end_0 = const()[name = tensor<string, []>("op_5100_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_5100_end_mask_0 = const()[name = tensor<string, []>("op_5100_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> var_5100_cast_fp16 = slice_by_index(begin = var_5100_begin_0, end = var_5100_end_0, end_mask = var_5100_end_mask_0, x = attn_weights_29_cast_fp16)[name = tensor<string, []>("op_5100_cast_fp16")];
            tensor<string, []> x_639_pad_type_0 = const()[name = tensor<string, []>("x_639_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_639_strides_0 = const()[name = tensor<string, []>("x_639_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_639_pad_0 = const()[name = tensor<string, []>("x_639_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_639_dilations_0 = const()[name = tensor<string, []>("x_639_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_639_groups_0 = const()[name = tensor<string, []>("x_639_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_4_layers_2_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299907904)))];
            tensor<fp16, [1152]> core_stacks_4_layers_2_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301087616)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_639_cast_fp16 = conv(bias = core_stacks_4_layers_2_nonlin_attention_in_proj_bias_to_fp16, dilations = x_639_dilations_0, groups = x_639_groups_0, pad = x_639_pad_0, pad_type = x_639_pad_type_0, strides = x_639_strides_0, weight = core_stacks_4_layers_2_nonlin_attention_in_proj_weight_to_fp16, x = input_457_cast_fp16)[name = tensor<string, []>("x_639_cast_fp16")];
            tensor<int32, [4]> var_5111_begin_0 = const()[name = tensor<string, []>("op_5111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5111_end_0 = const()[name = tensor<string, []>("op_5111_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_5111_end_mask_0 = const()[name = tensor<string, []>("op_5111_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_5111_cast_fp16 = slice_by_index(begin = var_5111_begin_0, end = var_5111_end_0, end_mask = var_5111_end_mask_0, x = x_639_cast_fp16)[name = tensor<string, []>("op_5111_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> s_gate_29_cast_fp16 = tanh(x = var_5111_cast_fp16)[name = tensor<string, []>("s_gate_29_cast_fp16")];
            tensor<int32, [4]> var_5114_begin_0 = const()[name = tensor<string, []>("op_5114_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_5114_end_0 = const()[name = tensor<string, []>("op_5114_end_0"), val = tensor<int32, [4]>([1, 768, 1, 1024])];
            tensor<bool, [4]> var_5114_end_mask_0 = const()[name = tensor<string, []>("op_5114_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_5114_cast_fp16 = slice_by_index(begin = var_5114_begin_0, end = var_5114_end_0, end_mask = var_5114_end_mask_0, x = x_639_cast_fp16)[name = tensor<string, []>("op_5114_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> var_5115_cast_fp16 = mul(x = var_5114_cast_fp16, y = s_gate_29_cast_fp16)[name = tensor<string, []>("op_5115_cast_fp16")];
            tensor<int32, [4]> y_29_begin_0 = const()[name = tensor<string, []>("y_29_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_29_end_0 = const()[name = tensor<string, []>("y_29_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 1024])];
            tensor<bool, [4]> y_29_end_mask_0 = const()[name = tensor<string, []>("y_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> y_29_cast_fp16 = slice_by_index(begin = y_29_begin_0, end = y_29_end_0, end_mask = y_29_end_mask_0, x = x_639_cast_fp16)[name = tensor<string, []>("y_29_cast_fp16")];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = tensor<string, []>("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_574 = const()[name = tensor<string, []>("concat_574"), val = tensor<int32, [3]>([1, 1024, 1024])];
            tensor<fp16, [1, 1, 1024, 1024]> transpose_114_cast_fp16 = transpose(perm = transpose_114_perm_0, x = var_5100_cast_fp16)[name = tensor<string, []>("transpose_148")];
            tensor<fp16, [1, 1024, 1024]> reshape_171_cast_fp16 = reshape(shape = concat_574, x = transpose_114_cast_fp16)[name = tensor<string, []>("reshape_171_cast_fp16")];
            tensor<int32, [4]> transpose_115_perm_0 = const()[name = tensor<string, []>("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_575 = const()[name = tensor<string, []>("concat_575"), val = tensor<int32, [3]>([1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> transpose_115_cast_fp16 = transpose(perm = transpose_115_perm_0, x = var_5115_cast_fp16)[name = tensor<string, []>("transpose_147")];
            tensor<fp16, [1, 1024, 384]> reshape_172_cast_fp16 = reshape(shape = concat_575, x = transpose_115_cast_fp16)[name = tensor<string, []>("reshape_172_cast_fp16")];
            tensor<bool, []> matmul_57_transpose_x_0 = const()[name = tensor<string, []>("matmul_57_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_57_transpose_y_0 = const()[name = tensor<string, []>("matmul_57_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1024, 384]> matmul_57_cast_fp16 = matmul(transpose_x = matmul_57_transpose_x_0, transpose_y = matmul_57_transpose_y_0, x = reshape_171_cast_fp16, y = reshape_172_cast_fp16)[name = tensor<string, []>("matmul_57_cast_fp16")];
            tensor<int32, [4]> concat_579 = const()[name = tensor<string, []>("concat_579"), val = tensor<int32, [4]>([1, 1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> reshape_173_cast_fp16 = reshape(shape = concat_579, x = matmul_57_cast_fp16)[name = tensor<string, []>("reshape_173_cast_fp16")];
            tensor<int32, [4]> v_29_perm_0 = const()[name = tensor<string, []>("v_29_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 1024]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = reshape_173_cast_fp16)[name = tensor<string, []>("transpose_146")];
            tensor<fp16, [1, 384, 1, 1024]> input_459_cast_fp16 = mul(x = v_29_cast_fp16, y = y_29_cast_fp16)[name = tensor<string, []>("input_459_cast_fp16")];
            tensor<string, []> var_5127_pad_type_0 = const()[name = tensor<string, []>("op_5127_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5127_strides_0 = const()[name = tensor<string, []>("op_5127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5127_pad_0 = const()[name = tensor<string, []>("op_5127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5127_dilations_0 = const()[name = tensor<string, []>("op_5127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5127_groups_0 = const()[name = tensor<string, []>("op_5127_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_4_layers_2_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301089984)))];
            tensor<fp16, [512]> core_stacks_4_layers_2_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301483264)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5127_cast_fp16 = conv(bias = core_stacks_4_layers_2_nonlin_attention_out_proj_bias_to_fp16, dilations = var_5127_dilations_0, groups = var_5127_groups_0, pad = var_5127_pad_0, pad_type = var_5127_pad_type_0, strides = var_5127_strides_0, weight = core_stacks_4_layers_2_nonlin_attention_out_proj_weight_to_fp16, x = input_459_cast_fp16)[name = tensor<string, []>("op_5127_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_641_cast_fp16 = add(x = input_457_cast_fp16, y = var_5127_cast_fp16)[name = tensor<string, []>("x_641_cast_fp16")];
            tensor<string, []> var_5138_pad_type_0 = const()[name = tensor<string, []>("op_5138_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5138_strides_0 = const()[name = tensor<string, []>("op_5138_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5138_pad_0 = const()[name = tensor<string, []>("op_5138_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5138_dilations_0 = const()[name = tensor<string, []>("op_5138_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5138_groups_0 = const()[name = tensor<string, []>("op_5138_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_4_layers_2_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301484352)))];
            tensor<fp16, [48]> core_stacks_4_layers_2_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301533568)))];
            tensor<fp16, [1, 48, 1, 1024]> var_5138_cast_fp16 = conv(bias = core_stacks_4_layers_2_self_attn1_in_proj_bias_to_fp16, dilations = var_5138_dilations_0, groups = var_5138_groups_0, pad = var_5138_pad_0, pad_type = var_5138_pad_type_0, strides = var_5138_strides_0, weight = core_stacks_4_layers_2_self_attn1_in_proj_weight_to_fp16, x = x_641_cast_fp16)[name = tensor<string, []>("op_5138_cast_fp16")];
            tensor<int32, [4]> var_5139 = const()[name = tensor<string, []>("op_5139"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_5140_cast_fp16 = reshape(shape = var_5139, x = var_5138_cast_fp16)[name = tensor<string, []>("op_5140_cast_fp16")];
            tensor<int32, [4]> transpose_116_perm_0 = const()[name = tensor<string, []>("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_584 = const()[name = tensor<string, []>("concat_584"), val = tensor<int32, [3]>([4, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> transpose_116_cast_fp16 = transpose(perm = transpose_116_perm_0, x = attn_weights_29_cast_fp16)[name = tensor<string, []>("transpose_145")];
            tensor<fp16, [4, 1024, 1024]> reshape_174_cast_fp16 = reshape(shape = concat_584, x = transpose_116_cast_fp16)[name = tensor<string, []>("reshape_174_cast_fp16")];
            tensor<int32, [4]> transpose_117_perm_0 = const()[name = tensor<string, []>("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_585 = const()[name = tensor<string, []>("concat_585"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_117_cast_fp16 = transpose(perm = transpose_117_perm_0, x = var_5140_cast_fp16)[name = tensor<string, []>("transpose_144")];
            tensor<fp16, [4, 1024, 12]> reshape_175_cast_fp16 = reshape(shape = concat_585, x = transpose_117_cast_fp16)[name = tensor<string, []>("reshape_175_cast_fp16")];
            tensor<bool, []> matmul_58_transpose_x_0 = const()[name = tensor<string, []>("matmul_58_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_58_transpose_y_0 = const()[name = tensor<string, []>("matmul_58_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_58_cast_fp16 = matmul(transpose_x = matmul_58_transpose_x_0, transpose_y = matmul_58_transpose_y_0, x = reshape_174_cast_fp16, y = reshape_175_cast_fp16)[name = tensor<string, []>("matmul_58_cast_fp16")];
            tensor<int32, [4]> concat_589 = const()[name = tensor<string, []>("concat_589"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_176_cast_fp16 = reshape(shape = concat_589, x = matmul_58_cast_fp16)[name = tensor<string, []>("reshape_176_cast_fp16")];
            tensor<int32, [4]> out_57_perm_0 = const()[name = tensor<string, []>("out_57_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_5143 = const()[name = tensor<string, []>("op_5143"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_57_cast_fp16 = transpose(perm = out_57_perm_0, x = reshape_176_cast_fp16)[name = tensor<string, []>("transpose_143")];
            tensor<fp16, [1, 48, 1, 1024]> input_461_cast_fp16 = reshape(shape = var_5143, x = out_57_cast_fp16)[name = tensor<string, []>("input_461_cast_fp16")];
            tensor<string, []> var_5151_pad_type_0 = const()[name = tensor<string, []>("op_5151_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5151_strides_0 = const()[name = tensor<string, []>("op_5151_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5151_pad_0 = const()[name = tensor<string, []>("op_5151_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5151_dilations_0 = const()[name = tensor<string, []>("op_5151_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5151_groups_0 = const()[name = tensor<string, []>("op_5151_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_4_layers_2_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301533760)))];
            tensor<fp16, [512]> core_stacks_4_layers_2_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301582976)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5151_cast_fp16 = conv(bias = core_stacks_4_layers_2_self_attn1_out_proj_bias_to_fp16, dilations = var_5151_dilations_0, groups = var_5151_groups_0, pad = var_5151_pad_0, pad_type = var_5151_pad_type_0, strides = var_5151_strides_0, weight = core_stacks_4_layers_2_self_attn1_out_proj_weight_to_fp16, x = input_461_cast_fp16)[name = tensor<string, []>("op_5151_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_643_cast_fp16 = add(x = x_641_cast_fp16, y = var_5151_cast_fp16)[name = tensor<string, []>("x_643_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_645_cast_fp16 = add(x = x_643_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("x_645_cast_fp16")];
            tensor<string, []> x_647_pad_type_0 = const()[name = tensor<string, []>("x_647_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_647_strides_0 = const()[name = tensor<string, []>("x_647_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_647_pad_0 = const()[name = tensor<string, []>("x_647_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_647_dilations_0 = const()[name = tensor<string, []>("x_647_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_647_groups_0 = const()[name = tensor<string, []>("x_647_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_4_layers_2_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301584064)))];
            tensor<fp16, [1024]> core_stacks_4_layers_2_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302632704)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_647_cast_fp16 = conv(bias = core_stacks_4_layers_2_conv_module1_in_proj_bias_to_fp16, dilations = x_647_dilations_0, groups = x_647_groups_0, pad = x_647_pad_0, pad_type = x_647_pad_type_0, strides = x_647_strides_0, weight = core_stacks_4_layers_2_conv_module1_in_proj_weight_to_fp16, x = x_645_cast_fp16)[name = tensor<string, []>("x_647_cast_fp16")];
            tensor<int32, [4]> var_5169_begin_0 = const()[name = tensor<string, []>("op_5169_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5169_end_0 = const()[name = tensor<string, []>("op_5169_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_5169_end_mask_0 = const()[name = tensor<string, []>("op_5169_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5169_cast_fp16 = slice_by_index(begin = var_5169_begin_0, end = var_5169_end_0, end_mask = var_5169_end_mask_0, x = x_647_cast_fp16)[name = tensor<string, []>("op_5169_cast_fp16")];
            tensor<int32, [4]> var_5171_begin_0 = const()[name = tensor<string, []>("op_5171_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_5171_end_0 = const()[name = tensor<string, []>("op_5171_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_5171_end_mask_0 = const()[name = tensor<string, []>("op_5171_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5171_cast_fp16 = slice_by_index(begin = var_5171_begin_0, end = var_5171_end_0, end_mask = var_5171_end_mask_0, x = x_647_cast_fp16)[name = tensor<string, []>("op_5171_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5172_cast_fp16 = sigmoid(x = var_5171_cast_fp16)[name = tensor<string, []>("op_5172_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_649_cast_fp16 = mul(x = var_5169_cast_fp16, y = var_5172_cast_fp16)[name = tensor<string, []>("x_649_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_463_cast_fp16 = mul(x = x_649_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_463_cast_fp16")];
            tensor<string, []> var_5181_pad_type_0 = const()[name = tensor<string, []>("op_5181_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5181_pad_0 = const()[name = tensor<string, []>("op_5181_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5181_groups_0 = const()[name = tensor<string, []>("op_5181_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5181_strides_0 = const()[name = tensor<string, []>("op_5181_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5181_dilations_0 = const()[name = tensor<string, []>("op_5181_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_2_conv_module1_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module1_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302634816)))];
            tensor<fp16, [512]> core_stacks_4_layers_2_conv_module1_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module1_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302650240)))];
            tensor<fp16, [1, 512, 1, 1040]> var_5181_cast_fp16 = conv(bias = core_stacks_4_layers_2_conv_module1_dw_a_bias_to_fp16, dilations = var_5181_dilations_0, groups = var_5181_groups_0, pad = var_5181_pad_0, pad_type = var_5181_pad_type_0, strides = var_5181_strides_0, weight = core_stacks_4_layers_2_conv_module1_dw_a_weight_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("op_5181_cast_fp16")];
            tensor<int32, [4]> var_5182_begin_0 = const()[name = tensor<string, []>("op_5182_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5182_end_0 = const()[name = tensor<string, []>("op_5182_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_5182_end_mask_0 = const()[name = tensor<string, []>("op_5182_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_5182_cast_fp16 = slice_by_index(begin = var_5182_begin_0, end = var_5182_end_0, end_mask = var_5182_end_mask_0, x = var_5181_cast_fp16)[name = tensor<string, []>("op_5182_cast_fp16")];
            tensor<string, []> var_5188_pad_type_0 = const()[name = tensor<string, []>("op_5188_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5188_pad_0 = const()[name = tensor<string, []>("op_5188_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5188_groups_0 = const()[name = tensor<string, []>("op_5188_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5188_strides_0 = const()[name = tensor<string, []>("op_5188_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5188_dilations_0 = const()[name = tensor<string, []>("op_5188_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_2_conv_module1_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module1_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302651328)))];
            tensor<fp16, [1, 512, 1, 1040]> var_5188_cast_fp16 = conv(dilations = var_5188_dilations_0, groups = var_5188_groups_0, pad = var_5188_pad_0, pad_type = var_5188_pad_type_0, strides = var_5188_strides_0, weight = core_stacks_4_layers_2_conv_module1_dw_b_weight_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("op_5188_cast_fp16")];
            tensor<int32, [4]> var_5192_begin_0 = const()[name = tensor<string, []>("op_5192_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_5192_end_0 = const()[name = tensor<string, []>("op_5192_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_5192_end_mask_0 = const()[name = tensor<string, []>("op_5192_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_5192_cast_fp16 = slice_by_index(begin = var_5192_begin_0, end = var_5192_end_0, end_mask = var_5192_end_mask_0, x = var_5188_cast_fp16)[name = tensor<string, []>("op_5192_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5193_cast_fp16 = add(x = var_5182_cast_fp16, y = var_5192_cast_fp16)[name = tensor<string, []>("op_5193_cast_fp16")];
            tensor<string, []> var_5199_pad_type_0 = const()[name = tensor<string, []>("op_5199_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5199_pad_0 = const()[name = tensor<string, []>("op_5199_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5199_groups_0 = const()[name = tensor<string, []>("op_5199_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5199_strides_0 = const()[name = tensor<string, []>("op_5199_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5199_dilations_0 = const()[name = tensor<string, []>("op_5199_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_4_layers_2_conv_module1_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module1_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302666752)))];
            tensor<fp16, [1, 512, 1, 1054]> var_5199_cast_fp16 = conv(dilations = var_5199_dilations_0, groups = var_5199_groups_0, pad = var_5199_pad_0, pad_type = var_5199_pad_type_0, strides = var_5199_strides_0, weight = core_stacks_4_layers_2_conv_module1_dw_c_weight_to_fp16, x = input_463_cast_fp16)[name = tensor<string, []>("op_5199_cast_fp16")];
            tensor<int32, [4]> var_5203_begin_0 = const()[name = tensor<string, []>("op_5203_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_5203_end_0 = const()[name = tensor<string, []>("op_5203_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_5203_end_mask_0 = const()[name = tensor<string, []>("op_5203_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5203_cast_fp16 = slice_by_index(begin = var_5203_begin_0, end = var_5203_end_0, end_mask = var_5203_end_mask_0, x = var_5199_cast_fp16)[name = tensor<string, []>("op_5203_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_651_cast_fp16 = add(x = var_5193_cast_fp16, y = var_5203_cast_fp16)[name = tensor<string, []>("x_651_cast_fp16")];
            tensor<fp16, []> var_5205_to_fp16 = const()[name = tensor<string, []>("op_5205_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_653_cast_fp16 = sub(x = x_651_cast_fp16, y = var_5205_to_fp16)[name = tensor<string, []>("x_653_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5207_cast_fp16 = relu(x = x_653_cast_fp16)[name = tensor<string, []>("op_5207_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5208_cast_fp16 = abs(x = x_653_cast_fp16)[name = tensor<string, []>("op_5208_cast_fp16")];
            tensor<fp16, []> const_115_promoted_to_fp16 = const()[name = tensor<string, []>("const_115_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_5209_cast_fp16 = mul(x = var_5208_cast_fp16, y = const_115_promoted_to_fp16)[name = tensor<string, []>("op_5209_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5210_cast_fp16 = exp(x = var_5209_cast_fp16)[name = tensor<string, []>("op_5210_cast_fp16")];
            tensor<fp32, []> var_5211_epsilon_0 = const()[name = tensor<string, []>("op_5211_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_5211_cast_fp16 = log(epsilon = var_5211_epsilon_0, x = var_5210_cast_fp16)[name = tensor<string, []>("op_5211_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5212_cast_fp16 = add(x = var_5207_cast_fp16, y = var_5211_cast_fp16)[name = tensor<string, []>("op_5212_cast_fp16")];
            tensor<fp16, []> var_5213_to_fp16 = const()[name = tensor<string, []>("op_5213_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_5214_cast_fp16 = mul(x = x_651_cast_fp16, y = var_5213_to_fp16)[name = tensor<string, []>("op_5214_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5215_cast_fp16 = sub(x = var_5212_cast_fp16, y = var_5214_cast_fp16)[name = tensor<string, []>("op_5215_cast_fp16")];
            tensor<fp16, []> var_5216_to_fp16 = const()[name = tensor<string, []>("op_5216_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_465_cast_fp16 = sub(x = var_5215_cast_fp16, y = var_5216_to_fp16)[name = tensor<string, []>("input_465_cast_fp16")];
            tensor<string, []> var_5224_pad_type_0 = const()[name = tensor<string, []>("op_5224_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5224_strides_0 = const()[name = tensor<string, []>("op_5224_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5224_pad_0 = const()[name = tensor<string, []>("op_5224_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5224_dilations_0 = const()[name = tensor<string, []>("op_5224_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5224_groups_0 = const()[name = tensor<string, []>("op_5224_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_4_layers_2_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302667840)))];
            tensor<fp16, [512]> core_stacks_4_layers_2_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303192192)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5224_cast_fp16 = conv(bias = core_stacks_4_layers_2_conv_module1_out_proj_bias_to_fp16, dilations = var_5224_dilations_0, groups = var_5224_groups_0, pad = var_5224_pad_0, pad_type = var_5224_pad_type_0, strides = var_5224_strides_0, weight = core_stacks_4_layers_2_conv_module1_out_proj_weight_to_fp16, x = input_465_cast_fp16)[name = tensor<string, []>("op_5224_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_467_cast_fp16 = add(x = x_645_cast_fp16, y = var_5224_cast_fp16)[name = tensor<string, []>("input_467_cast_fp16")];
            tensor<string, []> x_655_pad_type_0 = const()[name = tensor<string, []>("x_655_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_655_strides_0 = const()[name = tensor<string, []>("x_655_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_655_pad_0 = const()[name = tensor<string, []>("x_655_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_655_dilations_0 = const()[name = tensor<string, []>("x_655_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_655_groups_0 = const()[name = tensor<string, []>("x_655_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_4_layers_2_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303193280)))];
            tensor<fp16, [1536]> core_stacks_4_layers_2_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304766208)))];
            tensor<fp16, [1, 1536, 1, 1024]> x_655_cast_fp16 = conv(bias = core_stacks_4_layers_2_feed_forward2_in_proj_bias_to_fp16, dilations = x_655_dilations_0, groups = x_655_groups_0, pad = x_655_pad_0, pad_type = x_655_pad_type_0, strides = x_655_strides_0, weight = core_stacks_4_layers_2_feed_forward2_in_proj_weight_to_fp16, x = input_467_cast_fp16)[name = tensor<string, []>("x_655_cast_fp16")];
            tensor<fp16, []> var_5235_to_fp16 = const()[name = tensor<string, []>("op_5235_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 1024]> x_657_cast_fp16 = sub(x = x_655_cast_fp16, y = var_5235_to_fp16)[name = tensor<string, []>("x_657_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_5237_cast_fp16 = relu(x = x_657_cast_fp16)[name = tensor<string, []>("op_5237_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_5238_cast_fp16 = abs(x = x_657_cast_fp16)[name = tensor<string, []>("op_5238_cast_fp16")];
            tensor<fp16, []> const_116_promoted_to_fp16 = const()[name = tensor<string, []>("const_116_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_5239_cast_fp16 = mul(x = var_5238_cast_fp16, y = const_116_promoted_to_fp16)[name = tensor<string, []>("op_5239_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_5240_cast_fp16 = exp(x = var_5239_cast_fp16)[name = tensor<string, []>("op_5240_cast_fp16")];
            tensor<fp32, []> var_5241_epsilon_0 = const()[name = tensor<string, []>("op_5241_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_5241_cast_fp16 = log(epsilon = var_5241_epsilon_0, x = var_5240_cast_fp16)[name = tensor<string, []>("op_5241_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_5242_cast_fp16 = add(x = var_5237_cast_fp16, y = var_5241_cast_fp16)[name = tensor<string, []>("op_5242_cast_fp16")];
            tensor<fp16, []> var_5243_to_fp16 = const()[name = tensor<string, []>("op_5243_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 1024]> var_5244_cast_fp16 = mul(x = x_655_cast_fp16, y = var_5243_to_fp16)[name = tensor<string, []>("op_5244_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_5245_cast_fp16 = sub(x = var_5242_cast_fp16, y = var_5244_cast_fp16)[name = tensor<string, []>("op_5245_cast_fp16")];
            tensor<fp16, []> var_5246_to_fp16 = const()[name = tensor<string, []>("op_5246_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 1024]> input_469_cast_fp16 = sub(x = var_5245_cast_fp16, y = var_5246_to_fp16)[name = tensor<string, []>("input_469_cast_fp16")];
            tensor<string, []> var_5254_pad_type_0 = const()[name = tensor<string, []>("op_5254_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5254_strides_0 = const()[name = tensor<string, []>("op_5254_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5254_pad_0 = const()[name = tensor<string, []>("op_5254_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5254_dilations_0 = const()[name = tensor<string, []>("op_5254_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5254_groups_0 = const()[name = tensor<string, []>("op_5254_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_4_layers_2_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304769344)))];
            tensor<fp16, [512]> core_stacks_4_layers_2_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306342272)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5254_cast_fp16 = conv(bias = core_stacks_4_layers_2_feed_forward2_out_proj_bias_to_fp16, dilations = var_5254_dilations_0, groups = var_5254_groups_0, pad = var_5254_pad_0, pad_type = var_5254_pad_type_0, strides = var_5254_strides_0, weight = core_stacks_4_layers_2_feed_forward2_out_proj_weight_to_fp16, x = input_469_cast_fp16)[name = tensor<string, []>("op_5254_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> src_63_cast_fp16 = add(x = input_467_cast_fp16, y = var_5254_cast_fp16)[name = tensor<string, []>("src_63_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5257_cast_fp16 = sub(x = src_63_cast_fp16, y = input_451_cast_fp16)[name = tensor<string, []>("op_5257_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_2_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306343360)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5258_cast_fp16 = mul(x = var_5257_cast_fp16, y = core_stacks_4_layers_2_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_5258_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_659_cast_fp16 = add(x = input_451_cast_fp16, y = var_5258_cast_fp16)[name = tensor<string, []>("x_659_cast_fp16")];
            tensor<string, []> var_5269_pad_type_0 = const()[name = tensor<string, []>("op_5269_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5269_strides_0 = const()[name = tensor<string, []>("op_5269_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5269_pad_0 = const()[name = tensor<string, []>("op_5269_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5269_dilations_0 = const()[name = tensor<string, []>("op_5269_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5269_groups_0 = const()[name = tensor<string, []>("op_5269_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_4_layers_2_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306344448)))];
            tensor<fp16, [48]> core_stacks_4_layers_2_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306393664)))];
            tensor<fp16, [1, 48, 1, 1024]> var_5269_cast_fp16 = conv(bias = core_stacks_4_layers_2_self_attn2_in_proj_bias_to_fp16, dilations = var_5269_dilations_0, groups = var_5269_groups_0, pad = var_5269_pad_0, pad_type = var_5269_pad_type_0, strides = var_5269_strides_0, weight = core_stacks_4_layers_2_self_attn2_in_proj_weight_to_fp16, x = x_659_cast_fp16)[name = tensor<string, []>("op_5269_cast_fp16")];
            tensor<int32, [4]> var_5270 = const()[name = tensor<string, []>("op_5270"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_5271_cast_fp16 = reshape(shape = var_5270, x = var_5269_cast_fp16)[name = tensor<string, []>("op_5271_cast_fp16")];
            tensor<int32, [4]> transpose_119_perm_0 = const()[name = tensor<string, []>("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_595 = const()[name = tensor<string, []>("concat_595"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_119_cast_fp16 = transpose(perm = transpose_119_perm_0, x = var_5271_cast_fp16)[name = tensor<string, []>("transpose_142")];
            tensor<fp16, [4, 1024, 12]> reshape_178_cast_fp16 = reshape(shape = concat_595, x = transpose_119_cast_fp16)[name = tensor<string, []>("reshape_178_cast_fp16")];
            tensor<bool, []> matmul_59_transpose_x_0 = const()[name = tensor<string, []>("matmul_59_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_59_transpose_y_0 = const()[name = tensor<string, []>("matmul_59_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_59_cast_fp16 = matmul(transpose_x = matmul_59_transpose_x_0, transpose_y = matmul_59_transpose_y_0, x = reshape_174_cast_fp16, y = reshape_178_cast_fp16)[name = tensor<string, []>("matmul_59_cast_fp16")];
            tensor<int32, [4]> concat_599 = const()[name = tensor<string, []>("concat_599"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_179_cast_fp16 = reshape(shape = concat_599, x = matmul_59_cast_fp16)[name = tensor<string, []>("reshape_179_cast_fp16")];
            tensor<int32, [4]> out_59_perm_0 = const()[name = tensor<string, []>("out_59_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_5274 = const()[name = tensor<string, []>("op_5274"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_59_cast_fp16 = transpose(perm = out_59_perm_0, x = reshape_179_cast_fp16)[name = tensor<string, []>("transpose_141")];
            tensor<fp16, [1, 48, 1, 1024]> input_471_cast_fp16 = reshape(shape = var_5274, x = out_59_cast_fp16)[name = tensor<string, []>("input_471_cast_fp16")];
            tensor<string, []> var_5282_pad_type_0 = const()[name = tensor<string, []>("op_5282_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5282_strides_0 = const()[name = tensor<string, []>("op_5282_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5282_pad_0 = const()[name = tensor<string, []>("op_5282_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5282_dilations_0 = const()[name = tensor<string, []>("op_5282_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5282_groups_0 = const()[name = tensor<string, []>("op_5282_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_4_layers_2_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306393856)))];
            tensor<fp16, [512]> core_stacks_4_layers_2_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306443072)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5282_cast_fp16 = conv(bias = core_stacks_4_layers_2_self_attn2_out_proj_bias_to_fp16, dilations = var_5282_dilations_0, groups = var_5282_groups_0, pad = var_5282_pad_0, pad_type = var_5282_pad_type_0, strides = var_5282_strides_0, weight = core_stacks_4_layers_2_self_attn2_out_proj_weight_to_fp16, x = input_471_cast_fp16)[name = tensor<string, []>("op_5282_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_661_cast_fp16 = add(x = x_659_cast_fp16, y = var_5282_cast_fp16)[name = tensor<string, []>("x_661_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_663_cast_fp16 = add(x = x_661_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("x_663_cast_fp16")];
            tensor<string, []> x_665_pad_type_0 = const()[name = tensor<string, []>("x_665_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_665_strides_0 = const()[name = tensor<string, []>("x_665_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_665_pad_0 = const()[name = tensor<string, []>("x_665_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_665_dilations_0 = const()[name = tensor<string, []>("x_665_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_665_groups_0 = const()[name = tensor<string, []>("x_665_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_4_layers_2_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306444160)))];
            tensor<fp16, [1024]> core_stacks_4_layers_2_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307492800)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_665_cast_fp16 = conv(bias = core_stacks_4_layers_2_conv_module2_in_proj_bias_to_fp16, dilations = x_665_dilations_0, groups = x_665_groups_0, pad = x_665_pad_0, pad_type = x_665_pad_type_0, strides = x_665_strides_0, weight = core_stacks_4_layers_2_conv_module2_in_proj_weight_to_fp16, x = x_663_cast_fp16)[name = tensor<string, []>("x_665_cast_fp16")];
            tensor<int32, [4]> var_5300_begin_0 = const()[name = tensor<string, []>("op_5300_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5300_end_0 = const()[name = tensor<string, []>("op_5300_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_5300_end_mask_0 = const()[name = tensor<string, []>("op_5300_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5300_cast_fp16 = slice_by_index(begin = var_5300_begin_0, end = var_5300_end_0, end_mask = var_5300_end_mask_0, x = x_665_cast_fp16)[name = tensor<string, []>("op_5300_cast_fp16")];
            tensor<int32, [4]> var_5302_begin_0 = const()[name = tensor<string, []>("op_5302_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_5302_end_0 = const()[name = tensor<string, []>("op_5302_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_5302_end_mask_0 = const()[name = tensor<string, []>("op_5302_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5302_cast_fp16 = slice_by_index(begin = var_5302_begin_0, end = var_5302_end_0, end_mask = var_5302_end_mask_0, x = x_665_cast_fp16)[name = tensor<string, []>("op_5302_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5303_cast_fp16 = sigmoid(x = var_5302_cast_fp16)[name = tensor<string, []>("op_5303_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_667_cast_fp16 = mul(x = var_5300_cast_fp16, y = var_5303_cast_fp16)[name = tensor<string, []>("x_667_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_473_cast_fp16 = mul(x = x_667_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_473_cast_fp16")];
            tensor<string, []> var_5312_pad_type_0 = const()[name = tensor<string, []>("op_5312_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5312_pad_0 = const()[name = tensor<string, []>("op_5312_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5312_groups_0 = const()[name = tensor<string, []>("op_5312_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5312_strides_0 = const()[name = tensor<string, []>("op_5312_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5312_dilations_0 = const()[name = tensor<string, []>("op_5312_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_2_conv_module2_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module2_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307494912)))];
            tensor<fp16, [512]> core_stacks_4_layers_2_conv_module2_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module2_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307510336)))];
            tensor<fp16, [1, 512, 1, 1040]> var_5312_cast_fp16 = conv(bias = core_stacks_4_layers_2_conv_module2_dw_a_bias_to_fp16, dilations = var_5312_dilations_0, groups = var_5312_groups_0, pad = var_5312_pad_0, pad_type = var_5312_pad_type_0, strides = var_5312_strides_0, weight = core_stacks_4_layers_2_conv_module2_dw_a_weight_to_fp16, x = input_473_cast_fp16)[name = tensor<string, []>("op_5312_cast_fp16")];
            tensor<int32, [4]> var_5313_begin_0 = const()[name = tensor<string, []>("op_5313_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5313_end_0 = const()[name = tensor<string, []>("op_5313_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_5313_end_mask_0 = const()[name = tensor<string, []>("op_5313_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_5313_cast_fp16 = slice_by_index(begin = var_5313_begin_0, end = var_5313_end_0, end_mask = var_5313_end_mask_0, x = var_5312_cast_fp16)[name = tensor<string, []>("op_5313_cast_fp16")];
            tensor<string, []> var_5319_pad_type_0 = const()[name = tensor<string, []>("op_5319_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5319_pad_0 = const()[name = tensor<string, []>("op_5319_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5319_groups_0 = const()[name = tensor<string, []>("op_5319_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5319_strides_0 = const()[name = tensor<string, []>("op_5319_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5319_dilations_0 = const()[name = tensor<string, []>("op_5319_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_2_conv_module2_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module2_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307511424)))];
            tensor<fp16, [1, 512, 1, 1040]> var_5319_cast_fp16 = conv(dilations = var_5319_dilations_0, groups = var_5319_groups_0, pad = var_5319_pad_0, pad_type = var_5319_pad_type_0, strides = var_5319_strides_0, weight = core_stacks_4_layers_2_conv_module2_dw_b_weight_to_fp16, x = input_473_cast_fp16)[name = tensor<string, []>("op_5319_cast_fp16")];
            tensor<int32, [4]> var_5323_begin_0 = const()[name = tensor<string, []>("op_5323_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_5323_end_0 = const()[name = tensor<string, []>("op_5323_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_5323_end_mask_0 = const()[name = tensor<string, []>("op_5323_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_5323_cast_fp16 = slice_by_index(begin = var_5323_begin_0, end = var_5323_end_0, end_mask = var_5323_end_mask_0, x = var_5319_cast_fp16)[name = tensor<string, []>("op_5323_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5324_cast_fp16 = add(x = var_5313_cast_fp16, y = var_5323_cast_fp16)[name = tensor<string, []>("op_5324_cast_fp16")];
            tensor<string, []> var_5330_pad_type_0 = const()[name = tensor<string, []>("op_5330_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5330_pad_0 = const()[name = tensor<string, []>("op_5330_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5330_groups_0 = const()[name = tensor<string, []>("op_5330_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5330_strides_0 = const()[name = tensor<string, []>("op_5330_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5330_dilations_0 = const()[name = tensor<string, []>("op_5330_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_4_layers_2_conv_module2_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module2_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307526848)))];
            tensor<fp16, [1, 512, 1, 1054]> var_5330_cast_fp16 = conv(dilations = var_5330_dilations_0, groups = var_5330_groups_0, pad = var_5330_pad_0, pad_type = var_5330_pad_type_0, strides = var_5330_strides_0, weight = core_stacks_4_layers_2_conv_module2_dw_c_weight_to_fp16, x = input_473_cast_fp16)[name = tensor<string, []>("op_5330_cast_fp16")];
            tensor<int32, [4]> var_5334_begin_0 = const()[name = tensor<string, []>("op_5334_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_5334_end_0 = const()[name = tensor<string, []>("op_5334_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_5334_end_mask_0 = const()[name = tensor<string, []>("op_5334_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5334_cast_fp16 = slice_by_index(begin = var_5334_begin_0, end = var_5334_end_0, end_mask = var_5334_end_mask_0, x = var_5330_cast_fp16)[name = tensor<string, []>("op_5334_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_669_cast_fp16 = add(x = var_5324_cast_fp16, y = var_5334_cast_fp16)[name = tensor<string, []>("x_669_cast_fp16")];
            tensor<fp16, []> var_5336_to_fp16 = const()[name = tensor<string, []>("op_5336_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_671_cast_fp16 = sub(x = x_669_cast_fp16, y = var_5336_to_fp16)[name = tensor<string, []>("x_671_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5338_cast_fp16 = relu(x = x_671_cast_fp16)[name = tensor<string, []>("op_5338_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5339_cast_fp16 = abs(x = x_671_cast_fp16)[name = tensor<string, []>("op_5339_cast_fp16")];
            tensor<fp16, []> const_119_promoted_to_fp16 = const()[name = tensor<string, []>("const_119_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_5340_cast_fp16 = mul(x = var_5339_cast_fp16, y = const_119_promoted_to_fp16)[name = tensor<string, []>("op_5340_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5341_cast_fp16 = exp(x = var_5340_cast_fp16)[name = tensor<string, []>("op_5341_cast_fp16")];
            tensor<fp32, []> var_5342_epsilon_0 = const()[name = tensor<string, []>("op_5342_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_5342_cast_fp16 = log(epsilon = var_5342_epsilon_0, x = var_5341_cast_fp16)[name = tensor<string, []>("op_5342_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5343_cast_fp16 = add(x = var_5338_cast_fp16, y = var_5342_cast_fp16)[name = tensor<string, []>("op_5343_cast_fp16")];
            tensor<fp16, []> var_5344_to_fp16 = const()[name = tensor<string, []>("op_5344_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_5345_cast_fp16 = mul(x = x_669_cast_fp16, y = var_5344_to_fp16)[name = tensor<string, []>("op_5345_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5346_cast_fp16 = sub(x = var_5343_cast_fp16, y = var_5345_cast_fp16)[name = tensor<string, []>("op_5346_cast_fp16")];
            tensor<fp16, []> var_5347_to_fp16 = const()[name = tensor<string, []>("op_5347_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_475_cast_fp16 = sub(x = var_5346_cast_fp16, y = var_5347_to_fp16)[name = tensor<string, []>("input_475_cast_fp16")];
            tensor<string, []> var_5355_pad_type_0 = const()[name = tensor<string, []>("op_5355_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5355_strides_0 = const()[name = tensor<string, []>("op_5355_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5355_pad_0 = const()[name = tensor<string, []>("op_5355_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5355_dilations_0 = const()[name = tensor<string, []>("op_5355_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5355_groups_0 = const()[name = tensor<string, []>("op_5355_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_4_layers_2_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307527936)))];
            tensor<fp16, [512]> core_stacks_4_layers_2_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308052288)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5355_cast_fp16 = conv(bias = core_stacks_4_layers_2_conv_module2_out_proj_bias_to_fp16, dilations = var_5355_dilations_0, groups = var_5355_groups_0, pad = var_5355_pad_0, pad_type = var_5355_pad_type_0, strides = var_5355_strides_0, weight = core_stacks_4_layers_2_conv_module2_out_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor<string, []>("op_5355_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_477_cast_fp16 = add(x = x_663_cast_fp16, y = var_5355_cast_fp16)[name = tensor<string, []>("input_477_cast_fp16")];
            tensor<string, []> x_673_pad_type_0 = const()[name = tensor<string, []>("x_673_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_673_strides_0 = const()[name = tensor<string, []>("x_673_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_673_pad_0 = const()[name = tensor<string, []>("x_673_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_673_dilations_0 = const()[name = tensor<string, []>("x_673_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_673_groups_0 = const()[name = tensor<string, []>("x_673_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_4_layers_2_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308053376)))];
            tensor<fp16, [1920]> core_stacks_4_layers_2_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310019520)))];
            tensor<fp16, [1, 1920, 1, 1024]> x_673_cast_fp16 = conv(bias = core_stacks_4_layers_2_feed_forward3_in_proj_bias_to_fp16, dilations = x_673_dilations_0, groups = x_673_groups_0, pad = x_673_pad_0, pad_type = x_673_pad_type_0, strides = x_673_strides_0, weight = core_stacks_4_layers_2_feed_forward3_in_proj_weight_to_fp16, x = input_477_cast_fp16)[name = tensor<string, []>("x_673_cast_fp16")];
            tensor<fp16, []> var_5366_to_fp16 = const()[name = tensor<string, []>("op_5366_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 1024]> x_675_cast_fp16 = sub(x = x_673_cast_fp16, y = var_5366_to_fp16)[name = tensor<string, []>("x_675_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5368_cast_fp16 = relu(x = x_675_cast_fp16)[name = tensor<string, []>("op_5368_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5369_cast_fp16 = abs(x = x_675_cast_fp16)[name = tensor<string, []>("op_5369_cast_fp16")];
            tensor<fp16, []> const_120_promoted_to_fp16 = const()[name = tensor<string, []>("const_120_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_5370_cast_fp16 = mul(x = var_5369_cast_fp16, y = const_120_promoted_to_fp16)[name = tensor<string, []>("op_5370_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5371_cast_fp16 = exp(x = var_5370_cast_fp16)[name = tensor<string, []>("op_5371_cast_fp16")];
            tensor<fp32, []> var_5372_epsilon_0 = const()[name = tensor<string, []>("op_5372_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_5372_cast_fp16 = log(epsilon = var_5372_epsilon_0, x = var_5371_cast_fp16)[name = tensor<string, []>("op_5372_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5373_cast_fp16 = add(x = var_5368_cast_fp16, y = var_5372_cast_fp16)[name = tensor<string, []>("op_5373_cast_fp16")];
            tensor<fp16, []> var_5374_to_fp16 = const()[name = tensor<string, []>("op_5374_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 1024]> var_5375_cast_fp16 = mul(x = x_673_cast_fp16, y = var_5374_to_fp16)[name = tensor<string, []>("op_5375_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5376_cast_fp16 = sub(x = var_5373_cast_fp16, y = var_5375_cast_fp16)[name = tensor<string, []>("op_5376_cast_fp16")];
            tensor<fp16, []> var_5377_to_fp16 = const()[name = tensor<string, []>("op_5377_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 1024]> input_479_cast_fp16 = sub(x = var_5376_cast_fp16, y = var_5377_to_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
            tensor<string, []> var_5385_pad_type_0 = const()[name = tensor<string, []>("op_5385_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5385_strides_0 = const()[name = tensor<string, []>("op_5385_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5385_pad_0 = const()[name = tensor<string, []>("op_5385_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5385_dilations_0 = const()[name = tensor<string, []>("op_5385_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5385_groups_0 = const()[name = tensor<string, []>("op_5385_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_4_layers_2_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310023424)))];
            tensor<fp16, [512]> core_stacks_4_layers_2_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311989568)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5385_cast_fp16 = conv(bias = core_stacks_4_layers_2_feed_forward3_out_proj_bias_to_fp16, dilations = var_5385_dilations_0, groups = var_5385_groups_0, pad = var_5385_pad_0, pad_type = var_5385_pad_type_0, strides = var_5385_strides_0, weight = core_stacks_4_layers_2_feed_forward3_out_proj_weight_to_fp16, x = input_479_cast_fp16)[name = tensor<string, []>("op_5385_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_677_cast_fp16 = add(x = input_477_cast_fp16, y = var_5385_cast_fp16)[name = tensor<string, []>("x_677_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_2_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311990656)))];
            tensor<fp16, [1, 512, 1, 1024]> d_29_cast_fp16 = sub(x = x_677_cast_fp16, y = core_stacks_4_layers_2_norm_bias_to_fp16)[name = tensor<string, []>("d_29_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5390_cast_fp16 = mul(x = d_29_cast_fp16, y = d_29_cast_fp16)[name = tensor<string, []>("op_5390_cast_fp16")];
            tensor<int32, [1]> var_5392_axes_0 = const()[name = tensor<string, []>("op_5392_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_5392_keep_dims_0 = const()[name = tensor<string, []>("op_5392_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 1024]> var_5392_cast_fp16 = reduce_mean(axes = var_5392_axes_0, keep_dims = var_5392_keep_dims_0, x = var_5390_cast_fp16)[name = tensor<string, []>("op_5392_cast_fp16")];
            tensor<fp32, []> scales_29_epsilon_0 = const()[name = tensor<string, []>("scales_29_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1024]> scales_29_cast_fp16 = rsqrt(epsilon = scales_29_epsilon_0, x = var_5392_cast_fp16)[name = tensor<string, []>("scales_29_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5394_cast_fp16 = mul(x = x_677_cast_fp16, y = scales_29_cast_fp16)[name = tensor<string, []>("op_5394_cast_fp16")];
            tensor<fp16, []> core_stacks_4_layers_2_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.d0cp+0)];
            tensor<fp16, [1, 512, 1, 1024]> src_65_cast_fp16 = mul(x = var_5394_cast_fp16, y = core_stacks_4_layers_2_norm_scale_to_fp16)[name = tensor<string, []>("src_65_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5397_cast_fp16 = sub(x = src_65_cast_fp16, y = input_451_cast_fp16)[name = tensor<string, []>("op_5397_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_2_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_2_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311991744)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5398_cast_fp16 = mul(x = var_5397_cast_fp16, y = core_stacks_4_layers_2_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_5398_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_481_cast_fp16 = add(x = input_451_cast_fp16, y = var_5398_cast_fp16)[name = tensor<string, []>("input_481_cast_fp16")];
            tensor<string, []> proj_pad_type_0 = const()[name = tensor<string, []>("proj_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> proj_strides_0 = const()[name = tensor<string, []>("proj_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> proj_pad_0 = const()[name = tensor<string, []>("proj_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> proj_dilations_0 = const()[name = tensor<string, []>("proj_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> proj_groups_0 = const()[name = tensor<string, []>("proj_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 512, 1, 1]> core_stacks_4_layers_3_self_attn_weights_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_self_attn_weights_in_proj_weight_to_fp16"), val = tensor<fp16, [384, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311992832)))];
            tensor<fp16, [384]> core_stacks_4_layers_3_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312386112)))];
            tensor<fp16, [1, 384, 1, 1024]> proj_cast_fp16 = conv(bias = core_stacks_4_layers_3_self_attn_weights_in_proj_bias_to_fp16, dilations = proj_dilations_0, groups = proj_groups_0, pad = proj_pad_0, pad_type = proj_pad_type_0, strides = proj_strides_0, weight = core_stacks_4_layers_3_self_attn_weights_in_proj_weight_to_fp16, x = input_481_cast_fp16)[name = tensor<string, []>("proj_cast_fp16")];
            tensor<int32, [4]> var_5421_begin_0 = const()[name = tensor<string, []>("op_5421_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5421_end_0 = const()[name = tensor<string, []>("op_5421_end_0"), val = tensor<int32, [4]>([1, 128, 1, 1024])];
            tensor<bool, [4]> var_5421_end_mask_0 = const()[name = tensor<string, []>("op_5421_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_5421_cast_fp16 = slice_by_index(begin = var_5421_begin_0, end = var_5421_end_0, end_mask = var_5421_end_mask_0, x = proj_cast_fp16)[name = tensor<string, []>("op_5421_cast_fp16")];
            tensor<int32, [4]> var_5422 = const()[name = tensor<string, []>("op_5422"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_5423_cast_fp16 = reshape(shape = var_5422, x = var_5421_cast_fp16)[name = tensor<string, []>("op_5423_cast_fp16")];
            tensor<int32, [4]> var_5425_begin_0 = const()[name = tensor<string, []>("op_5425_begin_0"), val = tensor<int32, [4]>([0, 128, 0, 0])];
            tensor<int32, [4]> var_5425_end_0 = const()[name = tensor<string, []>("op_5425_end_0"), val = tensor<int32, [4]>([1, 256, 1, 1024])];
            tensor<bool, [4]> var_5425_end_mask_0 = const()[name = tensor<string, []>("op_5425_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_5425_cast_fp16 = slice_by_index(begin = var_5425_begin_0, end = var_5425_end_0, end_mask = var_5425_end_mask_0, x = proj_cast_fp16)[name = tensor<string, []>("op_5425_cast_fp16")];
            tensor<int32, [4]> var_5426 = const()[name = tensor<string, []>("op_5426"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> var_5427_cast_fp16 = reshape(shape = var_5426, x = var_5425_cast_fp16)[name = tensor<string, []>("op_5427_cast_fp16")];
            tensor<int32, [4]> var_5429_begin_0 = const()[name = tensor<string, []>("op_5429_begin_0"), val = tensor<int32, [4]>([0, 256, 0, 0])];
            tensor<int32, [4]> var_5429_end_0 = const()[name = tensor<string, []>("op_5429_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_5429_end_mask_0 = const()[name = tensor<string, []>("op_5429_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 128, 1, 1024]> var_5429_cast_fp16 = slice_by_index(begin = var_5429_begin_0, end = var_5429_end_0, end_mask = var_5429_end_mask_0, x = proj_cast_fp16)[name = tensor<string, []>("op_5429_cast_fp16")];
            tensor<int32, [4]> var_5430 = const()[name = tensor<string, []>("op_5430"), val = tensor<int32, [4]>([4, 32, 1, 1024])];
            tensor<fp16, [4, 32, 1, 1024]> pt_cast_fp16 = reshape(shape = var_5430, x = var_5429_cast_fp16)[name = tensor<string, []>("pt_cast_fp16")];
            tensor<int32, [4]> transpose_120_perm_0 = const()[name = tensor<string, []>("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_604 = const()[name = tensor<string, []>("concat_604"), val = tensor<int32, [3]>([4, 1024, 32])];
            tensor<fp16, [4, 1, 1024, 32]> transpose_120_cast_fp16 = transpose(perm = transpose_120_perm_0, x = var_5423_cast_fp16)[name = tensor<string, []>("transpose_140")];
            tensor<fp16, [4, 1024, 32]> reshape_180_cast_fp16 = reshape(shape = concat_604, x = transpose_120_cast_fp16)[name = tensor<string, []>("reshape_180_cast_fp16")];
            tensor<int32, [4]> transpose_121_perm_0 = const()[name = tensor<string, []>("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_605 = const()[name = tensor<string, []>("concat_605"), val = tensor<int32, [3]>([4, 32, 1024])];
            tensor<fp16, [4, 1, 32, 1024]> transpose_121_cast_fp16 = transpose(perm = transpose_121_perm_0, x = var_5427_cast_fp16)[name = tensor<string, []>("transpose_139")];
            tensor<fp16, [4, 32, 1024]> reshape_181_cast_fp16 = reshape(shape = concat_605, x = transpose_121_cast_fp16)[name = tensor<string, []>("reshape_181_cast_fp16")];
            tensor<bool, []> matmul_60_transpose_x_0 = const()[name = tensor<string, []>("matmul_60_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_60_transpose_y_0 = const()[name = tensor<string, []>("matmul_60_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 1024]> matmul_60_cast_fp16 = matmul(transpose_x = matmul_60_transpose_x_0, transpose_y = matmul_60_transpose_y_0, x = reshape_180_cast_fp16, y = reshape_181_cast_fp16)[name = tensor<string, []>("matmul_60_cast_fp16")];
            tensor<int32, [4]> concat_609 = const()[name = tensor<string, []>("concat_609"), val = tensor<int32, [4]>([4, 1, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> reshape_182_cast_fp16 = reshape(shape = concat_609, x = matmul_60_cast_fp16)[name = tensor<string, []>("reshape_182_cast_fp16")];
            tensor<int32, [4]> attn_scores_91_perm_0 = const()[name = tensor<string, []>("attn_scores_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> var_5434 = const()[name = tensor<string, []>("op_5434"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<fp16, [4, 32, 1024, 1]> var_5435_cast_fp16 = transpose(perm = var_5434, x = pt_cast_fp16)[name = tensor<string, []>("transpose_137")];
            tensor<fp16, [4, 32, 1024, 1024]> var_5436_cast_fp16 = mul(x = var_5435_cast_fp16, y = core_pos_basis_1024_to_fp16)[name = tensor<string, []>("op_5436_cast_fp16")];
            tensor<int32, [1]> pos_scores_axes_0 = const()[name = tensor<string, []>("pos_scores_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> pos_scores_keep_dims_0 = const()[name = tensor<string, []>("pos_scores_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 1024, 1024]> pos_scores_cast_fp16 = reduce_sum(axes = pos_scores_axes_0, keep_dims = pos_scores_keep_dims_0, x = var_5436_cast_fp16)[name = tensor<string, []>("pos_scores_cast_fp16")];
            tensor<int32, [4]> var_5439 = const()[name = tensor<string, []>("op_5439"), val = tensor<int32, [4]>([4, 1024, 1, 1024])];
            tensor<fp16, [4, 1024, 1, 1024]> var_5440_cast_fp16 = reshape(shape = var_5439, x = pos_scores_cast_fp16)[name = tensor<string, []>("op_5440_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_91_cast_fp16 = transpose(perm = attn_scores_91_perm_0, x = reshape_182_cast_fp16)[name = tensor<string, []>("transpose_138")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_93_cast_fp16 = add(x = attn_scores_91_cast_fp16, y = var_5440_cast_fp16)[name = tensor<string, []>("attn_scores_93_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_scores_cast_fp16 = add(x = attn_scores_93_cast_fp16, y = attn_bias_1_cast_fp16)[name = tensor<string, []>("attn_scores_cast_fp16")];
            tensor<fp16, [4, 1024, 1, 1024]> attn_weights_cast_fp16 = softmax(axis = var_41, x = attn_scores_cast_fp16)[name = tensor<string, []>("attn_weights_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_483_cast_fp16 = add(x = input_481_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("input_483_cast_fp16")];
            tensor<string, []> x_679_pad_type_0 = const()[name = tensor<string, []>("x_679_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_679_strides_0 = const()[name = tensor<string, []>("x_679_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_679_pad_0 = const()[name = tensor<string, []>("x_679_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_679_dilations_0 = const()[name = tensor<string, []>("x_679_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_679_groups_0 = const()[name = tensor<string, []>("x_679_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_4_layers_3_feed_forward1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward1_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312386944)))];
            tensor<fp16, [1152]> core_stacks_4_layers_3_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313566656)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_679_cast_fp16 = conv(bias = core_stacks_4_layers_3_feed_forward1_in_proj_bias_to_fp16, dilations = x_679_dilations_0, groups = x_679_groups_0, pad = x_679_pad_0, pad_type = x_679_pad_type_0, strides = x_679_strides_0, weight = core_stacks_4_layers_3_feed_forward1_in_proj_weight_to_fp16, x = input_483_cast_fp16)[name = tensor<string, []>("x_679_cast_fp16")];
            tensor<fp16, []> var_5454_to_fp16 = const()[name = tensor<string, []>("op_5454_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1152, 1, 1024]> x_681_cast_fp16 = sub(x = x_679_cast_fp16, y = var_5454_to_fp16)[name = tensor<string, []>("x_681_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_5456_cast_fp16 = relu(x = x_681_cast_fp16)[name = tensor<string, []>("op_5456_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_5457_cast_fp16 = abs(x = x_681_cast_fp16)[name = tensor<string, []>("op_5457_cast_fp16")];
            tensor<fp16, []> const_121_promoted_to_fp16 = const()[name = tensor<string, []>("const_121_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_5458_cast_fp16 = mul(x = var_5457_cast_fp16, y = const_121_promoted_to_fp16)[name = tensor<string, []>("op_5458_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_5459_cast_fp16 = exp(x = var_5458_cast_fp16)[name = tensor<string, []>("op_5459_cast_fp16")];
            tensor<fp32, []> var_5460_epsilon_0 = const()[name = tensor<string, []>("op_5460_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1152, 1, 1024]> var_5460_cast_fp16 = log(epsilon = var_5460_epsilon_0, x = var_5459_cast_fp16)[name = tensor<string, []>("op_5460_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_5461_cast_fp16 = add(x = var_5456_cast_fp16, y = var_5460_cast_fp16)[name = tensor<string, []>("op_5461_cast_fp16")];
            tensor<fp16, []> var_5462_to_fp16 = const()[name = tensor<string, []>("op_5462_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1152, 1, 1024]> var_5463_cast_fp16 = mul(x = x_679_cast_fp16, y = var_5462_to_fp16)[name = tensor<string, []>("op_5463_cast_fp16")];
            tensor<fp16, [1, 1152, 1, 1024]> var_5464_cast_fp16 = sub(x = var_5461_cast_fp16, y = var_5463_cast_fp16)[name = tensor<string, []>("op_5464_cast_fp16")];
            tensor<fp16, []> var_5465_to_fp16 = const()[name = tensor<string, []>("op_5465_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1152, 1, 1024]> input_485_cast_fp16 = sub(x = var_5464_cast_fp16, y = var_5465_to_fp16)[name = tensor<string, []>("input_485_cast_fp16")];
            tensor<string, []> var_5473_pad_type_0 = const()[name = tensor<string, []>("op_5473_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5473_strides_0 = const()[name = tensor<string, []>("op_5473_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5473_pad_0 = const()[name = tensor<string, []>("op_5473_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5473_dilations_0 = const()[name = tensor<string, []>("op_5473_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5473_groups_0 = const()[name = tensor<string, []>("op_5473_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1152, 1, 1]> core_stacks_4_layers_3_feed_forward1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1152, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313569024)))];
            tensor<fp16, [512]> core_stacks_4_layers_3_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314748736)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5473_cast_fp16 = conv(bias = core_stacks_4_layers_3_feed_forward1_out_proj_bias_to_fp16, dilations = var_5473_dilations_0, groups = var_5473_groups_0, pad = var_5473_pad_0, pad_type = var_5473_pad_type_0, strides = var_5473_strides_0, weight = core_stacks_4_layers_3_feed_forward1_out_proj_weight_to_fp16, x = input_485_cast_fp16)[name = tensor<string, []>("op_5473_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_487_cast_fp16 = add(x = input_483_cast_fp16, y = var_5473_cast_fp16)[name = tensor<string, []>("input_487_cast_fp16")];
            tensor<int32, [4]> var_5475_begin_0 = const()[name = tensor<string, []>("op_5475_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5475_end_0 = const()[name = tensor<string, []>("op_5475_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_5475_end_mask_0 = const()[name = tensor<string, []>("op_5475_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> var_5475_cast_fp16 = slice_by_index(begin = var_5475_begin_0, end = var_5475_end_0, end_mask = var_5475_end_mask_0, x = attn_weights_cast_fp16)[name = tensor<string, []>("op_5475_cast_fp16")];
            tensor<string, []> x_683_pad_type_0 = const()[name = tensor<string, []>("x_683_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_683_strides_0 = const()[name = tensor<string, []>("x_683_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_683_pad_0 = const()[name = tensor<string, []>("x_683_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_683_dilations_0 = const()[name = tensor<string, []>("x_683_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_683_groups_0 = const()[name = tensor<string, []>("x_683_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1152, 512, 1, 1]> core_stacks_4_layers_3_nonlin_attention_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_nonlin_attention_in_proj_weight_to_fp16"), val = tensor<fp16, [1152, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314749824)))];
            tensor<fp16, [1152]> core_stacks_4_layers_3_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315929536)))];
            tensor<fp16, [1, 1152, 1, 1024]> x_683_cast_fp16 = conv(bias = core_stacks_4_layers_3_nonlin_attention_in_proj_bias_to_fp16, dilations = x_683_dilations_0, groups = x_683_groups_0, pad = x_683_pad_0, pad_type = x_683_pad_type_0, strides = x_683_strides_0, weight = core_stacks_4_layers_3_nonlin_attention_in_proj_weight_to_fp16, x = input_487_cast_fp16)[name = tensor<string, []>("x_683_cast_fp16")];
            tensor<int32, [4]> var_5486_begin_0 = const()[name = tensor<string, []>("op_5486_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5486_end_0 = const()[name = tensor<string, []>("op_5486_end_0"), val = tensor<int32, [4]>([1, 384, 1, 1024])];
            tensor<bool, [4]> var_5486_end_mask_0 = const()[name = tensor<string, []>("op_5486_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_5486_cast_fp16 = slice_by_index(begin = var_5486_begin_0, end = var_5486_end_0, end_mask = var_5486_end_mask_0, x = x_683_cast_fp16)[name = tensor<string, []>("op_5486_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> s_gate_cast_fp16 = tanh(x = var_5486_cast_fp16)[name = tensor<string, []>("s_gate_cast_fp16")];
            tensor<int32, [4]> var_5489_begin_0 = const()[name = tensor<string, []>("op_5489_begin_0"), val = tensor<int32, [4]>([0, 384, 0, 0])];
            tensor<int32, [4]> var_5489_end_0 = const()[name = tensor<string, []>("op_5489_end_0"), val = tensor<int32, [4]>([1, 768, 1, 1024])];
            tensor<bool, [4]> var_5489_end_mask_0 = const()[name = tensor<string, []>("op_5489_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> var_5489_cast_fp16 = slice_by_index(begin = var_5489_begin_0, end = var_5489_end_0, end_mask = var_5489_end_mask_0, x = x_683_cast_fp16)[name = tensor<string, []>("op_5489_cast_fp16")];
            tensor<fp16, [1, 384, 1, 1024]> var_5490_cast_fp16 = mul(x = var_5489_cast_fp16, y = s_gate_cast_fp16)[name = tensor<string, []>("op_5490_cast_fp16")];
            tensor<int32, [4]> y_begin_0 = const()[name = tensor<string, []>("y_begin_0"), val = tensor<int32, [4]>([0, 768, 0, 0])];
            tensor<int32, [4]> y_end_0 = const()[name = tensor<string, []>("y_end_0"), val = tensor<int32, [4]>([1, 1152, 1, 1024])];
            tensor<bool, [4]> y_end_mask_0 = const()[name = tensor<string, []>("y_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 384, 1, 1024]> y_cast_fp16 = slice_by_index(begin = y_begin_0, end = y_end_0, end_mask = y_end_mask_0, x = x_683_cast_fp16)[name = tensor<string, []>("y_cast_fp16")];
            tensor<int32, [4]> transpose_122_perm_0 = const()[name = tensor<string, []>("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_614 = const()[name = tensor<string, []>("concat_614"), val = tensor<int32, [3]>([1, 1024, 1024])];
            tensor<fp16, [1, 1, 1024, 1024]> transpose_122_cast_fp16 = transpose(perm = transpose_122_perm_0, x = var_5475_cast_fp16)[name = tensor<string, []>("transpose_136")];
            tensor<fp16, [1, 1024, 1024]> reshape_183_cast_fp16 = reshape(shape = concat_614, x = transpose_122_cast_fp16)[name = tensor<string, []>("reshape_183_cast_fp16")];
            tensor<int32, [4]> transpose_123_perm_0 = const()[name = tensor<string, []>("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_615 = const()[name = tensor<string, []>("concat_615"), val = tensor<int32, [3]>([1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> transpose_123_cast_fp16 = transpose(perm = transpose_123_perm_0, x = var_5490_cast_fp16)[name = tensor<string, []>("transpose_135")];
            tensor<fp16, [1, 1024, 384]> reshape_184_cast_fp16 = reshape(shape = concat_615, x = transpose_123_cast_fp16)[name = tensor<string, []>("reshape_184_cast_fp16")];
            tensor<bool, []> matmul_61_transpose_x_0 = const()[name = tensor<string, []>("matmul_61_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_61_transpose_y_0 = const()[name = tensor<string, []>("matmul_61_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1024, 384]> matmul_61_cast_fp16 = matmul(transpose_x = matmul_61_transpose_x_0, transpose_y = matmul_61_transpose_y_0, x = reshape_183_cast_fp16, y = reshape_184_cast_fp16)[name = tensor<string, []>("matmul_61_cast_fp16")];
            tensor<int32, [4]> concat_619 = const()[name = tensor<string, []>("concat_619"), val = tensor<int32, [4]>([1, 1, 1024, 384])];
            tensor<fp16, [1, 1, 1024, 384]> reshape_185_cast_fp16 = reshape(shape = concat_619, x = matmul_61_cast_fp16)[name = tensor<string, []>("reshape_185_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = tensor<string, []>("v_31_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<fp16, [1, 384, 1, 1024]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = reshape_185_cast_fp16)[name = tensor<string, []>("transpose_134")];
            tensor<fp16, [1, 384, 1, 1024]> input_489_cast_fp16 = mul(x = v_31_cast_fp16, y = y_cast_fp16)[name = tensor<string, []>("input_489_cast_fp16")];
            tensor<string, []> var_5502_pad_type_0 = const()[name = tensor<string, []>("op_5502_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5502_strides_0 = const()[name = tensor<string, []>("op_5502_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5502_pad_0 = const()[name = tensor<string, []>("op_5502_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5502_dilations_0 = const()[name = tensor<string, []>("op_5502_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5502_groups_0 = const()[name = tensor<string, []>("op_5502_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 384, 1, 1]> core_stacks_4_layers_3_nonlin_attention_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_nonlin_attention_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315931904)))];
            tensor<fp16, [512]> core_stacks_4_layers_3_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316325184)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5502_cast_fp16 = conv(bias = core_stacks_4_layers_3_nonlin_attention_out_proj_bias_to_fp16, dilations = var_5502_dilations_0, groups = var_5502_groups_0, pad = var_5502_pad_0, pad_type = var_5502_pad_type_0, strides = var_5502_strides_0, weight = core_stacks_4_layers_3_nonlin_attention_out_proj_weight_to_fp16, x = input_489_cast_fp16)[name = tensor<string, []>("op_5502_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_685_cast_fp16 = add(x = input_487_cast_fp16, y = var_5502_cast_fp16)[name = tensor<string, []>("x_685_cast_fp16")];
            tensor<string, []> var_5513_pad_type_0 = const()[name = tensor<string, []>("op_5513_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5513_strides_0 = const()[name = tensor<string, []>("op_5513_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5513_pad_0 = const()[name = tensor<string, []>("op_5513_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5513_dilations_0 = const()[name = tensor<string, []>("op_5513_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5513_groups_0 = const()[name = tensor<string, []>("op_5513_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_4_layers_3_self_attn1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_self_attn1_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316326272)))];
            tensor<fp16, [48]> core_stacks_4_layers_3_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316375488)))];
            tensor<fp16, [1, 48, 1, 1024]> var_5513_cast_fp16 = conv(bias = core_stacks_4_layers_3_self_attn1_in_proj_bias_to_fp16, dilations = var_5513_dilations_0, groups = var_5513_groups_0, pad = var_5513_pad_0, pad_type = var_5513_pad_type_0, strides = var_5513_strides_0, weight = core_stacks_4_layers_3_self_attn1_in_proj_weight_to_fp16, x = x_685_cast_fp16)[name = tensor<string, []>("op_5513_cast_fp16")];
            tensor<int32, [4]> var_5514 = const()[name = tensor<string, []>("op_5514"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_5515_cast_fp16 = reshape(shape = var_5514, x = var_5513_cast_fp16)[name = tensor<string, []>("op_5515_cast_fp16")];
            tensor<int32, [4]> transpose_124_perm_0 = const()[name = tensor<string, []>("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> concat_624 = const()[name = tensor<string, []>("concat_624"), val = tensor<int32, [3]>([4, 1024, 1024])];
            tensor<fp16, [4, 1, 1024, 1024]> transpose_124_cast_fp16 = transpose(perm = transpose_124_perm_0, x = attn_weights_cast_fp16)[name = tensor<string, []>("transpose_133")];
            tensor<fp16, [4, 1024, 1024]> reshape_186_cast_fp16 = reshape(shape = concat_624, x = transpose_124_cast_fp16)[name = tensor<string, []>("reshape_186_cast_fp16")];
            tensor<int32, [4]> transpose_125_perm_0 = const()[name = tensor<string, []>("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_625 = const()[name = tensor<string, []>("concat_625"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_125_cast_fp16 = transpose(perm = transpose_125_perm_0, x = var_5515_cast_fp16)[name = tensor<string, []>("transpose_132")];
            tensor<fp16, [4, 1024, 12]> reshape_187_cast_fp16 = reshape(shape = concat_625, x = transpose_125_cast_fp16)[name = tensor<string, []>("reshape_187_cast_fp16")];
            tensor<bool, []> matmul_62_transpose_x_0 = const()[name = tensor<string, []>("matmul_62_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_62_transpose_y_0 = const()[name = tensor<string, []>("matmul_62_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_62_cast_fp16 = matmul(transpose_x = matmul_62_transpose_x_0, transpose_y = matmul_62_transpose_y_0, x = reshape_186_cast_fp16, y = reshape_187_cast_fp16)[name = tensor<string, []>("matmul_62_cast_fp16")];
            tensor<int32, [4]> concat_629 = const()[name = tensor<string, []>("concat_629"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_188_cast_fp16 = reshape(shape = concat_629, x = matmul_62_cast_fp16)[name = tensor<string, []>("reshape_188_cast_fp16")];
            tensor<int32, [4]> out_61_perm_0 = const()[name = tensor<string, []>("out_61_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_5518 = const()[name = tensor<string, []>("op_5518"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_61_cast_fp16 = transpose(perm = out_61_perm_0, x = reshape_188_cast_fp16)[name = tensor<string, []>("transpose_131")];
            tensor<fp16, [1, 48, 1, 1024]> input_491_cast_fp16 = reshape(shape = var_5518, x = out_61_cast_fp16)[name = tensor<string, []>("input_491_cast_fp16")];
            tensor<string, []> var_5526_pad_type_0 = const()[name = tensor<string, []>("op_5526_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5526_strides_0 = const()[name = tensor<string, []>("op_5526_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5526_pad_0 = const()[name = tensor<string, []>("op_5526_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5526_dilations_0 = const()[name = tensor<string, []>("op_5526_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5526_groups_0 = const()[name = tensor<string, []>("op_5526_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_4_layers_3_self_attn1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_self_attn1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316375680)))];
            tensor<fp16, [512]> core_stacks_4_layers_3_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316424896)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5526_cast_fp16 = conv(bias = core_stacks_4_layers_3_self_attn1_out_proj_bias_to_fp16, dilations = var_5526_dilations_0, groups = var_5526_groups_0, pad = var_5526_pad_0, pad_type = var_5526_pad_type_0, strides = var_5526_strides_0, weight = core_stacks_4_layers_3_self_attn1_out_proj_weight_to_fp16, x = input_491_cast_fp16)[name = tensor<string, []>("op_5526_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_687_cast_fp16 = add(x = x_685_cast_fp16, y = var_5526_cast_fp16)[name = tensor<string, []>("x_687_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_689_cast_fp16 = add(x = x_687_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("x_689_cast_fp16")];
            tensor<string, []> x_691_pad_type_0 = const()[name = tensor<string, []>("x_691_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_691_strides_0 = const()[name = tensor<string, []>("x_691_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_691_pad_0 = const()[name = tensor<string, []>("x_691_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_691_dilations_0 = const()[name = tensor<string, []>("x_691_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_691_groups_0 = const()[name = tensor<string, []>("x_691_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_4_layers_3_conv_module1_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module1_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316425984)))];
            tensor<fp16, [1024]> core_stacks_4_layers_3_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317474624)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_691_cast_fp16 = conv(bias = core_stacks_4_layers_3_conv_module1_in_proj_bias_to_fp16, dilations = x_691_dilations_0, groups = x_691_groups_0, pad = x_691_pad_0, pad_type = x_691_pad_type_0, strides = x_691_strides_0, weight = core_stacks_4_layers_3_conv_module1_in_proj_weight_to_fp16, x = x_689_cast_fp16)[name = tensor<string, []>("x_691_cast_fp16")];
            tensor<int32, [4]> var_5544_begin_0 = const()[name = tensor<string, []>("op_5544_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5544_end_0 = const()[name = tensor<string, []>("op_5544_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_5544_end_mask_0 = const()[name = tensor<string, []>("op_5544_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5544_cast_fp16 = slice_by_index(begin = var_5544_begin_0, end = var_5544_end_0, end_mask = var_5544_end_mask_0, x = x_691_cast_fp16)[name = tensor<string, []>("op_5544_cast_fp16")];
            tensor<int32, [4]> var_5546_begin_0 = const()[name = tensor<string, []>("op_5546_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_5546_end_0 = const()[name = tensor<string, []>("op_5546_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_5546_end_mask_0 = const()[name = tensor<string, []>("op_5546_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5546_cast_fp16 = slice_by_index(begin = var_5546_begin_0, end = var_5546_end_0, end_mask = var_5546_end_mask_0, x = x_691_cast_fp16)[name = tensor<string, []>("op_5546_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5547_cast_fp16 = sigmoid(x = var_5546_cast_fp16)[name = tensor<string, []>("op_5547_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_693_cast_fp16 = mul(x = var_5544_cast_fp16, y = var_5547_cast_fp16)[name = tensor<string, []>("x_693_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_493_cast_fp16 = mul(x = x_693_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_493_cast_fp16")];
            tensor<string, []> var_5556_pad_type_0 = const()[name = tensor<string, []>("op_5556_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5556_pad_0 = const()[name = tensor<string, []>("op_5556_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5556_groups_0 = const()[name = tensor<string, []>("op_5556_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5556_strides_0 = const()[name = tensor<string, []>("op_5556_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5556_dilations_0 = const()[name = tensor<string, []>("op_5556_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_3_conv_module1_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module1_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317476736)))];
            tensor<fp16, [512]> core_stacks_4_layers_3_conv_module1_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module1_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317492160)))];
            tensor<fp16, [1, 512, 1, 1040]> var_5556_cast_fp16 = conv(bias = core_stacks_4_layers_3_conv_module1_dw_a_bias_to_fp16, dilations = var_5556_dilations_0, groups = var_5556_groups_0, pad = var_5556_pad_0, pad_type = var_5556_pad_type_0, strides = var_5556_strides_0, weight = core_stacks_4_layers_3_conv_module1_dw_a_weight_to_fp16, x = input_493_cast_fp16)[name = tensor<string, []>("op_5556_cast_fp16")];
            tensor<int32, [4]> var_5557_begin_0 = const()[name = tensor<string, []>("op_5557_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5557_end_0 = const()[name = tensor<string, []>("op_5557_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_5557_end_mask_0 = const()[name = tensor<string, []>("op_5557_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_5557_cast_fp16 = slice_by_index(begin = var_5557_begin_0, end = var_5557_end_0, end_mask = var_5557_end_mask_0, x = var_5556_cast_fp16)[name = tensor<string, []>("op_5557_cast_fp16")];
            tensor<string, []> var_5563_pad_type_0 = const()[name = tensor<string, []>("op_5563_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5563_pad_0 = const()[name = tensor<string, []>("op_5563_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5563_groups_0 = const()[name = tensor<string, []>("op_5563_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5563_strides_0 = const()[name = tensor<string, []>("op_5563_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5563_dilations_0 = const()[name = tensor<string, []>("op_5563_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_3_conv_module1_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module1_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317493248)))];
            tensor<fp16, [1, 512, 1, 1040]> var_5563_cast_fp16 = conv(dilations = var_5563_dilations_0, groups = var_5563_groups_0, pad = var_5563_pad_0, pad_type = var_5563_pad_type_0, strides = var_5563_strides_0, weight = core_stacks_4_layers_3_conv_module1_dw_b_weight_to_fp16, x = input_493_cast_fp16)[name = tensor<string, []>("op_5563_cast_fp16")];
            tensor<int32, [4]> var_5567_begin_0 = const()[name = tensor<string, []>("op_5567_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_5567_end_0 = const()[name = tensor<string, []>("op_5567_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_5567_end_mask_0 = const()[name = tensor<string, []>("op_5567_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_5567_cast_fp16 = slice_by_index(begin = var_5567_begin_0, end = var_5567_end_0, end_mask = var_5567_end_mask_0, x = var_5563_cast_fp16)[name = tensor<string, []>("op_5567_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5568_cast_fp16 = add(x = var_5557_cast_fp16, y = var_5567_cast_fp16)[name = tensor<string, []>("op_5568_cast_fp16")];
            tensor<string, []> var_5574_pad_type_0 = const()[name = tensor<string, []>("op_5574_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5574_pad_0 = const()[name = tensor<string, []>("op_5574_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5574_groups_0 = const()[name = tensor<string, []>("op_5574_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5574_strides_0 = const()[name = tensor<string, []>("op_5574_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5574_dilations_0 = const()[name = tensor<string, []>("op_5574_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_4_layers_3_conv_module1_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module1_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317508672)))];
            tensor<fp16, [1, 512, 1, 1054]> var_5574_cast_fp16 = conv(dilations = var_5574_dilations_0, groups = var_5574_groups_0, pad = var_5574_pad_0, pad_type = var_5574_pad_type_0, strides = var_5574_strides_0, weight = core_stacks_4_layers_3_conv_module1_dw_c_weight_to_fp16, x = input_493_cast_fp16)[name = tensor<string, []>("op_5574_cast_fp16")];
            tensor<int32, [4]> var_5578_begin_0 = const()[name = tensor<string, []>("op_5578_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_5578_end_0 = const()[name = tensor<string, []>("op_5578_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_5578_end_mask_0 = const()[name = tensor<string, []>("op_5578_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5578_cast_fp16 = slice_by_index(begin = var_5578_begin_0, end = var_5578_end_0, end_mask = var_5578_end_mask_0, x = var_5574_cast_fp16)[name = tensor<string, []>("op_5578_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_695_cast_fp16 = add(x = var_5568_cast_fp16, y = var_5578_cast_fp16)[name = tensor<string, []>("x_695_cast_fp16")];
            tensor<fp16, []> var_5580_to_fp16 = const()[name = tensor<string, []>("op_5580_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_697_cast_fp16 = sub(x = x_695_cast_fp16, y = var_5580_to_fp16)[name = tensor<string, []>("x_697_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5582_cast_fp16 = relu(x = x_697_cast_fp16)[name = tensor<string, []>("op_5582_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5583_cast_fp16 = abs(x = x_697_cast_fp16)[name = tensor<string, []>("op_5583_cast_fp16")];
            tensor<fp16, []> const_124_promoted_to_fp16 = const()[name = tensor<string, []>("const_124_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_5584_cast_fp16 = mul(x = var_5583_cast_fp16, y = const_124_promoted_to_fp16)[name = tensor<string, []>("op_5584_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5585_cast_fp16 = exp(x = var_5584_cast_fp16)[name = tensor<string, []>("op_5585_cast_fp16")];
            tensor<fp32, []> var_5586_epsilon_0 = const()[name = tensor<string, []>("op_5586_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_5586_cast_fp16 = log(epsilon = var_5586_epsilon_0, x = var_5585_cast_fp16)[name = tensor<string, []>("op_5586_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5587_cast_fp16 = add(x = var_5582_cast_fp16, y = var_5586_cast_fp16)[name = tensor<string, []>("op_5587_cast_fp16")];
            tensor<fp16, []> var_5588_to_fp16 = const()[name = tensor<string, []>("op_5588_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_5589_cast_fp16 = mul(x = x_695_cast_fp16, y = var_5588_to_fp16)[name = tensor<string, []>("op_5589_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5590_cast_fp16 = sub(x = var_5587_cast_fp16, y = var_5589_cast_fp16)[name = tensor<string, []>("op_5590_cast_fp16")];
            tensor<fp16, []> var_5591_to_fp16 = const()[name = tensor<string, []>("op_5591_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_495_cast_fp16 = sub(x = var_5590_cast_fp16, y = var_5591_to_fp16)[name = tensor<string, []>("input_495_cast_fp16")];
            tensor<string, []> var_5599_pad_type_0 = const()[name = tensor<string, []>("op_5599_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5599_strides_0 = const()[name = tensor<string, []>("op_5599_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5599_pad_0 = const()[name = tensor<string, []>("op_5599_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5599_dilations_0 = const()[name = tensor<string, []>("op_5599_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5599_groups_0 = const()[name = tensor<string, []>("op_5599_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_4_layers_3_conv_module1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317509760)))];
            tensor<fp16, [512]> core_stacks_4_layers_3_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318034112)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5599_cast_fp16 = conv(bias = core_stacks_4_layers_3_conv_module1_out_proj_bias_to_fp16, dilations = var_5599_dilations_0, groups = var_5599_groups_0, pad = var_5599_pad_0, pad_type = var_5599_pad_type_0, strides = var_5599_strides_0, weight = core_stacks_4_layers_3_conv_module1_out_proj_weight_to_fp16, x = input_495_cast_fp16)[name = tensor<string, []>("op_5599_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_497_cast_fp16 = add(x = x_689_cast_fp16, y = var_5599_cast_fp16)[name = tensor<string, []>("input_497_cast_fp16")];
            tensor<string, []> x_699_pad_type_0 = const()[name = tensor<string, []>("x_699_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_699_strides_0 = const()[name = tensor<string, []>("x_699_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_699_pad_0 = const()[name = tensor<string, []>("x_699_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_699_dilations_0 = const()[name = tensor<string, []>("x_699_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_699_groups_0 = const()[name = tensor<string, []>("x_699_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1536, 512, 1, 1]> core_stacks_4_layers_3_feed_forward2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward2_in_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318035200)))];
            tensor<fp16, [1536]> core_stacks_4_layers_3_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319608128)))];
            tensor<fp16, [1, 1536, 1, 1024]> x_699_cast_fp16 = conv(bias = core_stacks_4_layers_3_feed_forward2_in_proj_bias_to_fp16, dilations = x_699_dilations_0, groups = x_699_groups_0, pad = x_699_pad_0, pad_type = x_699_pad_type_0, strides = x_699_strides_0, weight = core_stacks_4_layers_3_feed_forward2_in_proj_weight_to_fp16, x = input_497_cast_fp16)[name = tensor<string, []>("x_699_cast_fp16")];
            tensor<fp16, []> var_5610_to_fp16 = const()[name = tensor<string, []>("op_5610_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1536, 1, 1024]> x_701_cast_fp16 = sub(x = x_699_cast_fp16, y = var_5610_to_fp16)[name = tensor<string, []>("x_701_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_5612_cast_fp16 = relu(x = x_701_cast_fp16)[name = tensor<string, []>("op_5612_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_5613_cast_fp16 = abs(x = x_701_cast_fp16)[name = tensor<string, []>("op_5613_cast_fp16")];
            tensor<fp16, []> const_125_promoted_to_fp16 = const()[name = tensor<string, []>("const_125_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_5614_cast_fp16 = mul(x = var_5613_cast_fp16, y = const_125_promoted_to_fp16)[name = tensor<string, []>("op_5614_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_5615_cast_fp16 = exp(x = var_5614_cast_fp16)[name = tensor<string, []>("op_5615_cast_fp16")];
            tensor<fp32, []> var_5616_epsilon_0 = const()[name = tensor<string, []>("op_5616_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1536, 1, 1024]> var_5616_cast_fp16 = log(epsilon = var_5616_epsilon_0, x = var_5615_cast_fp16)[name = tensor<string, []>("op_5616_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_5617_cast_fp16 = add(x = var_5612_cast_fp16, y = var_5616_cast_fp16)[name = tensor<string, []>("op_5617_cast_fp16")];
            tensor<fp16, []> var_5618_to_fp16 = const()[name = tensor<string, []>("op_5618_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1536, 1, 1024]> var_5619_cast_fp16 = mul(x = x_699_cast_fp16, y = var_5618_to_fp16)[name = tensor<string, []>("op_5619_cast_fp16")];
            tensor<fp16, [1, 1536, 1, 1024]> var_5620_cast_fp16 = sub(x = var_5617_cast_fp16, y = var_5619_cast_fp16)[name = tensor<string, []>("op_5620_cast_fp16")];
            tensor<fp16, []> var_5621_to_fp16 = const()[name = tensor<string, []>("op_5621_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1536, 1, 1024]> input_499_cast_fp16 = sub(x = var_5620_cast_fp16, y = var_5621_to_fp16)[name = tensor<string, []>("input_499_cast_fp16")];
            tensor<string, []> var_5629_pad_type_0 = const()[name = tensor<string, []>("op_5629_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5629_strides_0 = const()[name = tensor<string, []>("op_5629_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5629_pad_0 = const()[name = tensor<string, []>("op_5629_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5629_dilations_0 = const()[name = tensor<string, []>("op_5629_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5629_groups_0 = const()[name = tensor<string, []>("op_5629_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1536, 1, 1]> core_stacks_4_layers_3_feed_forward2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319611264)))];
            tensor<fp16, [512]> core_stacks_4_layers_3_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321184192)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5629_cast_fp16 = conv(bias = core_stacks_4_layers_3_feed_forward2_out_proj_bias_to_fp16, dilations = var_5629_dilations_0, groups = var_5629_groups_0, pad = var_5629_pad_0, pad_type = var_5629_pad_type_0, strides = var_5629_strides_0, weight = core_stacks_4_layers_3_feed_forward2_out_proj_weight_to_fp16, x = input_499_cast_fp16)[name = tensor<string, []>("op_5629_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> src_67_cast_fp16 = add(x = input_497_cast_fp16, y = var_5629_cast_fp16)[name = tensor<string, []>("src_67_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5632_cast_fp16 = sub(x = src_67_cast_fp16, y = input_481_cast_fp16)[name = tensor<string, []>("op_5632_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_3_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321185280)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5633_cast_fp16 = mul(x = var_5632_cast_fp16, y = core_stacks_4_layers_3_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_5633_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_703_cast_fp16 = add(x = input_481_cast_fp16, y = var_5633_cast_fp16)[name = tensor<string, []>("x_703_cast_fp16")];
            tensor<string, []> var_5644_pad_type_0 = const()[name = tensor<string, []>("op_5644_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5644_strides_0 = const()[name = tensor<string, []>("op_5644_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5644_pad_0 = const()[name = tensor<string, []>("op_5644_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5644_dilations_0 = const()[name = tensor<string, []>("op_5644_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5644_groups_0 = const()[name = tensor<string, []>("op_5644_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [48, 512, 1, 1]> core_stacks_4_layers_3_self_attn2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_self_attn2_in_proj_weight_to_fp16"), val = tensor<fp16, [48, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321186368)))];
            tensor<fp16, [48]> core_stacks_4_layers_3_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321235584)))];
            tensor<fp16, [1, 48, 1, 1024]> var_5644_cast_fp16 = conv(bias = core_stacks_4_layers_3_self_attn2_in_proj_bias_to_fp16, dilations = var_5644_dilations_0, groups = var_5644_groups_0, pad = var_5644_pad_0, pad_type = var_5644_pad_type_0, strides = var_5644_strides_0, weight = core_stacks_4_layers_3_self_attn2_in_proj_weight_to_fp16, x = x_703_cast_fp16)[name = tensor<string, []>("op_5644_cast_fp16")];
            tensor<int32, [4]> var_5645 = const()[name = tensor<string, []>("op_5645"), val = tensor<int32, [4]>([4, 12, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> var_5646_cast_fp16 = reshape(shape = var_5645, x = var_5644_cast_fp16)[name = tensor<string, []>("op_5646_cast_fp16")];
            tensor<int32, [4]> transpose_127_perm_0 = const()[name = tensor<string, []>("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [3]> concat_635 = const()[name = tensor<string, []>("concat_635"), val = tensor<int32, [3]>([4, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> transpose_127_cast_fp16 = transpose(perm = transpose_127_perm_0, x = var_5646_cast_fp16)[name = tensor<string, []>("transpose_130")];
            tensor<fp16, [4, 1024, 12]> reshape_190_cast_fp16 = reshape(shape = concat_635, x = transpose_127_cast_fp16)[name = tensor<string, []>("reshape_190_cast_fp16")];
            tensor<bool, []> matmul_63_transpose_x_0 = const()[name = tensor<string, []>("matmul_63_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> matmul_63_transpose_y_0 = const()[name = tensor<string, []>("matmul_63_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1024, 12]> matmul_63_cast_fp16 = matmul(transpose_x = matmul_63_transpose_x_0, transpose_y = matmul_63_transpose_y_0, x = reshape_186_cast_fp16, y = reshape_190_cast_fp16)[name = tensor<string, []>("matmul_63_cast_fp16")];
            tensor<int32, [4]> concat_639 = const()[name = tensor<string, []>("concat_639"), val = tensor<int32, [4]>([4, 1, 1024, 12])];
            tensor<fp16, [4, 1, 1024, 12]> reshape_191_cast_fp16 = reshape(shape = concat_639, x = matmul_63_cast_fp16)[name = tensor<string, []>("reshape_191_cast_fp16")];
            tensor<int32, [4]> out_perm_0 = const()[name = tensor<string, []>("out_perm_0"), val = tensor<int32, [4]>([0, 3, 1, 2])];
            tensor<int32, [4]> var_5649 = const()[name = tensor<string, []>("op_5649"), val = tensor<int32, [4]>([1, 48, 1, 1024])];
            tensor<fp16, [4, 12, 1, 1024]> out_cast_fp16 = transpose(perm = out_perm_0, x = reshape_191_cast_fp16)[name = tensor<string, []>("transpose_129")];
            tensor<fp16, [1, 48, 1, 1024]> input_501_cast_fp16 = reshape(shape = var_5649, x = out_cast_fp16)[name = tensor<string, []>("input_501_cast_fp16")];
            tensor<string, []> var_5657_pad_type_0 = const()[name = tensor<string, []>("op_5657_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5657_strides_0 = const()[name = tensor<string, []>("op_5657_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5657_pad_0 = const()[name = tensor<string, []>("op_5657_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5657_dilations_0 = const()[name = tensor<string, []>("op_5657_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5657_groups_0 = const()[name = tensor<string, []>("op_5657_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 48, 1, 1]> core_stacks_4_layers_3_self_attn2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_self_attn2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 48, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321235776)))];
            tensor<fp16, [512]> core_stacks_4_layers_3_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321284992)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5657_cast_fp16 = conv(bias = core_stacks_4_layers_3_self_attn2_out_proj_bias_to_fp16, dilations = var_5657_dilations_0, groups = var_5657_groups_0, pad = var_5657_pad_0, pad_type = var_5657_pad_type_0, strides = var_5657_strides_0, weight = core_stacks_4_layers_3_self_attn2_out_proj_weight_to_fp16, x = input_501_cast_fp16)[name = tensor<string, []>("op_5657_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_705_cast_fp16 = add(x = x_703_cast_fp16, y = var_5657_cast_fp16)[name = tensor<string, []>("x_705_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_707_cast_fp16 = add(x = x_705_cast_fp16, y = time_emb_cast_fp16)[name = tensor<string, []>("x_707_cast_fp16")];
            tensor<string, []> x_709_pad_type_0 = const()[name = tensor<string, []>("x_709_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_709_strides_0 = const()[name = tensor<string, []>("x_709_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_709_pad_0 = const()[name = tensor<string, []>("x_709_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_709_dilations_0 = const()[name = tensor<string, []>("x_709_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_709_groups_0 = const()[name = tensor<string, []>("x_709_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1024, 512, 1, 1]> core_stacks_4_layers_3_conv_module2_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module2_in_proj_weight_to_fp16"), val = tensor<fp16, [1024, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321286080)))];
            tensor<fp16, [1024]> core_stacks_4_layers_3_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322334720)))];
            tensor<fp16, [1, 1024, 1, 1024]> x_709_cast_fp16 = conv(bias = core_stacks_4_layers_3_conv_module2_in_proj_bias_to_fp16, dilations = x_709_dilations_0, groups = x_709_groups_0, pad = x_709_pad_0, pad_type = x_709_pad_type_0, strides = x_709_strides_0, weight = core_stacks_4_layers_3_conv_module2_in_proj_weight_to_fp16, x = x_707_cast_fp16)[name = tensor<string, []>("x_709_cast_fp16")];
            tensor<int32, [4]> var_5675_begin_0 = const()[name = tensor<string, []>("op_5675_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5675_end_0 = const()[name = tensor<string, []>("op_5675_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_5675_end_mask_0 = const()[name = tensor<string, []>("op_5675_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5675_cast_fp16 = slice_by_index(begin = var_5675_begin_0, end = var_5675_end_0, end_mask = var_5675_end_mask_0, x = x_709_cast_fp16)[name = tensor<string, []>("op_5675_cast_fp16")];
            tensor<int32, [4]> var_5677_begin_0 = const()[name = tensor<string, []>("op_5677_begin_0"), val = tensor<int32, [4]>([0, 512, 0, 0])];
            tensor<int32, [4]> var_5677_end_0 = const()[name = tensor<string, []>("op_5677_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> var_5677_end_mask_0 = const()[name = tensor<string, []>("op_5677_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5677_cast_fp16 = slice_by_index(begin = var_5677_begin_0, end = var_5677_end_0, end_mask = var_5677_end_mask_0, x = x_709_cast_fp16)[name = tensor<string, []>("op_5677_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5678_cast_fp16 = sigmoid(x = var_5677_cast_fp16)[name = tensor<string, []>("op_5678_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_711_cast_fp16 = mul(x = var_5675_cast_fp16, y = var_5678_cast_fp16)[name = tensor<string, []>("x_711_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_503_cast_fp16 = mul(x = x_711_cast_fp16, y = conv_keep_1_cast_fp16)[name = tensor<string, []>("input_503_cast_fp16")];
            tensor<string, []> var_5687_pad_type_0 = const()[name = tensor<string, []>("op_5687_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5687_pad_0 = const()[name = tensor<string, []>("op_5687_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5687_groups_0 = const()[name = tensor<string, []>("op_5687_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5687_strides_0 = const()[name = tensor<string, []>("op_5687_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5687_dilations_0 = const()[name = tensor<string, []>("op_5687_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_3_conv_module2_dw_a_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module2_dw_a_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322336832)))];
            tensor<fp16, [512]> core_stacks_4_layers_3_conv_module2_dw_a_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module2_dw_a_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322352256)))];
            tensor<fp16, [1, 512, 1, 1040]> var_5687_cast_fp16 = conv(bias = core_stacks_4_layers_3_conv_module2_dw_a_bias_to_fp16, dilations = var_5687_dilations_0, groups = var_5687_groups_0, pad = var_5687_pad_0, pad_type = var_5687_pad_type_0, strides = var_5687_strides_0, weight = core_stacks_4_layers_3_conv_module2_dw_a_weight_to_fp16, x = input_503_cast_fp16)[name = tensor<string, []>("op_5687_cast_fp16")];
            tensor<int32, [4]> var_5688_begin_0 = const()[name = tensor<string, []>("op_5688_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5688_end_0 = const()[name = tensor<string, []>("op_5688_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1024])];
            tensor<bool, [4]> var_5688_end_mask_0 = const()[name = tensor<string, []>("op_5688_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_5688_cast_fp16 = slice_by_index(begin = var_5688_begin_0, end = var_5688_end_0, end_mask = var_5688_end_mask_0, x = var_5687_cast_fp16)[name = tensor<string, []>("op_5688_cast_fp16")];
            tensor<string, []> var_5694_pad_type_0 = const()[name = tensor<string, []>("op_5694_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5694_pad_0 = const()[name = tensor<string, []>("op_5694_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5694_groups_0 = const()[name = tensor<string, []>("op_5694_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5694_strides_0 = const()[name = tensor<string, []>("op_5694_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5694_dilations_0 = const()[name = tensor<string, []>("op_5694_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 15]> core_stacks_4_layers_3_conv_module2_dw_b_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module2_dw_b_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322353344)))];
            tensor<fp16, [1, 512, 1, 1040]> var_5694_cast_fp16 = conv(dilations = var_5694_dilations_0, groups = var_5694_groups_0, pad = var_5694_pad_0, pad_type = var_5694_pad_type_0, strides = var_5694_strides_0, weight = core_stacks_4_layers_3_conv_module2_dw_b_weight_to_fp16, x = input_503_cast_fp16)[name = tensor<string, []>("op_5694_cast_fp16")];
            tensor<int32, [4]> var_5698_begin_0 = const()[name = tensor<string, []>("op_5698_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 15])];
            tensor<int32, [4]> var_5698_end_0 = const()[name = tensor<string, []>("op_5698_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1039])];
            tensor<bool, [4]> var_5698_end_mask_0 = const()[name = tensor<string, []>("op_5698_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 512, 1, 1024]> var_5698_cast_fp16 = slice_by_index(begin = var_5698_begin_0, end = var_5698_end_0, end_mask = var_5698_end_mask_0, x = var_5694_cast_fp16)[name = tensor<string, []>("op_5698_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5699_cast_fp16 = add(x = var_5688_cast_fp16, y = var_5698_cast_fp16)[name = tensor<string, []>("op_5699_cast_fp16")];
            tensor<string, []> var_5705_pad_type_0 = const()[name = tensor<string, []>("op_5705_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> var_5705_pad_0 = const()[name = tensor<string, []>("op_5705_pad_0"), val = tensor<int32, [4]>([0, 0, 15, 15])];
            tensor<int32, []> var_5705_groups_0 = const()[name = tensor<string, []>("op_5705_groups_0"), val = tensor<int32, []>(512)];
            tensor<int32, [2]> var_5705_strides_0 = const()[name = tensor<string, []>("op_5705_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_5705_dilations_0 = const()[name = tensor<string, []>("op_5705_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [512, 1, 1, 1]> core_stacks_4_layers_3_conv_module2_dw_c_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module2_dw_c_weight_to_fp16"), val = tensor<fp16, [512, 1, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322368768)))];
            tensor<fp16, [1, 512, 1, 1054]> var_5705_cast_fp16 = conv(dilations = var_5705_dilations_0, groups = var_5705_groups_0, pad = var_5705_pad_0, pad_type = var_5705_pad_type_0, strides = var_5705_strides_0, weight = core_stacks_4_layers_3_conv_module2_dw_c_weight_to_fp16, x = input_503_cast_fp16)[name = tensor<string, []>("op_5705_cast_fp16")];
            tensor<int32, [4]> var_5709_begin_0 = const()[name = tensor<string, []>("op_5709_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 30])];
            tensor<int32, [4]> var_5709_end_0 = const()[name = tensor<string, []>("op_5709_end_0"), val = tensor<int32, [4]>([1, 512, 1, 1])];
            tensor<bool, [4]> var_5709_end_mask_0 = const()[name = tensor<string, []>("op_5709_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 512, 1, 1024]> var_5709_cast_fp16 = slice_by_index(begin = var_5709_begin_0, end = var_5709_end_0, end_mask = var_5709_end_mask_0, x = var_5705_cast_fp16)[name = tensor<string, []>("op_5709_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_713_cast_fp16 = add(x = var_5699_cast_fp16, y = var_5709_cast_fp16)[name = tensor<string, []>("x_713_cast_fp16")];
            tensor<fp16, []> var_5711_to_fp16 = const()[name = tensor<string, []>("op_5711_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> x_715_cast_fp16 = sub(x = x_713_cast_fp16, y = var_5711_to_fp16)[name = tensor<string, []>("x_715_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5713_cast_fp16 = relu(x = x_715_cast_fp16)[name = tensor<string, []>("op_5713_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5714_cast_fp16 = abs(x = x_715_cast_fp16)[name = tensor<string, []>("op_5714_cast_fp16")];
            tensor<fp16, []> const_128_promoted_to_fp16 = const()[name = tensor<string, []>("const_128_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_5715_cast_fp16 = mul(x = var_5714_cast_fp16, y = const_128_promoted_to_fp16)[name = tensor<string, []>("op_5715_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5716_cast_fp16 = exp(x = var_5715_cast_fp16)[name = tensor<string, []>("op_5716_cast_fp16")];
            tensor<fp32, []> var_5717_epsilon_0 = const()[name = tensor<string, []>("op_5717_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 512, 1, 1024]> var_5717_cast_fp16 = log(epsilon = var_5717_epsilon_0, x = var_5716_cast_fp16)[name = tensor<string, []>("op_5717_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5718_cast_fp16 = add(x = var_5713_cast_fp16, y = var_5717_cast_fp16)[name = tensor<string, []>("op_5718_cast_fp16")];
            tensor<fp16, []> var_5719_to_fp16 = const()[name = tensor<string, []>("op_5719_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 512, 1, 1024]> var_5720_cast_fp16 = mul(x = x_713_cast_fp16, y = var_5719_to_fp16)[name = tensor<string, []>("op_5720_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5721_cast_fp16 = sub(x = var_5718_cast_fp16, y = var_5720_cast_fp16)[name = tensor<string, []>("op_5721_cast_fp16")];
            tensor<fp16, []> var_5722_to_fp16 = const()[name = tensor<string, []>("op_5722_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 512, 1, 1024]> input_505_cast_fp16 = sub(x = var_5721_cast_fp16, y = var_5722_to_fp16)[name = tensor<string, []>("input_505_cast_fp16")];
            tensor<string, []> var_5730_pad_type_0 = const()[name = tensor<string, []>("op_5730_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5730_strides_0 = const()[name = tensor<string, []>("op_5730_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5730_pad_0 = const()[name = tensor<string, []>("op_5730_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5730_dilations_0 = const()[name = tensor<string, []>("op_5730_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5730_groups_0 = const()[name = tensor<string, []>("op_5730_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 512, 1, 1]> core_stacks_4_layers_3_conv_module2_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module2_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322369856)))];
            tensor<fp16, [512]> core_stacks_4_layers_3_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322894208)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5730_cast_fp16 = conv(bias = core_stacks_4_layers_3_conv_module2_out_proj_bias_to_fp16, dilations = var_5730_dilations_0, groups = var_5730_groups_0, pad = var_5730_pad_0, pad_type = var_5730_pad_type_0, strides = var_5730_strides_0, weight = core_stacks_4_layers_3_conv_module2_out_proj_weight_to_fp16, x = input_505_cast_fp16)[name = tensor<string, []>("op_5730_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_507_cast_fp16 = add(x = x_707_cast_fp16, y = var_5730_cast_fp16)[name = tensor<string, []>("input_507_cast_fp16")];
            tensor<string, []> x_717_pad_type_0 = const()[name = tensor<string, []>("x_717_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_717_strides_0 = const()[name = tensor<string, []>("x_717_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_717_pad_0 = const()[name = tensor<string, []>("x_717_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_717_dilations_0 = const()[name = tensor<string, []>("x_717_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_717_groups_0 = const()[name = tensor<string, []>("x_717_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [1920, 512, 1, 1]> core_stacks_4_layers_3_feed_forward3_in_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward3_in_proj_weight_to_fp16"), val = tensor<fp16, [1920, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322895296)))];
            tensor<fp16, [1920]> core_stacks_4_layers_3_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [1920]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324861440)))];
            tensor<fp16, [1, 1920, 1, 1024]> x_717_cast_fp16 = conv(bias = core_stacks_4_layers_3_feed_forward3_in_proj_bias_to_fp16, dilations = x_717_dilations_0, groups = x_717_groups_0, pad = x_717_pad_0, pad_type = x_717_pad_type_0, strides = x_717_strides_0, weight = core_stacks_4_layers_3_feed_forward3_in_proj_weight_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("x_717_cast_fp16")];
            tensor<fp16, []> var_5741_to_fp16 = const()[name = tensor<string, []>("op_5741_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 1920, 1, 1024]> x_719_cast_fp16 = sub(x = x_717_cast_fp16, y = var_5741_to_fp16)[name = tensor<string, []>("x_719_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5743_cast_fp16 = relu(x = x_719_cast_fp16)[name = tensor<string, []>("op_5743_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5744_cast_fp16 = abs(x = x_719_cast_fp16)[name = tensor<string, []>("op_5744_cast_fp16")];
            tensor<fp16, []> const_129_promoted_to_fp16 = const()[name = tensor<string, []>("const_129_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_5745_cast_fp16 = mul(x = var_5744_cast_fp16, y = const_129_promoted_to_fp16)[name = tensor<string, []>("op_5745_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5746_cast_fp16 = exp(x = var_5745_cast_fp16)[name = tensor<string, []>("op_5746_cast_fp16")];
            tensor<fp32, []> var_5747_epsilon_0 = const()[name = tensor<string, []>("op_5747_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 1920, 1, 1024]> var_5747_cast_fp16 = log(epsilon = var_5747_epsilon_0, x = var_5746_cast_fp16)[name = tensor<string, []>("op_5747_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5748_cast_fp16 = add(x = var_5743_cast_fp16, y = var_5747_cast_fp16)[name = tensor<string, []>("op_5748_cast_fp16")];
            tensor<fp16, []> var_5749_to_fp16 = const()[name = tensor<string, []>("op_5749_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 1920, 1, 1024]> var_5750_cast_fp16 = mul(x = x_717_cast_fp16, y = var_5749_to_fp16)[name = tensor<string, []>("op_5750_cast_fp16")];
            tensor<fp16, [1, 1920, 1, 1024]> var_5751_cast_fp16 = sub(x = var_5748_cast_fp16, y = var_5750_cast_fp16)[name = tensor<string, []>("op_5751_cast_fp16")];
            tensor<fp16, []> var_5752_to_fp16 = const()[name = tensor<string, []>("op_5752_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 1920, 1, 1024]> input_509_cast_fp16 = sub(x = var_5751_cast_fp16, y = var_5752_to_fp16)[name = tensor<string, []>("input_509_cast_fp16")];
            tensor<string, []> var_5760_pad_type_0 = const()[name = tensor<string, []>("op_5760_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> var_5760_strides_0 = const()[name = tensor<string, []>("op_5760_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5760_pad_0 = const()[name = tensor<string, []>("op_5760_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5760_dilations_0 = const()[name = tensor<string, []>("op_5760_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> var_5760_groups_0 = const()[name = tensor<string, []>("op_5760_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [512, 1920, 1, 1]> core_stacks_4_layers_3_feed_forward3_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward3_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 1920, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324865344)))];
            tensor<fp16, [512]> core_stacks_4_layers_3_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326831488)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5760_cast_fp16 = conv(bias = core_stacks_4_layers_3_feed_forward3_out_proj_bias_to_fp16, dilations = var_5760_dilations_0, groups = var_5760_groups_0, pad = var_5760_pad_0, pad_type = var_5760_pad_type_0, strides = var_5760_strides_0, weight = core_stacks_4_layers_3_feed_forward3_out_proj_weight_to_fp16, x = input_509_cast_fp16)[name = tensor<string, []>("op_5760_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> x_cast_fp16 = add(x = input_507_cast_fp16, y = var_5760_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_3_norm_bias_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_norm_bias_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326832576)))];
            tensor<fp16, [1, 512, 1, 1024]> d_cast_fp16 = sub(x = x_cast_fp16, y = core_stacks_4_layers_3_norm_bias_to_fp16)[name = tensor<string, []>("d_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5765_cast_fp16 = mul(x = d_cast_fp16, y = d_cast_fp16)[name = tensor<string, []>("op_5765_cast_fp16")];
            tensor<int32, [1]> var_5767_axes_0 = const()[name = tensor<string, []>("op_5767_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> var_5767_keep_dims_0 = const()[name = tensor<string, []>("op_5767_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 1, 1, 1024]> var_5767_cast_fp16 = reduce_mean(axes = var_5767_axes_0, keep_dims = var_5767_keep_dims_0, x = var_5765_cast_fp16)[name = tensor<string, []>("op_5767_cast_fp16")];
            tensor<fp32, []> scales_epsilon_0 = const()[name = tensor<string, []>("scales_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1, 1024]> scales_cast_fp16 = rsqrt(epsilon = scales_epsilon_0, x = var_5767_cast_fp16)[name = tensor<string, []>("scales_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5769_cast_fp16 = mul(x = x_cast_fp16, y = scales_cast_fp16)[name = tensor<string, []>("op_5769_cast_fp16")];
            tensor<fp16, []> core_stacks_4_layers_3_norm_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_norm_scale_to_fp16"), val = tensor<fp16, []>(0x1.90cp+0)];
            tensor<fp16, [1, 512, 1, 1024]> src_cast_fp16 = mul(x = var_5769_cast_fp16, y = core_stacks_4_layers_3_norm_scale_to_fp16)[name = tensor<string, []>("src_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> var_5772_cast_fp16 = sub(x = src_cast_fp16, y = input_481_cast_fp16)[name = tensor<string, []>("op_5772_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1]> core_stacks_4_layers_3_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("core_stacks_4_layers_3_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [1, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326833664)))];
            tensor<fp16, [1, 512, 1, 1024]> var_5773_cast_fp16 = mul(x = var_5772_cast_fp16, y = core_stacks_4_layers_3_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_5773_cast_fp16")];
            tensor<fp16, [1, 512, 1, 1024]> input_cast_fp16 = add(x = input_481_cast_fp16, y = var_5773_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
            tensor<string, []> v_pad_type_0 = const()[name = tensor<string, []>("v_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> v_strides_0 = const()[name = tensor<string, []>("v_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> v_pad_0 = const()[name = tensor<string, []>("v_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> v_dilations_0 = const()[name = tensor<string, []>("v_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> v_groups_0 = const()[name = tensor<string, []>("v_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [100, 512, 1, 1]> core_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("core_out_proj_weight_to_fp16"), val = tensor<fp16, [100, 512, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326834752)))];
            tensor<fp16, [100]> core_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("core_out_proj_bias_to_fp16"), val = tensor<fp16, [100]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326937216)))];
            tensor<fp16, [1, 100, 1, 1024]> v_cast_fp16 = conv(bias = core_out_proj_bias_to_fp16, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = core_out_proj_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
            tensor<int32, [1]> var_5783_axes_0 = const()[name = tensor<string, []>("op_5783_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 100, 1024]> var_5783_cast_fp16 = squeeze(axes = var_5783_axes_0, x = v_cast_fp16)[name = tensor<string, []>("op_5783_cast_fp16")];
            tensor<int32, [3]> var_5787 = const()[name = tensor<string, []>("op_5787"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<string, []> var_5788_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_5788_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp16, [1, 1024, 100]> var_5788_cast_fp16 = transpose(perm = var_5787, x = var_5783_cast_fp16)[name = tensor<string, []>("transpose_128")];
            tensor<fp32, [1, 1024, 100]> v = cast(dtype = var_5788_cast_fp16_to_fp32_dtype_0, x = var_5788_cast_fp16)[name = tensor<string, []>("cast_131")];
        } -> (v);
}