Instructions to use openbmb/MiniCPM-V-4-gguf with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use openbmb/MiniCPM-V-4-gguf with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="openbmb/MiniCPM-V-4-gguf", trust_remote_code=True)
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("openbmb/MiniCPM-V-4-gguf", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use openbmb/MiniCPM-V-4-gguf with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf openbmb/MiniCPM-V-4-gguf:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf openbmb/MiniCPM-V-4-gguf:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf openbmb/MiniCPM-V-4-gguf:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf openbmb/MiniCPM-V-4-gguf:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf openbmb/MiniCPM-V-4-gguf:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf openbmb/MiniCPM-V-4-gguf:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf openbmb/MiniCPM-V-4-gguf:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf openbmb/MiniCPM-V-4-gguf:Q4_K_M

Use Docker

docker model run hf.co/openbmb/MiniCPM-V-4-gguf:Q4_K_M

LM Studio
Jan

vLLM

How to use openbmb/MiniCPM-V-4-gguf with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "openbmb/MiniCPM-V-4-gguf"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "openbmb/MiniCPM-V-4-gguf",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/openbmb/MiniCPM-V-4-gguf:Q4_K_M

SGLang

How to use openbmb/MiniCPM-V-4-gguf with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "openbmb/MiniCPM-V-4-gguf" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "openbmb/MiniCPM-V-4-gguf",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "openbmb/MiniCPM-V-4-gguf" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "openbmb/MiniCPM-V-4-gguf",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Ollama
How to use openbmb/MiniCPM-V-4-gguf with Ollama:
```
ollama run hf.co/openbmb/MiniCPM-V-4-gguf:Q4_K_M
```
Unsloth Desktop
Docker Model Runner
How to use openbmb/MiniCPM-V-4-gguf with Docker Model Runner:
```
docker model run hf.co/openbmb/MiniCPM-V-4-gguf:Q4_K_M
```

Lemonade

How to use openbmb/MiniCPM-V-4-gguf with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull openbmb/MiniCPM-V-4-gguf:Q4_K_M

Run and chat with the model

lemonade run user.MiniCPM-V-4-gguf-Q4_K_M

List all available models

lemonade list

Atomic Chat

MiniCPM-V-4-gguf

File size: 349,418 Bytes

08b971f

program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
{
    func main<ios15>(tensor<fp32, [1, 3, 14, 14336]> pixel_values, tensor<fp32, [1024, 1, 2560]> pos_embed, tensor<int32, [1, 1024]> position_ids) {
            tensor<int32, []> var_14 = const()[name = tensor<string, []>("op_14"), val = tensor<int32, []>(1)];
            tensor<int32, []> var_18 = const()[name = tensor<string, []>("op_18"), val = tensor<int32, []>(-1)];
            tensor<int32, [2]> var_32 = const()[name = tensor<string, []>("op_32"), val = tensor<int32, [2]>([14, 14])];
            tensor<int32, [2]> var_33 = const()[name = tensor<string, []>("op_33"), val = tensor<int32, [2]>([1, 1])];
            tensor<string, []> patch_embeds_pad_type_0 = const()[name = tensor<string, []>("patch_embeds_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [4]> patch_embeds_pad_0 = const()[name = tensor<string, []>("patch_embeds_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<string, []> pixel_values_to_fp16_dtype_0 = const()[name = tensor<string, []>("pixel_values_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1152, 3, 14, 14]> vpm_embeddings_patch_embedding_weight_to_fp16 = const()[name = tensor<string, []>("vpm_embeddings_patch_embedding_weight_to_fp16"), val = tensor<fp16, [1152, 3, 14, 14]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [1152]> vpm_embeddings_patch_embedding_bias_to_fp16 = const()[name = tensor<string, []>("vpm_embeddings_patch_embedding_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1354880)))];
            tensor<fp16, [1, 3, 14, 14336]> pixel_values_to_fp16 = cast(dtype = pixel_values_to_fp16_dtype_0, x = pixel_values)[name = tensor<string, []>("cast_150")];
            tensor<fp16, [1, 1152, 1, 1024]> patch_embeds_cast_fp16 = conv(bias = vpm_embeddings_patch_embedding_bias_to_fp16, dilations = var_33, groups = var_14, pad = patch_embeds_pad_0, pad_type = patch_embeds_pad_type_0, strides = var_32, weight = vpm_embeddings_patch_embedding_weight_to_fp16, x = pixel_values_to_fp16)[name = tensor<string, []>("patch_embeds_cast_fp16")];
            tensor<int32, [3]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<int32, [3]>([1, 1152, 1024])];
            tensor<fp16, [1, 1152, 1024]> var_35_cast_fp16 = reshape(shape = concat_0, x = patch_embeds_cast_fp16)[name = tensor<string, []>("op_35_cast_fp16")];
            tensor<int32, [3]> embeddings_1_perm_0 = const()[name = tensor<string, []>("embeddings_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, []> var_40_axis_0 = const()[name = tensor<string, []>("op_40_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [4900, 1152]> vpm_embeddings_position_embedding_weight_to_fp16 = const()[name = tensor<string, []>("vpm_embeddings_position_embedding_weight_to_fp16"), val = tensor<fp16, [4900, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1357248)))];
            tensor<fp16, [1, 1024, 1152]> var_40_cast_fp16 = gather(axis = var_40_axis_0, indices = position_ids, x = vpm_embeddings_position_embedding_weight_to_fp16)[name = tensor<string, []>("op_40_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> embeddings_1_cast_fp16 = transpose(perm = embeddings_1_perm_0, x = var_35_cast_fp16)[name = tensor<string, []>("transpose_250")];
            tensor<fp16, [1, 1024, 1152]> input_3_cast_fp16 = add(x = embeddings_1_cast_fp16, y = var_40_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<int32, [1]> hidden_states_1_axes_0 = const()[name = tensor<string, []>("hidden_states_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_0_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12646912)))];
            tensor<fp16, [1152]> vpm_encoder_layers_0_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12649280)))];
            tensor<fp16, []> var_12_to_fp16 = const()[name = tensor<string, []>("op_12_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1152]> hidden_states_1_cast_fp16 = layer_norm(axes = hidden_states_1_axes_0, beta = vpm_encoder_layers_0_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_0_layer_norm1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12651648)))];
            tensor<fp16, [1152]> vpm_encoder_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15305920)))];
            tensor<fp16, [1, 1024, 1152]> linear_0_cast_fp16 = linear(bias = vpm_encoder_layers_0_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_0_self_attn_q_proj_weight_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15308288)))];
            tensor<fp16, [1152]> vpm_encoder_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17962560)))];
            tensor<fp16, [1, 1024, 1152]> linear_1_cast_fp16 = linear(bias = vpm_encoder_layers_0_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_0_self_attn_k_proj_weight_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17964928)))];
            tensor<fp16, [1152]> vpm_encoder_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20619200)))];
            tensor<fp16, [1, 1024, 1152]> linear_2_cast_fp16 = linear(bias = vpm_encoder_layers_0_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_0_self_attn_v_proj_weight_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<int32, [4]> var_119 = const()[name = tensor<string, []>("op_119"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_120_cast_fp16 = reshape(shape = var_119, x = linear_0_cast_fp16)[name = tensor<string, []>("op_120_cast_fp16")];
            tensor<int32, [4]> var_122 = const()[name = tensor<string, []>("op_122"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_123_cast_fp16 = reshape(shape = var_122, x = linear_1_cast_fp16)[name = tensor<string, []>("op_123_cast_fp16")];
            tensor<int32, [4]> var_125 = const()[name = tensor<string, []>("op_125"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_126_cast_fp16 = reshape(shape = var_125, x = linear_2_cast_fp16)[name = tensor<string, []>("op_126_cast_fp16")];
            tensor<int32, [4]> value_states_3_perm_0 = const()[name = tensor<string, []>("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_129_transpose_x_0 = const()[name = tensor<string, []>("op_129_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_129_transpose_y_0 = const()[name = tensor<string, []>("op_129_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_84_perm_0 = const()[name = tensor<string, []>("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_84 = transpose(perm = transpose_84_perm_0, x = var_123_cast_fp16)[name = tensor<string, []>("transpose_247")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_120_cast_fp16)[name = tensor<string, []>("transpose_248")];
            tensor<fp16, [1, 16, 1024, 1024]> var_129_cast_fp16 = matmul(transpose_x = var_129_transpose_x_0, transpose_y = var_129_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor<string, []>("op_129_cast_fp16")];
            tensor<fp16, []> var_130_to_fp16 = const()[name = tensor<string, []>("op_130_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_1_cast_fp16 = mul(x = var_129_cast_fp16, y = var_130_to_fp16)[name = tensor<string, []>("attn_weights_1_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_132_cast_fp16 = softmax(axis = var_18, x = attn_weights_1_cast_fp16)[name = tensor<string, []>("op_132_cast_fp16")];
            tensor<bool, []> attn_output_1_transpose_x_0 = const()[name = tensor<string, []>("attn_output_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_1_transpose_y_0 = const()[name = tensor<string, []>("attn_output_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_3_cast_fp16 = transpose(perm = value_states_3_perm_0, x = var_126_cast_fp16)[name = tensor<string, []>("transpose_249")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = var_132_cast_fp16, y = value_states_3_cast_fp16)[name = tensor<string, []>("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_136_perm_0 = const()[name = tensor<string, []>("op_136_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_138 = const()[name = tensor<string, []>("op_138"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_136_cast_fp16 = transpose(perm = var_136_perm_0, x = attn_output_1_cast_fp16)[name = tensor<string, []>("transpose_246")];
            tensor<fp16, [1, 1024, 1152]> input_7_cast_fp16 = reshape(shape = var_138, x = var_136_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_0_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20621568)))];
            tensor<fp16, [1152]> vpm_encoder_layers_0_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23275840)))];
            tensor<fp16, [1, 1024, 1152]> linear_3_cast_fp16 = linear(bias = vpm_encoder_layers_0_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_0_self_attn_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_9_cast_fp16 = add(x = input_3_cast_fp16, y = linear_3_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
            tensor<int32, [1]> input_11_axes_0 = const()[name = tensor<string, []>("input_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_0_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23278208)))];
            tensor<fp16, [1152]> vpm_encoder_layers_0_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23280576)))];
            tensor<fp16, [1, 1024, 1152]> input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = vpm_encoder_layers_0_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_0_layer_norm2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_0_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23282944)))];
            tensor<fp16, [4304]> vpm_encoder_layers_0_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33199424)))];
            tensor<fp16, [1, 1024, 4304]> linear_4_cast_fp16 = linear(bias = vpm_encoder_layers_0_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_0_mlp_fc1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<string, []> input_15_mode_0 = const()[name = tensor<string, []>("input_15_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_15_cast_fp16 = gelu(mode = input_15_mode_0, x = linear_4_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_0_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33208128)))];
            tensor<fp16, [1152]> vpm_encoder_layers_0_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_0_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43124608)))];
            tensor<fp16, [1, 1024, 1152]> linear_5_cast_fp16 = linear(bias = vpm_encoder_layers_0_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_0_mlp_fc2_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_17_cast_fp16 = add(x = input_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
            tensor<int32, [1]> hidden_states_7_axes_0 = const()[name = tensor<string, []>("hidden_states_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_1_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43126976)))];
            tensor<fp16, [1152]> vpm_encoder_layers_1_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43129344)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_7_cast_fp16 = layer_norm(axes = hidden_states_7_axes_0, beta = vpm_encoder_layers_1_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_1_layer_norm1_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43131712)))];
            tensor<fp16, [1152]> vpm_encoder_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45785984)))];
            tensor<fp16, [1, 1024, 1152]> linear_6_cast_fp16 = linear(bias = vpm_encoder_layers_1_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_1_self_attn_q_proj_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45788352)))];
            tensor<fp16, [1152]> vpm_encoder_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48442624)))];
            tensor<fp16, [1, 1024, 1152]> linear_7_cast_fp16 = linear(bias = vpm_encoder_layers_1_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_1_self_attn_k_proj_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48444992)))];
            tensor<fp16, [1152]> vpm_encoder_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51099264)))];
            tensor<fp16, [1, 1024, 1152]> linear_8_cast_fp16 = linear(bias = vpm_encoder_layers_1_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_1_self_attn_v_proj_weight_to_fp16, x = hidden_states_7_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<int32, [4]> var_181 = const()[name = tensor<string, []>("op_181"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_182_cast_fp16 = reshape(shape = var_181, x = linear_6_cast_fp16)[name = tensor<string, []>("op_182_cast_fp16")];
            tensor<int32, [4]> var_184 = const()[name = tensor<string, []>("op_184"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_185_cast_fp16 = reshape(shape = var_184, x = linear_7_cast_fp16)[name = tensor<string, []>("op_185_cast_fp16")];
            tensor<int32, [4]> var_187 = const()[name = tensor<string, []>("op_187"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_188_cast_fp16 = reshape(shape = var_187, x = linear_8_cast_fp16)[name = tensor<string, []>("op_188_cast_fp16")];
            tensor<int32, [4]> value_states_7_perm_0 = const()[name = tensor<string, []>("value_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_191_transpose_x_0 = const()[name = tensor<string, []>("op_191_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_191_transpose_y_0 = const()[name = tensor<string, []>("op_191_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_86_perm_0 = const()[name = tensor<string, []>("transpose_86_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_86 = transpose(perm = transpose_86_perm_0, x = var_185_cast_fp16)[name = tensor<string, []>("transpose_243")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_85 = transpose(perm = transpose_85_perm_0, x = var_182_cast_fp16)[name = tensor<string, []>("transpose_244")];
            tensor<fp16, [1, 16, 1024, 1024]> var_191_cast_fp16 = matmul(transpose_x = var_191_transpose_x_0, transpose_y = var_191_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor<string, []>("op_191_cast_fp16")];
            tensor<fp16, []> var_192_to_fp16 = const()[name = tensor<string, []>("op_192_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_5_cast_fp16 = mul(x = var_191_cast_fp16, y = var_192_to_fp16)[name = tensor<string, []>("attn_weights_5_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_194_cast_fp16 = softmax(axis = var_18, x = attn_weights_5_cast_fp16)[name = tensor<string, []>("op_194_cast_fp16")];
            tensor<bool, []> attn_output_5_transpose_x_0 = const()[name = tensor<string, []>("attn_output_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_5_transpose_y_0 = const()[name = tensor<string, []>("attn_output_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_7_cast_fp16 = transpose(perm = value_states_7_perm_0, x = var_188_cast_fp16)[name = tensor<string, []>("transpose_245")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = var_194_cast_fp16, y = value_states_7_cast_fp16)[name = tensor<string, []>("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_198_perm_0 = const()[name = tensor<string, []>("op_198_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_200 = const()[name = tensor<string, []>("op_200"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_198_cast_fp16 = transpose(perm = var_198_perm_0, x = attn_output_5_cast_fp16)[name = tensor<string, []>("transpose_242")];
            tensor<fp16, [1, 1024, 1152]> input_21_cast_fp16 = reshape(shape = var_200, x = var_198_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_1_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51101632)))];
            tensor<fp16, [1152]> vpm_encoder_layers_1_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53755904)))];
            tensor<fp16, [1, 1024, 1152]> linear_9_cast_fp16 = linear(bias = vpm_encoder_layers_1_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_1_self_attn_out_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_9_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<int32, [1]> input_25_axes_0 = const()[name = tensor<string, []>("input_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_1_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53758272)))];
            tensor<fp16, [1152]> vpm_encoder_layers_1_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53760640)))];
            tensor<fp16, [1, 1024, 1152]> input_25_cast_fp16 = layer_norm(axes = input_25_axes_0, beta = vpm_encoder_layers_1_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_1_layer_norm2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_1_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53763008)))];
            tensor<fp16, [4304]> vpm_encoder_layers_1_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63679488)))];
            tensor<fp16, [1, 1024, 4304]> linear_10_cast_fp16 = linear(bias = vpm_encoder_layers_1_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_1_mlp_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<string, []> input_29_mode_0 = const()[name = tensor<string, []>("input_29_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = linear_10_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_1_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63688192)))];
            tensor<fp16, [1152]> vpm_encoder_layers_1_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_1_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73604672)))];
            tensor<fp16, [1, 1024, 1152]> linear_11_cast_fp16 = linear(bias = vpm_encoder_layers_1_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_1_mlp_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_31_cast_fp16 = add(x = input_23_cast_fp16, y = linear_11_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<int32, [1]> hidden_states_13_axes_0 = const()[name = tensor<string, []>("hidden_states_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_2_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73607040)))];
            tensor<fp16, [1152]> vpm_encoder_layers_2_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73609408)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_13_cast_fp16 = layer_norm(axes = hidden_states_13_axes_0, beta = vpm_encoder_layers_2_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_2_layer_norm1_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("hidden_states_13_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73611776)))];
            tensor<fp16, [1152]> vpm_encoder_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76266048)))];
            tensor<fp16, [1, 1024, 1152]> linear_12_cast_fp16 = linear(bias = vpm_encoder_layers_2_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_2_self_attn_q_proj_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76268416)))];
            tensor<fp16, [1152]> vpm_encoder_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78922688)))];
            tensor<fp16, [1, 1024, 1152]> linear_13_cast_fp16 = linear(bias = vpm_encoder_layers_2_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_2_self_attn_k_proj_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78925056)))];
            tensor<fp16, [1152]> vpm_encoder_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81579328)))];
            tensor<fp16, [1, 1024, 1152]> linear_14_cast_fp16 = linear(bias = vpm_encoder_layers_2_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_2_self_attn_v_proj_weight_to_fp16, x = hidden_states_13_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<int32, [4]> var_243 = const()[name = tensor<string, []>("op_243"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_244_cast_fp16 = reshape(shape = var_243, x = linear_12_cast_fp16)[name = tensor<string, []>("op_244_cast_fp16")];
            tensor<int32, [4]> var_246 = const()[name = tensor<string, []>("op_246"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_247_cast_fp16 = reshape(shape = var_246, x = linear_13_cast_fp16)[name = tensor<string, []>("op_247_cast_fp16")];
            tensor<int32, [4]> var_249 = const()[name = tensor<string, []>("op_249"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_250_cast_fp16 = reshape(shape = var_249, x = linear_14_cast_fp16)[name = tensor<string, []>("op_250_cast_fp16")];
            tensor<int32, [4]> value_states_11_perm_0 = const()[name = tensor<string, []>("value_states_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_253_transpose_x_0 = const()[name = tensor<string, []>("op_253_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_253_transpose_y_0 = const()[name = tensor<string, []>("op_253_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_88_perm_0 = const()[name = tensor<string, []>("transpose_88_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_88 = transpose(perm = transpose_88_perm_0, x = var_247_cast_fp16)[name = tensor<string, []>("transpose_239")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_87 = transpose(perm = transpose_87_perm_0, x = var_244_cast_fp16)[name = tensor<string, []>("transpose_240")];
            tensor<fp16, [1, 16, 1024, 1024]> var_253_cast_fp16 = matmul(transpose_x = var_253_transpose_x_0, transpose_y = var_253_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor<string, []>("op_253_cast_fp16")];
            tensor<fp16, []> var_254_to_fp16 = const()[name = tensor<string, []>("op_254_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_9_cast_fp16 = mul(x = var_253_cast_fp16, y = var_254_to_fp16)[name = tensor<string, []>("attn_weights_9_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_256_cast_fp16 = softmax(axis = var_18, x = attn_weights_9_cast_fp16)[name = tensor<string, []>("op_256_cast_fp16")];
            tensor<bool, []> attn_output_9_transpose_x_0 = const()[name = tensor<string, []>("attn_output_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_9_transpose_y_0 = const()[name = tensor<string, []>("attn_output_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_11_cast_fp16 = transpose(perm = value_states_11_perm_0, x = var_250_cast_fp16)[name = tensor<string, []>("transpose_241")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = var_256_cast_fp16, y = value_states_11_cast_fp16)[name = tensor<string, []>("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_260_perm_0 = const()[name = tensor<string, []>("op_260_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_262 = const()[name = tensor<string, []>("op_262"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_260_cast_fp16 = transpose(perm = var_260_perm_0, x = attn_output_9_cast_fp16)[name = tensor<string, []>("transpose_238")];
            tensor<fp16, [1, 1024, 1152]> input_35_cast_fp16 = reshape(shape = var_262, x = var_260_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_2_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81581696)))];
            tensor<fp16, [1152]> vpm_encoder_layers_2_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84235968)))];
            tensor<fp16, [1, 1024, 1152]> linear_15_cast_fp16 = linear(bias = vpm_encoder_layers_2_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_2_self_attn_out_proj_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_15_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
            tensor<int32, [1]> input_39_axes_0 = const()[name = tensor<string, []>("input_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_2_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84238336)))];
            tensor<fp16, [1152]> vpm_encoder_layers_2_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84240704)))];
            tensor<fp16, [1, 1024, 1152]> input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = vpm_encoder_layers_2_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_2_layer_norm2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_2_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84243072)))];
            tensor<fp16, [4304]> vpm_encoder_layers_2_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94159552)))];
            tensor<fp16, [1, 1024, 4304]> linear_16_cast_fp16 = linear(bias = vpm_encoder_layers_2_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_2_mlp_fc1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<string, []> input_43_mode_0 = const()[name = tensor<string, []>("input_43_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_43_cast_fp16 = gelu(mode = input_43_mode_0, x = linear_16_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_2_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94168256)))];
            tensor<fp16, [1152]> vpm_encoder_layers_2_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_2_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104084736)))];
            tensor<fp16, [1, 1024, 1152]> linear_17_cast_fp16 = linear(bias = vpm_encoder_layers_2_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_2_mlp_fc2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_45_cast_fp16 = add(x = input_37_cast_fp16, y = linear_17_cast_fp16)[name = tensor<string, []>("input_45_cast_fp16")];
            tensor<int32, [1]> hidden_states_19_axes_0 = const()[name = tensor<string, []>("hidden_states_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_3_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104087104)))];
            tensor<fp16, [1152]> vpm_encoder_layers_3_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104089472)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_19_cast_fp16 = layer_norm(axes = hidden_states_19_axes_0, beta = vpm_encoder_layers_3_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_3_layer_norm1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("hidden_states_19_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104091840)))];
            tensor<fp16, [1152]> vpm_encoder_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106746112)))];
            tensor<fp16, [1, 1024, 1152]> linear_18_cast_fp16 = linear(bias = vpm_encoder_layers_3_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_3_self_attn_q_proj_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106748480)))];
            tensor<fp16, [1152]> vpm_encoder_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109402752)))];
            tensor<fp16, [1, 1024, 1152]> linear_19_cast_fp16 = linear(bias = vpm_encoder_layers_3_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_3_self_attn_k_proj_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109405120)))];
            tensor<fp16, [1152]> vpm_encoder_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112059392)))];
            tensor<fp16, [1, 1024, 1152]> linear_20_cast_fp16 = linear(bias = vpm_encoder_layers_3_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_3_self_attn_v_proj_weight_to_fp16, x = hidden_states_19_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<int32, [4]> var_305 = const()[name = tensor<string, []>("op_305"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_306_cast_fp16 = reshape(shape = var_305, x = linear_18_cast_fp16)[name = tensor<string, []>("op_306_cast_fp16")];
            tensor<int32, [4]> var_308 = const()[name = tensor<string, []>("op_308"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_309_cast_fp16 = reshape(shape = var_308, x = linear_19_cast_fp16)[name = tensor<string, []>("op_309_cast_fp16")];
            tensor<int32, [4]> var_311 = const()[name = tensor<string, []>("op_311"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_312_cast_fp16 = reshape(shape = var_311, x = linear_20_cast_fp16)[name = tensor<string, []>("op_312_cast_fp16")];
            tensor<int32, [4]> value_states_15_perm_0 = const()[name = tensor<string, []>("value_states_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_315_transpose_x_0 = const()[name = tensor<string, []>("op_315_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_315_transpose_y_0 = const()[name = tensor<string, []>("op_315_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_89_perm_0 = const()[name = tensor<string, []>("transpose_89_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = tensor<string, []>("transpose_90_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_90 = transpose(perm = transpose_90_perm_0, x = var_309_cast_fp16)[name = tensor<string, []>("transpose_235")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_89 = transpose(perm = transpose_89_perm_0, x = var_306_cast_fp16)[name = tensor<string, []>("transpose_236")];
            tensor<fp16, [1, 16, 1024, 1024]> var_315_cast_fp16 = matmul(transpose_x = var_315_transpose_x_0, transpose_y = var_315_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor<string, []>("op_315_cast_fp16")];
            tensor<fp16, []> var_316_to_fp16 = const()[name = tensor<string, []>("op_316_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_13_cast_fp16 = mul(x = var_315_cast_fp16, y = var_316_to_fp16)[name = tensor<string, []>("attn_weights_13_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_318_cast_fp16 = softmax(axis = var_18, x = attn_weights_13_cast_fp16)[name = tensor<string, []>("op_318_cast_fp16")];
            tensor<bool, []> attn_output_13_transpose_x_0 = const()[name = tensor<string, []>("attn_output_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_13_transpose_y_0 = const()[name = tensor<string, []>("attn_output_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_15_cast_fp16 = transpose(perm = value_states_15_perm_0, x = var_312_cast_fp16)[name = tensor<string, []>("transpose_237")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = var_318_cast_fp16, y = value_states_15_cast_fp16)[name = tensor<string, []>("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_322_perm_0 = const()[name = tensor<string, []>("op_322_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_324 = const()[name = tensor<string, []>("op_324"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_322_cast_fp16 = transpose(perm = var_322_perm_0, x = attn_output_13_cast_fp16)[name = tensor<string, []>("transpose_234")];
            tensor<fp16, [1, 1024, 1152]> input_49_cast_fp16 = reshape(shape = var_324, x = var_322_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_3_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112061760)))];
            tensor<fp16, [1152]> vpm_encoder_layers_3_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114716032)))];
            tensor<fp16, [1, 1024, 1152]> linear_21_cast_fp16 = linear(bias = vpm_encoder_layers_3_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_3_self_attn_out_proj_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_21_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<int32, [1]> input_53_axes_0 = const()[name = tensor<string, []>("input_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_3_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114718400)))];
            tensor<fp16, [1152]> vpm_encoder_layers_3_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114720768)))];
            tensor<fp16, [1, 1024, 1152]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = vpm_encoder_layers_3_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_3_layer_norm2_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_3_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114723136)))];
            tensor<fp16, [4304]> vpm_encoder_layers_3_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124639616)))];
            tensor<fp16, [1, 1024, 4304]> linear_22_cast_fp16 = linear(bias = vpm_encoder_layers_3_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_3_mlp_fc1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<string, []> input_57_mode_0 = const()[name = tensor<string, []>("input_57_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_57_cast_fp16 = gelu(mode = input_57_mode_0, x = linear_22_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_3_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124648320)))];
            tensor<fp16, [1152]> vpm_encoder_layers_3_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_3_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134564800)))];
            tensor<fp16, [1, 1024, 1152]> linear_23_cast_fp16 = linear(bias = vpm_encoder_layers_3_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_3_mlp_fc2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_59_cast_fp16 = add(x = input_51_cast_fp16, y = linear_23_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<int32, [1]> hidden_states_25_axes_0 = const()[name = tensor<string, []>("hidden_states_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_4_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134567168)))];
            tensor<fp16, [1152]> vpm_encoder_layers_4_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134569536)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_25_cast_fp16 = layer_norm(axes = hidden_states_25_axes_0, beta = vpm_encoder_layers_4_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_4_layer_norm1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("hidden_states_25_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134571904)))];
            tensor<fp16, [1152]> vpm_encoder_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137226176)))];
            tensor<fp16, [1, 1024, 1152]> linear_24_cast_fp16 = linear(bias = vpm_encoder_layers_4_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_4_self_attn_q_proj_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137228544)))];
            tensor<fp16, [1152]> vpm_encoder_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139882816)))];
            tensor<fp16, [1, 1024, 1152]> linear_25_cast_fp16 = linear(bias = vpm_encoder_layers_4_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_4_self_attn_k_proj_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139885184)))];
            tensor<fp16, [1152]> vpm_encoder_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142539456)))];
            tensor<fp16, [1, 1024, 1152]> linear_26_cast_fp16 = linear(bias = vpm_encoder_layers_4_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_4_self_attn_v_proj_weight_to_fp16, x = hidden_states_25_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<int32, [4]> var_367 = const()[name = tensor<string, []>("op_367"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_368_cast_fp16 = reshape(shape = var_367, x = linear_24_cast_fp16)[name = tensor<string, []>("op_368_cast_fp16")];
            tensor<int32, [4]> var_370 = const()[name = tensor<string, []>("op_370"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_371_cast_fp16 = reshape(shape = var_370, x = linear_25_cast_fp16)[name = tensor<string, []>("op_371_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = tensor<string, []>("op_373"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_374_cast_fp16 = reshape(shape = var_373, x = linear_26_cast_fp16)[name = tensor<string, []>("op_374_cast_fp16")];
            tensor<int32, [4]> value_states_19_perm_0 = const()[name = tensor<string, []>("value_states_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_377_transpose_x_0 = const()[name = tensor<string, []>("op_377_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_377_transpose_y_0 = const()[name = tensor<string, []>("op_377_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_92_perm_0 = const()[name = tensor<string, []>("transpose_92_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_92 = transpose(perm = transpose_92_perm_0, x = var_371_cast_fp16)[name = tensor<string, []>("transpose_231")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_91 = transpose(perm = transpose_91_perm_0, x = var_368_cast_fp16)[name = tensor<string, []>("transpose_232")];
            tensor<fp16, [1, 16, 1024, 1024]> var_377_cast_fp16 = matmul(transpose_x = var_377_transpose_x_0, transpose_y = var_377_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor<string, []>("op_377_cast_fp16")];
            tensor<fp16, []> var_378_to_fp16 = const()[name = tensor<string, []>("op_378_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_17_cast_fp16 = mul(x = var_377_cast_fp16, y = var_378_to_fp16)[name = tensor<string, []>("attn_weights_17_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_380_cast_fp16 = softmax(axis = var_18, x = attn_weights_17_cast_fp16)[name = tensor<string, []>("op_380_cast_fp16")];
            tensor<bool, []> attn_output_17_transpose_x_0 = const()[name = tensor<string, []>("attn_output_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_17_transpose_y_0 = const()[name = tensor<string, []>("attn_output_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_19_cast_fp16 = transpose(perm = value_states_19_perm_0, x = var_374_cast_fp16)[name = tensor<string, []>("transpose_233")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = var_380_cast_fp16, y = value_states_19_cast_fp16)[name = tensor<string, []>("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_384_perm_0 = const()[name = tensor<string, []>("op_384_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_386 = const()[name = tensor<string, []>("op_386"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_384_cast_fp16 = transpose(perm = var_384_perm_0, x = attn_output_17_cast_fp16)[name = tensor<string, []>("transpose_230")];
            tensor<fp16, [1, 1024, 1152]> input_63_cast_fp16 = reshape(shape = var_386, x = var_384_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_4_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142541824)))];
            tensor<fp16, [1152]> vpm_encoder_layers_4_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145196096)))];
            tensor<fp16, [1, 1024, 1152]> linear_27_cast_fp16 = linear(bias = vpm_encoder_layers_4_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_4_self_attn_out_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_65_cast_fp16 = add(x = input_59_cast_fp16, y = linear_27_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
            tensor<int32, [1]> input_67_axes_0 = const()[name = tensor<string, []>("input_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_4_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145198464)))];
            tensor<fp16, [1152]> vpm_encoder_layers_4_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145200832)))];
            tensor<fp16, [1, 1024, 1152]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = vpm_encoder_layers_4_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_4_layer_norm2_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_4_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145203200)))];
            tensor<fp16, [4304]> vpm_encoder_layers_4_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155119680)))];
            tensor<fp16, [1, 1024, 4304]> linear_28_cast_fp16 = linear(bias = vpm_encoder_layers_4_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_4_mlp_fc1_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<string, []> input_71_mode_0 = const()[name = tensor<string, []>("input_71_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_28_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_4_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155128384)))];
            tensor<fp16, [1152]> vpm_encoder_layers_4_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_4_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165044864)))];
            tensor<fp16, [1, 1024, 1152]> linear_29_cast_fp16 = linear(bias = vpm_encoder_layers_4_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_4_mlp_fc2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_73_cast_fp16 = add(x = input_65_cast_fp16, y = linear_29_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
            tensor<int32, [1]> hidden_states_31_axes_0 = const()[name = tensor<string, []>("hidden_states_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_5_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165047232)))];
            tensor<fp16, [1152]> vpm_encoder_layers_5_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165049600)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_31_cast_fp16 = layer_norm(axes = hidden_states_31_axes_0, beta = vpm_encoder_layers_5_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_5_layer_norm1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("hidden_states_31_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165051968)))];
            tensor<fp16, [1152]> vpm_encoder_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167706240)))];
            tensor<fp16, [1, 1024, 1152]> linear_30_cast_fp16 = linear(bias = vpm_encoder_layers_5_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_5_self_attn_q_proj_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167708608)))];
            tensor<fp16, [1152]> vpm_encoder_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170362880)))];
            tensor<fp16, [1, 1024, 1152]> linear_31_cast_fp16 = linear(bias = vpm_encoder_layers_5_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_5_self_attn_k_proj_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170365248)))];
            tensor<fp16, [1152]> vpm_encoder_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173019520)))];
            tensor<fp16, [1, 1024, 1152]> linear_32_cast_fp16 = linear(bias = vpm_encoder_layers_5_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_5_self_attn_v_proj_weight_to_fp16, x = hidden_states_31_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<int32, [4]> var_429 = const()[name = tensor<string, []>("op_429"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_430_cast_fp16 = reshape(shape = var_429, x = linear_30_cast_fp16)[name = tensor<string, []>("op_430_cast_fp16")];
            tensor<int32, [4]> var_432 = const()[name = tensor<string, []>("op_432"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_433_cast_fp16 = reshape(shape = var_432, x = linear_31_cast_fp16)[name = tensor<string, []>("op_433_cast_fp16")];
            tensor<int32, [4]> var_435 = const()[name = tensor<string, []>("op_435"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_436_cast_fp16 = reshape(shape = var_435, x = linear_32_cast_fp16)[name = tensor<string, []>("op_436_cast_fp16")];
            tensor<int32, [4]> value_states_23_perm_0 = const()[name = tensor<string, []>("value_states_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_439_transpose_x_0 = const()[name = tensor<string, []>("op_439_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_439_transpose_y_0 = const()[name = tensor<string, []>("op_439_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_94_perm_0 = const()[name = tensor<string, []>("transpose_94_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_94 = transpose(perm = transpose_94_perm_0, x = var_433_cast_fp16)[name = tensor<string, []>("transpose_227")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_93 = transpose(perm = transpose_93_perm_0, x = var_430_cast_fp16)[name = tensor<string, []>("transpose_228")];
            tensor<fp16, [1, 16, 1024, 1024]> var_439_cast_fp16 = matmul(transpose_x = var_439_transpose_x_0, transpose_y = var_439_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor<string, []>("op_439_cast_fp16")];
            tensor<fp16, []> var_440_to_fp16 = const()[name = tensor<string, []>("op_440_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_21_cast_fp16 = mul(x = var_439_cast_fp16, y = var_440_to_fp16)[name = tensor<string, []>("attn_weights_21_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_442_cast_fp16 = softmax(axis = var_18, x = attn_weights_21_cast_fp16)[name = tensor<string, []>("op_442_cast_fp16")];
            tensor<bool, []> attn_output_21_transpose_x_0 = const()[name = tensor<string, []>("attn_output_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_21_transpose_y_0 = const()[name = tensor<string, []>("attn_output_21_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_23_cast_fp16 = transpose(perm = value_states_23_perm_0, x = var_436_cast_fp16)[name = tensor<string, []>("transpose_229")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = var_442_cast_fp16, y = value_states_23_cast_fp16)[name = tensor<string, []>("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_446_perm_0 = const()[name = tensor<string, []>("op_446_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_448 = const()[name = tensor<string, []>("op_448"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_446_cast_fp16 = transpose(perm = var_446_perm_0, x = attn_output_21_cast_fp16)[name = tensor<string, []>("transpose_226")];
            tensor<fp16, [1, 1024, 1152]> input_77_cast_fp16 = reshape(shape = var_448, x = var_446_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_5_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173021888)))];
            tensor<fp16, [1152]> vpm_encoder_layers_5_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175676160)))];
            tensor<fp16, [1, 1024, 1152]> linear_33_cast_fp16 = linear(bias = vpm_encoder_layers_5_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_5_self_attn_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_33_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<int32, [1]> input_81_axes_0 = const()[name = tensor<string, []>("input_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_5_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175678528)))];
            tensor<fp16, [1152]> vpm_encoder_layers_5_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175680896)))];
            tensor<fp16, [1, 1024, 1152]> input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = vpm_encoder_layers_5_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_5_layer_norm2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_5_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175683264)))];
            tensor<fp16, [4304]> vpm_encoder_layers_5_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185599744)))];
            tensor<fp16, [1, 1024, 4304]> linear_34_cast_fp16 = linear(bias = vpm_encoder_layers_5_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_5_mlp_fc1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<string, []> input_85_mode_0 = const()[name = tensor<string, []>("input_85_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_34_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_5_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185608448)))];
            tensor<fp16, [1152]> vpm_encoder_layers_5_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_5_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195524928)))];
            tensor<fp16, [1, 1024, 1152]> linear_35_cast_fp16 = linear(bias = vpm_encoder_layers_5_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_5_mlp_fc2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_87_cast_fp16 = add(x = input_79_cast_fp16, y = linear_35_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<int32, [1]> hidden_states_37_axes_0 = const()[name = tensor<string, []>("hidden_states_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_6_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195527296)))];
            tensor<fp16, [1152]> vpm_encoder_layers_6_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195529664)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_37_cast_fp16 = layer_norm(axes = hidden_states_37_axes_0, beta = vpm_encoder_layers_6_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_6_layer_norm1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("hidden_states_37_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195532032)))];
            tensor<fp16, [1152]> vpm_encoder_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198186304)))];
            tensor<fp16, [1, 1024, 1152]> linear_36_cast_fp16 = linear(bias = vpm_encoder_layers_6_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_6_self_attn_q_proj_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198188672)))];
            tensor<fp16, [1152]> vpm_encoder_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200842944)))];
            tensor<fp16, [1, 1024, 1152]> linear_37_cast_fp16 = linear(bias = vpm_encoder_layers_6_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_6_self_attn_k_proj_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(200845312)))];
            tensor<fp16, [1152]> vpm_encoder_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203499584)))];
            tensor<fp16, [1, 1024, 1152]> linear_38_cast_fp16 = linear(bias = vpm_encoder_layers_6_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_6_self_attn_v_proj_weight_to_fp16, x = hidden_states_37_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<int32, [4]> var_491 = const()[name = tensor<string, []>("op_491"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_492_cast_fp16 = reshape(shape = var_491, x = linear_36_cast_fp16)[name = tensor<string, []>("op_492_cast_fp16")];
            tensor<int32, [4]> var_494 = const()[name = tensor<string, []>("op_494"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_495_cast_fp16 = reshape(shape = var_494, x = linear_37_cast_fp16)[name = tensor<string, []>("op_495_cast_fp16")];
            tensor<int32, [4]> var_497 = const()[name = tensor<string, []>("op_497"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_498_cast_fp16 = reshape(shape = var_497, x = linear_38_cast_fp16)[name = tensor<string, []>("op_498_cast_fp16")];
            tensor<int32, [4]> value_states_27_perm_0 = const()[name = tensor<string, []>("value_states_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_501_transpose_x_0 = const()[name = tensor<string, []>("op_501_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_501_transpose_y_0 = const()[name = tensor<string, []>("op_501_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_96_perm_0 = const()[name = tensor<string, []>("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_495_cast_fp16)[name = tensor<string, []>("transpose_223")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_95 = transpose(perm = transpose_95_perm_0, x = var_492_cast_fp16)[name = tensor<string, []>("transpose_224")];
            tensor<fp16, [1, 16, 1024, 1024]> var_501_cast_fp16 = matmul(transpose_x = var_501_transpose_x_0, transpose_y = var_501_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor<string, []>("op_501_cast_fp16")];
            tensor<fp16, []> var_502_to_fp16 = const()[name = tensor<string, []>("op_502_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_25_cast_fp16 = mul(x = var_501_cast_fp16, y = var_502_to_fp16)[name = tensor<string, []>("attn_weights_25_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_504_cast_fp16 = softmax(axis = var_18, x = attn_weights_25_cast_fp16)[name = tensor<string, []>("op_504_cast_fp16")];
            tensor<bool, []> attn_output_25_transpose_x_0 = const()[name = tensor<string, []>("attn_output_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_25_transpose_y_0 = const()[name = tensor<string, []>("attn_output_25_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_27_cast_fp16 = transpose(perm = value_states_27_perm_0, x = var_498_cast_fp16)[name = tensor<string, []>("transpose_225")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = var_504_cast_fp16, y = value_states_27_cast_fp16)[name = tensor<string, []>("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_508_perm_0 = const()[name = tensor<string, []>("op_508_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_510 = const()[name = tensor<string, []>("op_510"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_508_cast_fp16 = transpose(perm = var_508_perm_0, x = attn_output_25_cast_fp16)[name = tensor<string, []>("transpose_222")];
            tensor<fp16, [1, 1024, 1152]> input_91_cast_fp16 = reshape(shape = var_510, x = var_508_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_6_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203501952)))];
            tensor<fp16, [1152]> vpm_encoder_layers_6_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206156224)))];
            tensor<fp16, [1, 1024, 1152]> linear_39_cast_fp16 = linear(bias = vpm_encoder_layers_6_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_6_self_attn_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_39_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")];
            tensor<int32, [1]> input_95_axes_0 = const()[name = tensor<string, []>("input_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_6_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206158592)))];
            tensor<fp16, [1152]> vpm_encoder_layers_6_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206160960)))];
            tensor<fp16, [1, 1024, 1152]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = vpm_encoder_layers_6_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_6_layer_norm2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_6_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(206163328)))];
            tensor<fp16, [4304]> vpm_encoder_layers_6_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216079808)))];
            tensor<fp16, [1, 1024, 4304]> linear_40_cast_fp16 = linear(bias = vpm_encoder_layers_6_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_6_mlp_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<string, []> input_99_mode_0 = const()[name = tensor<string, []>("input_99_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = linear_40_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_6_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216088512)))];
            tensor<fp16, [1152]> vpm_encoder_layers_6_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_6_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226004992)))];
            tensor<fp16, [1, 1024, 1152]> linear_41_cast_fp16 = linear(bias = vpm_encoder_layers_6_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_6_mlp_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_101_cast_fp16 = add(x = input_93_cast_fp16, y = linear_41_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")];
            tensor<int32, [1]> hidden_states_43_axes_0 = const()[name = tensor<string, []>("hidden_states_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_7_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226007360)))];
            tensor<fp16, [1152]> vpm_encoder_layers_7_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226009728)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_43_cast_fp16 = layer_norm(axes = hidden_states_43_axes_0, beta = vpm_encoder_layers_7_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_7_layer_norm1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("hidden_states_43_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226012096)))];
            tensor<fp16, [1152]> vpm_encoder_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228666368)))];
            tensor<fp16, [1, 1024, 1152]> linear_42_cast_fp16 = linear(bias = vpm_encoder_layers_7_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_7_self_attn_q_proj_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228668736)))];
            tensor<fp16, [1152]> vpm_encoder_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231323008)))];
            tensor<fp16, [1, 1024, 1152]> linear_43_cast_fp16 = linear(bias = vpm_encoder_layers_7_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_7_self_attn_k_proj_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231325376)))];
            tensor<fp16, [1152]> vpm_encoder_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233979648)))];
            tensor<fp16, [1, 1024, 1152]> linear_44_cast_fp16 = linear(bias = vpm_encoder_layers_7_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_7_self_attn_v_proj_weight_to_fp16, x = hidden_states_43_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<int32, [4]> var_553 = const()[name = tensor<string, []>("op_553"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_554_cast_fp16 = reshape(shape = var_553, x = linear_42_cast_fp16)[name = tensor<string, []>("op_554_cast_fp16")];
            tensor<int32, [4]> var_556 = const()[name = tensor<string, []>("op_556"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_557_cast_fp16 = reshape(shape = var_556, x = linear_43_cast_fp16)[name = tensor<string, []>("op_557_cast_fp16")];
            tensor<int32, [4]> var_559 = const()[name = tensor<string, []>("op_559"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_560_cast_fp16 = reshape(shape = var_559, x = linear_44_cast_fp16)[name = tensor<string, []>("op_560_cast_fp16")];
            tensor<int32, [4]> value_states_31_perm_0 = const()[name = tensor<string, []>("value_states_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_563_transpose_x_0 = const()[name = tensor<string, []>("op_563_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_563_transpose_y_0 = const()[name = tensor<string, []>("op_563_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_97_perm_0 = const()[name = tensor<string, []>("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_98_perm_0 = const()[name = tensor<string, []>("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_557_cast_fp16)[name = tensor<string, []>("transpose_219")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_97 = transpose(perm = transpose_97_perm_0, x = var_554_cast_fp16)[name = tensor<string, []>("transpose_220")];
            tensor<fp16, [1, 16, 1024, 1024]> var_563_cast_fp16 = matmul(transpose_x = var_563_transpose_x_0, transpose_y = var_563_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor<string, []>("op_563_cast_fp16")];
            tensor<fp16, []> var_564_to_fp16 = const()[name = tensor<string, []>("op_564_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_29_cast_fp16 = mul(x = var_563_cast_fp16, y = var_564_to_fp16)[name = tensor<string, []>("attn_weights_29_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_566_cast_fp16 = softmax(axis = var_18, x = attn_weights_29_cast_fp16)[name = tensor<string, []>("op_566_cast_fp16")];
            tensor<bool, []> attn_output_29_transpose_x_0 = const()[name = tensor<string, []>("attn_output_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_29_transpose_y_0 = const()[name = tensor<string, []>("attn_output_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_31_cast_fp16 = transpose(perm = value_states_31_perm_0, x = var_560_cast_fp16)[name = tensor<string, []>("transpose_221")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = var_566_cast_fp16, y = value_states_31_cast_fp16)[name = tensor<string, []>("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_570_perm_0 = const()[name = tensor<string, []>("op_570_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_572 = const()[name = tensor<string, []>("op_572"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_570_cast_fp16 = transpose(perm = var_570_perm_0, x = attn_output_29_cast_fp16)[name = tensor<string, []>("transpose_218")];
            tensor<fp16, [1, 1024, 1152]> input_105_cast_fp16 = reshape(shape = var_572, x = var_570_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_7_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233982016)))];
            tensor<fp16, [1152]> vpm_encoder_layers_7_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236636288)))];
            tensor<fp16, [1, 1024, 1152]> linear_45_cast_fp16 = linear(bias = vpm_encoder_layers_7_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_7_self_attn_out_proj_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_107_cast_fp16 = add(x = input_101_cast_fp16, y = linear_45_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<int32, [1]> input_109_axes_0 = const()[name = tensor<string, []>("input_109_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_7_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236638656)))];
            tensor<fp16, [1152]> vpm_encoder_layers_7_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236641024)))];
            tensor<fp16, [1, 1024, 1152]> input_109_cast_fp16 = layer_norm(axes = input_109_axes_0, beta = vpm_encoder_layers_7_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_7_layer_norm2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_7_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236643392)))];
            tensor<fp16, [4304]> vpm_encoder_layers_7_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246559872)))];
            tensor<fp16, [1, 1024, 4304]> linear_46_cast_fp16 = linear(bias = vpm_encoder_layers_7_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_7_mlp_fc1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<string, []> input_113_mode_0 = const()[name = tensor<string, []>("input_113_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_113_cast_fp16 = gelu(mode = input_113_mode_0, x = linear_46_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_7_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246568576)))];
            tensor<fp16, [1152]> vpm_encoder_layers_7_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_7_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256485056)))];
            tensor<fp16, [1, 1024, 1152]> linear_47_cast_fp16 = linear(bias = vpm_encoder_layers_7_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_7_mlp_fc2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_115_cast_fp16 = add(x = input_107_cast_fp16, y = linear_47_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
            tensor<int32, [1]> hidden_states_49_axes_0 = const()[name = tensor<string, []>("hidden_states_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_8_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256487424)))];
            tensor<fp16, [1152]> vpm_encoder_layers_8_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256489792)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_49_cast_fp16 = layer_norm(axes = hidden_states_49_axes_0, beta = vpm_encoder_layers_8_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_8_layer_norm1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor<string, []>("hidden_states_49_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256492160)))];
            tensor<fp16, [1152]> vpm_encoder_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259146432)))];
            tensor<fp16, [1, 1024, 1152]> linear_48_cast_fp16 = linear(bias = vpm_encoder_layers_8_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_8_self_attn_q_proj_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259148800)))];
            tensor<fp16, [1152]> vpm_encoder_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261803072)))];
            tensor<fp16, [1, 1024, 1152]> linear_49_cast_fp16 = linear(bias = vpm_encoder_layers_8_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_8_self_attn_k_proj_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261805440)))];
            tensor<fp16, [1152]> vpm_encoder_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264459712)))];
            tensor<fp16, [1, 1024, 1152]> linear_50_cast_fp16 = linear(bias = vpm_encoder_layers_8_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_8_self_attn_v_proj_weight_to_fp16, x = hidden_states_49_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
            tensor<int32, [4]> var_615 = const()[name = tensor<string, []>("op_615"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_616_cast_fp16 = reshape(shape = var_615, x = linear_48_cast_fp16)[name = tensor<string, []>("op_616_cast_fp16")];
            tensor<int32, [4]> var_618 = const()[name = tensor<string, []>("op_618"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_619_cast_fp16 = reshape(shape = var_618, x = linear_49_cast_fp16)[name = tensor<string, []>("op_619_cast_fp16")];
            tensor<int32, [4]> var_621 = const()[name = tensor<string, []>("op_621"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_622_cast_fp16 = reshape(shape = var_621, x = linear_50_cast_fp16)[name = tensor<string, []>("op_622_cast_fp16")];
            tensor<int32, [4]> value_states_35_perm_0 = const()[name = tensor<string, []>("value_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_625_transpose_x_0 = const()[name = tensor<string, []>("op_625_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_625_transpose_y_0 = const()[name = tensor<string, []>("op_625_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_100_perm_0 = const()[name = tensor<string, []>("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_619_cast_fp16)[name = tensor<string, []>("transpose_215")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_99 = transpose(perm = transpose_99_perm_0, x = var_616_cast_fp16)[name = tensor<string, []>("transpose_216")];
            tensor<fp16, [1, 16, 1024, 1024]> var_625_cast_fp16 = matmul(transpose_x = var_625_transpose_x_0, transpose_y = var_625_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor<string, []>("op_625_cast_fp16")];
            tensor<fp16, []> var_626_to_fp16 = const()[name = tensor<string, []>("op_626_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_33_cast_fp16 = mul(x = var_625_cast_fp16, y = var_626_to_fp16)[name = tensor<string, []>("attn_weights_33_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_628_cast_fp16 = softmax(axis = var_18, x = attn_weights_33_cast_fp16)[name = tensor<string, []>("op_628_cast_fp16")];
            tensor<bool, []> attn_output_33_transpose_x_0 = const()[name = tensor<string, []>("attn_output_33_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_33_transpose_y_0 = const()[name = tensor<string, []>("attn_output_33_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_35_cast_fp16 = transpose(perm = value_states_35_perm_0, x = var_622_cast_fp16)[name = tensor<string, []>("transpose_217")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = var_628_cast_fp16, y = value_states_35_cast_fp16)[name = tensor<string, []>("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_632_perm_0 = const()[name = tensor<string, []>("op_632_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_634 = const()[name = tensor<string, []>("op_634"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_632_cast_fp16 = transpose(perm = var_632_perm_0, x = attn_output_33_cast_fp16)[name = tensor<string, []>("transpose_214")];
            tensor<fp16, [1, 1024, 1152]> input_119_cast_fp16 = reshape(shape = var_634, x = var_632_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_8_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264462080)))];
            tensor<fp16, [1152]> vpm_encoder_layers_8_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267116352)))];
            tensor<fp16, [1, 1024, 1152]> linear_51_cast_fp16 = linear(bias = vpm_encoder_layers_8_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_8_self_attn_out_proj_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("linear_51_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_51_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
            tensor<int32, [1]> input_123_axes_0 = const()[name = tensor<string, []>("input_123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_8_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267118720)))];
            tensor<fp16, [1152]> vpm_encoder_layers_8_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267121088)))];
            tensor<fp16, [1, 1024, 1152]> input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = vpm_encoder_layers_8_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_8_layer_norm2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_8_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267123456)))];
            tensor<fp16, [4304]> vpm_encoder_layers_8_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277039936)))];
            tensor<fp16, [1, 1024, 4304]> linear_52_cast_fp16 = linear(bias = vpm_encoder_layers_8_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_8_mlp_fc1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
            tensor<string, []> input_127_mode_0 = const()[name = tensor<string, []>("input_127_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = linear_52_cast_fp16)[name = tensor<string, []>("input_127_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_8_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277048640)))];
            tensor<fp16, [1152]> vpm_encoder_layers_8_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_8_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286965120)))];
            tensor<fp16, [1, 1024, 1152]> linear_53_cast_fp16 = linear(bias = vpm_encoder_layers_8_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_8_mlp_fc2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_129_cast_fp16 = add(x = input_121_cast_fp16, y = linear_53_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
            tensor<int32, [1]> hidden_states_55_axes_0 = const()[name = tensor<string, []>("hidden_states_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_9_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286967488)))];
            tensor<fp16, [1152]> vpm_encoder_layers_9_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286969856)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_55_cast_fp16 = layer_norm(axes = hidden_states_55_axes_0, beta = vpm_encoder_layers_9_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_9_layer_norm1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("hidden_states_55_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286972224)))];
            tensor<fp16, [1152]> vpm_encoder_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289626496)))];
            tensor<fp16, [1, 1024, 1152]> linear_54_cast_fp16 = linear(bias = vpm_encoder_layers_9_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_9_self_attn_q_proj_weight_to_fp16, x = hidden_states_55_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289628864)))];
            tensor<fp16, [1152]> vpm_encoder_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292283136)))];
            tensor<fp16, [1, 1024, 1152]> linear_55_cast_fp16 = linear(bias = vpm_encoder_layers_9_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_9_self_attn_k_proj_weight_to_fp16, x = hidden_states_55_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292285504)))];
            tensor<fp16, [1152]> vpm_encoder_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294939776)))];
            tensor<fp16, [1, 1024, 1152]> linear_56_cast_fp16 = linear(bias = vpm_encoder_layers_9_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_9_self_attn_v_proj_weight_to_fp16, x = hidden_states_55_cast_fp16)[name = tensor<string, []>("linear_56_cast_fp16")];
            tensor<int32, [4]> var_677 = const()[name = tensor<string, []>("op_677"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_678_cast_fp16 = reshape(shape = var_677, x = linear_54_cast_fp16)[name = tensor<string, []>("op_678_cast_fp16")];
            tensor<int32, [4]> var_680 = const()[name = tensor<string, []>("op_680"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_681_cast_fp16 = reshape(shape = var_680, x = linear_55_cast_fp16)[name = tensor<string, []>("op_681_cast_fp16")];
            tensor<int32, [4]> var_683 = const()[name = tensor<string, []>("op_683"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_684_cast_fp16 = reshape(shape = var_683, x = linear_56_cast_fp16)[name = tensor<string, []>("op_684_cast_fp16")];
            tensor<int32, [4]> value_states_39_perm_0 = const()[name = tensor<string, []>("value_states_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_687_transpose_x_0 = const()[name = tensor<string, []>("op_687_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_687_transpose_y_0 = const()[name = tensor<string, []>("op_687_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_102_perm_0 = const()[name = tensor<string, []>("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_681_cast_fp16)[name = tensor<string, []>("transpose_211")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_101 = transpose(perm = transpose_101_perm_0, x = var_678_cast_fp16)[name = tensor<string, []>("transpose_212")];
            tensor<fp16, [1, 16, 1024, 1024]> var_687_cast_fp16 = matmul(transpose_x = var_687_transpose_x_0, transpose_y = var_687_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor<string, []>("op_687_cast_fp16")];
            tensor<fp16, []> var_688_to_fp16 = const()[name = tensor<string, []>("op_688_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_37_cast_fp16 = mul(x = var_687_cast_fp16, y = var_688_to_fp16)[name = tensor<string, []>("attn_weights_37_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_690_cast_fp16 = softmax(axis = var_18, x = attn_weights_37_cast_fp16)[name = tensor<string, []>("op_690_cast_fp16")];
            tensor<bool, []> attn_output_37_transpose_x_0 = const()[name = tensor<string, []>("attn_output_37_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_37_transpose_y_0 = const()[name = tensor<string, []>("attn_output_37_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_39_cast_fp16 = transpose(perm = value_states_39_perm_0, x = var_684_cast_fp16)[name = tensor<string, []>("transpose_213")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = var_690_cast_fp16, y = value_states_39_cast_fp16)[name = tensor<string, []>("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_694_perm_0 = const()[name = tensor<string, []>("op_694_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_696 = const()[name = tensor<string, []>("op_696"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_694_cast_fp16 = transpose(perm = var_694_perm_0, x = attn_output_37_cast_fp16)[name = tensor<string, []>("transpose_210")];
            tensor<fp16, [1, 1024, 1152]> input_133_cast_fp16 = reshape(shape = var_696, x = var_694_cast_fp16)[name = tensor<string, []>("input_133_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_9_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(294942144)))];
            tensor<fp16, [1152]> vpm_encoder_layers_9_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297596416)))];
            tensor<fp16, [1, 1024, 1152]> linear_57_cast_fp16 = linear(bias = vpm_encoder_layers_9_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_9_self_attn_out_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_57_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
            tensor<int32, [1]> input_137_axes_0 = const()[name = tensor<string, []>("input_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_9_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297598784)))];
            tensor<fp16, [1152]> vpm_encoder_layers_9_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297601152)))];
            tensor<fp16, [1, 1024, 1152]> input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = vpm_encoder_layers_9_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_9_layer_norm2_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("input_137_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_9_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297603520)))];
            tensor<fp16, [4304]> vpm_encoder_layers_9_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307520000)))];
            tensor<fp16, [1, 1024, 4304]> linear_58_cast_fp16 = linear(bias = vpm_encoder_layers_9_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_9_mlp_fc1_weight_to_fp16, x = input_137_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")];
            tensor<string, []> input_141_mode_0 = const()[name = tensor<string, []>("input_141_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = linear_58_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_9_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307528704)))];
            tensor<fp16, [1152]> vpm_encoder_layers_9_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_9_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317445184)))];
            tensor<fp16, [1, 1024, 1152]> linear_59_cast_fp16 = linear(bias = vpm_encoder_layers_9_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_9_mlp_fc2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_143_cast_fp16 = add(x = input_135_cast_fp16, y = linear_59_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
            tensor<int32, [1]> hidden_states_61_axes_0 = const()[name = tensor<string, []>("hidden_states_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_10_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317447552)))];
            tensor<fp16, [1152]> vpm_encoder_layers_10_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317449920)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_61_cast_fp16 = layer_norm(axes = hidden_states_61_axes_0, beta = vpm_encoder_layers_10_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_10_layer_norm1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("hidden_states_61_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317452288)))];
            tensor<fp16, [1152]> vpm_encoder_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320106560)))];
            tensor<fp16, [1, 1024, 1152]> linear_60_cast_fp16 = linear(bias = vpm_encoder_layers_10_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_10_self_attn_q_proj_weight_to_fp16, x = hidden_states_61_cast_fp16)[name = tensor<string, []>("linear_60_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320108928)))];
            tensor<fp16, [1152]> vpm_encoder_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322763200)))];
            tensor<fp16, [1, 1024, 1152]> linear_61_cast_fp16 = linear(bias = vpm_encoder_layers_10_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_10_self_attn_k_proj_weight_to_fp16, x = hidden_states_61_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322765568)))];
            tensor<fp16, [1152]> vpm_encoder_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(325419840)))];
            tensor<fp16, [1, 1024, 1152]> linear_62_cast_fp16 = linear(bias = vpm_encoder_layers_10_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_10_self_attn_v_proj_weight_to_fp16, x = hidden_states_61_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")];
            tensor<int32, [4]> var_739 = const()[name = tensor<string, []>("op_739"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_740_cast_fp16 = reshape(shape = var_739, x = linear_60_cast_fp16)[name = tensor<string, []>("op_740_cast_fp16")];
            tensor<int32, [4]> var_742 = const()[name = tensor<string, []>("op_742"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_743_cast_fp16 = reshape(shape = var_742, x = linear_61_cast_fp16)[name = tensor<string, []>("op_743_cast_fp16")];
            tensor<int32, [4]> var_745 = const()[name = tensor<string, []>("op_745"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_746_cast_fp16 = reshape(shape = var_745, x = linear_62_cast_fp16)[name = tensor<string, []>("op_746_cast_fp16")];
            tensor<int32, [4]> value_states_43_perm_0 = const()[name = tensor<string, []>("value_states_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_749_transpose_x_0 = const()[name = tensor<string, []>("op_749_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_749_transpose_y_0 = const()[name = tensor<string, []>("op_749_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_104_perm_0 = const()[name = tensor<string, []>("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_743_cast_fp16)[name = tensor<string, []>("transpose_207")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_103 = transpose(perm = transpose_103_perm_0, x = var_740_cast_fp16)[name = tensor<string, []>("transpose_208")];
            tensor<fp16, [1, 16, 1024, 1024]> var_749_cast_fp16 = matmul(transpose_x = var_749_transpose_x_0, transpose_y = var_749_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor<string, []>("op_749_cast_fp16")];
            tensor<fp16, []> var_750_to_fp16 = const()[name = tensor<string, []>("op_750_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_41_cast_fp16 = mul(x = var_749_cast_fp16, y = var_750_to_fp16)[name = tensor<string, []>("attn_weights_41_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_752_cast_fp16 = softmax(axis = var_18, x = attn_weights_41_cast_fp16)[name = tensor<string, []>("op_752_cast_fp16")];
            tensor<bool, []> attn_output_41_transpose_x_0 = const()[name = tensor<string, []>("attn_output_41_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_41_transpose_y_0 = const()[name = tensor<string, []>("attn_output_41_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_43_cast_fp16 = transpose(perm = value_states_43_perm_0, x = var_746_cast_fp16)[name = tensor<string, []>("transpose_209")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = var_752_cast_fp16, y = value_states_43_cast_fp16)[name = tensor<string, []>("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_756_perm_0 = const()[name = tensor<string, []>("op_756_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_758 = const()[name = tensor<string, []>("op_758"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_756_cast_fp16 = transpose(perm = var_756_perm_0, x = attn_output_41_cast_fp16)[name = tensor<string, []>("transpose_206")];
            tensor<fp16, [1, 1024, 1152]> input_147_cast_fp16 = reshape(shape = var_758, x = var_756_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_10_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(325422208)))];
            tensor<fp16, [1152]> vpm_encoder_layers_10_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328076480)))];
            tensor<fp16, [1, 1024, 1152]> linear_63_cast_fp16 = linear(bias = vpm_encoder_layers_10_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_10_self_attn_out_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_149_cast_fp16 = add(x = input_143_cast_fp16, y = linear_63_cast_fp16)[name = tensor<string, []>("input_149_cast_fp16")];
            tensor<int32, [1]> input_151_axes_0 = const()[name = tensor<string, []>("input_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_10_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328078848)))];
            tensor<fp16, [1152]> vpm_encoder_layers_10_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328081216)))];
            tensor<fp16, [1, 1024, 1152]> input_151_cast_fp16 = layer_norm(axes = input_151_axes_0, beta = vpm_encoder_layers_10_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_10_layer_norm2_weight_to_fp16, x = input_149_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_10_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328083584)))];
            tensor<fp16, [4304]> vpm_encoder_layers_10_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338000064)))];
            tensor<fp16, [1, 1024, 4304]> linear_64_cast_fp16 = linear(bias = vpm_encoder_layers_10_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_10_mlp_fc1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")];
            tensor<string, []> input_155_mode_0 = const()[name = tensor<string, []>("input_155_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_64_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_10_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338008768)))];
            tensor<fp16, [1152]> vpm_encoder_layers_10_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_10_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347925248)))];
            tensor<fp16, [1, 1024, 1152]> linear_65_cast_fp16 = linear(bias = vpm_encoder_layers_10_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_10_mlp_fc2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_157_cast_fp16 = add(x = input_149_cast_fp16, y = linear_65_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
            tensor<int32, [1]> hidden_states_67_axes_0 = const()[name = tensor<string, []>("hidden_states_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_11_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347927616)))];
            tensor<fp16, [1152]> vpm_encoder_layers_11_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347929984)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_67_cast_fp16 = layer_norm(axes = hidden_states_67_axes_0, beta = vpm_encoder_layers_11_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_11_layer_norm1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor<string, []>("hidden_states_67_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347932352)))];
            tensor<fp16, [1152]> vpm_encoder_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350586624)))];
            tensor<fp16, [1, 1024, 1152]> linear_66_cast_fp16 = linear(bias = vpm_encoder_layers_11_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_11_self_attn_q_proj_weight_to_fp16, x = hidden_states_67_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350588992)))];
            tensor<fp16, [1152]> vpm_encoder_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353243264)))];
            tensor<fp16, [1, 1024, 1152]> linear_67_cast_fp16 = linear(bias = vpm_encoder_layers_11_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_11_self_attn_k_proj_weight_to_fp16, x = hidden_states_67_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353245632)))];
            tensor<fp16, [1152]> vpm_encoder_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355899904)))];
            tensor<fp16, [1, 1024, 1152]> linear_68_cast_fp16 = linear(bias = vpm_encoder_layers_11_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_11_self_attn_v_proj_weight_to_fp16, x = hidden_states_67_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")];
            tensor<int32, [4]> var_801 = const()[name = tensor<string, []>("op_801"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_802_cast_fp16 = reshape(shape = var_801, x = linear_66_cast_fp16)[name = tensor<string, []>("op_802_cast_fp16")];
            tensor<int32, [4]> var_804 = const()[name = tensor<string, []>("op_804"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_805_cast_fp16 = reshape(shape = var_804, x = linear_67_cast_fp16)[name = tensor<string, []>("op_805_cast_fp16")];
            tensor<int32, [4]> var_807 = const()[name = tensor<string, []>("op_807"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_808_cast_fp16 = reshape(shape = var_807, x = linear_68_cast_fp16)[name = tensor<string, []>("op_808_cast_fp16")];
            tensor<int32, [4]> value_states_47_perm_0 = const()[name = tensor<string, []>("value_states_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_811_transpose_x_0 = const()[name = tensor<string, []>("op_811_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_811_transpose_y_0 = const()[name = tensor<string, []>("op_811_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_105_perm_0 = const()[name = tensor<string, []>("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_106_perm_0 = const()[name = tensor<string, []>("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_805_cast_fp16)[name = tensor<string, []>("transpose_203")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_105 = transpose(perm = transpose_105_perm_0, x = var_802_cast_fp16)[name = tensor<string, []>("transpose_204")];
            tensor<fp16, [1, 16, 1024, 1024]> var_811_cast_fp16 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor<string, []>("op_811_cast_fp16")];
            tensor<fp16, []> var_812_to_fp16 = const()[name = tensor<string, []>("op_812_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_45_cast_fp16 = mul(x = var_811_cast_fp16, y = var_812_to_fp16)[name = tensor<string, []>("attn_weights_45_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_814_cast_fp16 = softmax(axis = var_18, x = attn_weights_45_cast_fp16)[name = tensor<string, []>("op_814_cast_fp16")];
            tensor<bool, []> attn_output_45_transpose_x_0 = const()[name = tensor<string, []>("attn_output_45_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_45_transpose_y_0 = const()[name = tensor<string, []>("attn_output_45_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_47_cast_fp16 = transpose(perm = value_states_47_perm_0, x = var_808_cast_fp16)[name = tensor<string, []>("transpose_205")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = var_814_cast_fp16, y = value_states_47_cast_fp16)[name = tensor<string, []>("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_818_perm_0 = const()[name = tensor<string, []>("op_818_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_820 = const()[name = tensor<string, []>("op_820"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_818_cast_fp16 = transpose(perm = var_818_perm_0, x = attn_output_45_cast_fp16)[name = tensor<string, []>("transpose_202")];
            tensor<fp16, [1, 1024, 1152]> input_161_cast_fp16 = reshape(shape = var_820, x = var_818_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_11_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355902272)))];
            tensor<fp16, [1152]> vpm_encoder_layers_11_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358556544)))];
            tensor<fp16, [1, 1024, 1152]> linear_69_cast_fp16 = linear(bias = vpm_encoder_layers_11_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_11_self_attn_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("linear_69_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_69_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
            tensor<int32, [1]> input_165_axes_0 = const()[name = tensor<string, []>("input_165_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_11_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358558912)))];
            tensor<fp16, [1152]> vpm_encoder_layers_11_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358561280)))];
            tensor<fp16, [1, 1024, 1152]> input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = vpm_encoder_layers_11_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_11_layer_norm2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("input_165_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_11_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358563648)))];
            tensor<fp16, [4304]> vpm_encoder_layers_11_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368480128)))];
            tensor<fp16, [1, 1024, 4304]> linear_70_cast_fp16 = linear(bias = vpm_encoder_layers_11_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_11_mlp_fc1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")];
            tensor<string, []> input_169_mode_0 = const()[name = tensor<string, []>("input_169_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_70_cast_fp16)[name = tensor<string, []>("input_169_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_11_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368488832)))];
            tensor<fp16, [1152]> vpm_encoder_layers_11_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_11_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378405312)))];
            tensor<fp16, [1, 1024, 1152]> linear_71_cast_fp16 = linear(bias = vpm_encoder_layers_11_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_11_mlp_fc2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_171_cast_fp16 = add(x = input_163_cast_fp16, y = linear_71_cast_fp16)[name = tensor<string, []>("input_171_cast_fp16")];
            tensor<int32, [1]> hidden_states_73_axes_0 = const()[name = tensor<string, []>("hidden_states_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_12_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378407680)))];
            tensor<fp16, [1152]> vpm_encoder_layers_12_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378410048)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_73_cast_fp16 = layer_norm(axes = hidden_states_73_axes_0, beta = vpm_encoder_layers_12_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_12_layer_norm1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor<string, []>("hidden_states_73_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378412416)))];
            tensor<fp16, [1152]> vpm_encoder_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381066688)))];
            tensor<fp16, [1, 1024, 1152]> linear_72_cast_fp16 = linear(bias = vpm_encoder_layers_12_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_12_self_attn_q_proj_weight_to_fp16, x = hidden_states_73_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381069056)))];
            tensor<fp16, [1152]> vpm_encoder_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383723328)))];
            tensor<fp16, [1, 1024, 1152]> linear_73_cast_fp16 = linear(bias = vpm_encoder_layers_12_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_12_self_attn_k_proj_weight_to_fp16, x = hidden_states_73_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383725696)))];
            tensor<fp16, [1152]> vpm_encoder_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386379968)))];
            tensor<fp16, [1, 1024, 1152]> linear_74_cast_fp16 = linear(bias = vpm_encoder_layers_12_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_12_self_attn_v_proj_weight_to_fp16, x = hidden_states_73_cast_fp16)[name = tensor<string, []>("linear_74_cast_fp16")];
            tensor<int32, [4]> var_863 = const()[name = tensor<string, []>("op_863"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_864_cast_fp16 = reshape(shape = var_863, x = linear_72_cast_fp16)[name = tensor<string, []>("op_864_cast_fp16")];
            tensor<int32, [4]> var_866 = const()[name = tensor<string, []>("op_866"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_867_cast_fp16 = reshape(shape = var_866, x = linear_73_cast_fp16)[name = tensor<string, []>("op_867_cast_fp16")];
            tensor<int32, [4]> var_869 = const()[name = tensor<string, []>("op_869"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_870_cast_fp16 = reshape(shape = var_869, x = linear_74_cast_fp16)[name = tensor<string, []>("op_870_cast_fp16")];
            tensor<int32, [4]> value_states_51_perm_0 = const()[name = tensor<string, []>("value_states_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_873_transpose_x_0 = const()[name = tensor<string, []>("op_873_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_873_transpose_y_0 = const()[name = tensor<string, []>("op_873_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_108_perm_0 = const()[name = tensor<string, []>("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_867_cast_fp16)[name = tensor<string, []>("transpose_199")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_107 = transpose(perm = transpose_107_perm_0, x = var_864_cast_fp16)[name = tensor<string, []>("transpose_200")];
            tensor<fp16, [1, 16, 1024, 1024]> var_873_cast_fp16 = matmul(transpose_x = var_873_transpose_x_0, transpose_y = var_873_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor<string, []>("op_873_cast_fp16")];
            tensor<fp16, []> var_874_to_fp16 = const()[name = tensor<string, []>("op_874_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_49_cast_fp16 = mul(x = var_873_cast_fp16, y = var_874_to_fp16)[name = tensor<string, []>("attn_weights_49_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_876_cast_fp16 = softmax(axis = var_18, x = attn_weights_49_cast_fp16)[name = tensor<string, []>("op_876_cast_fp16")];
            tensor<bool, []> attn_output_49_transpose_x_0 = const()[name = tensor<string, []>("attn_output_49_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_49_transpose_y_0 = const()[name = tensor<string, []>("attn_output_49_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_51_cast_fp16 = transpose(perm = value_states_51_perm_0, x = var_870_cast_fp16)[name = tensor<string, []>("transpose_201")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = var_876_cast_fp16, y = value_states_51_cast_fp16)[name = tensor<string, []>("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_880_perm_0 = const()[name = tensor<string, []>("op_880_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_882 = const()[name = tensor<string, []>("op_882"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_880_cast_fp16 = transpose(perm = var_880_perm_0, x = attn_output_49_cast_fp16)[name = tensor<string, []>("transpose_198")];
            tensor<fp16, [1, 1024, 1152]> input_175_cast_fp16 = reshape(shape = var_882, x = var_880_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_12_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386382336)))];
            tensor<fp16, [1152]> vpm_encoder_layers_12_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389036608)))];
            tensor<fp16, [1, 1024, 1152]> linear_75_cast_fp16 = linear(bias = vpm_encoder_layers_12_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_12_self_attn_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor<string, []>("linear_75_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_75_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
            tensor<int32, [1]> input_179_axes_0 = const()[name = tensor<string, []>("input_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_12_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389038976)))];
            tensor<fp16, [1152]> vpm_encoder_layers_12_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389041344)))];
            tensor<fp16, [1, 1024, 1152]> input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = vpm_encoder_layers_12_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_12_layer_norm2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor<string, []>("input_179_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_12_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(389043712)))];
            tensor<fp16, [4304]> vpm_encoder_layers_12_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(398960192)))];
            tensor<fp16, [1, 1024, 4304]> linear_76_cast_fp16 = linear(bias = vpm_encoder_layers_12_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_12_mlp_fc1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor<string, []>("linear_76_cast_fp16")];
            tensor<string, []> input_183_mode_0 = const()[name = tensor<string, []>("input_183_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_183_cast_fp16 = gelu(mode = input_183_mode_0, x = linear_76_cast_fp16)[name = tensor<string, []>("input_183_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_12_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(398968896)))];
            tensor<fp16, [1152]> vpm_encoder_layers_12_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_12_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408885376)))];
            tensor<fp16, [1, 1024, 1152]> linear_77_cast_fp16 = linear(bias = vpm_encoder_layers_12_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_12_mlp_fc2_weight_to_fp16, x = input_183_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_185_cast_fp16 = add(x = input_177_cast_fp16, y = linear_77_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")];
            tensor<int32, [1]> hidden_states_79_axes_0 = const()[name = tensor<string, []>("hidden_states_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_13_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408887744)))];
            tensor<fp16, [1152]> vpm_encoder_layers_13_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408890112)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_79_cast_fp16 = layer_norm(axes = hidden_states_79_axes_0, beta = vpm_encoder_layers_13_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_13_layer_norm1_weight_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("hidden_states_79_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408892480)))];
            tensor<fp16, [1152]> vpm_encoder_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411546752)))];
            tensor<fp16, [1, 1024, 1152]> linear_78_cast_fp16 = linear(bias = vpm_encoder_layers_13_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_13_self_attn_q_proj_weight_to_fp16, x = hidden_states_79_cast_fp16)[name = tensor<string, []>("linear_78_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411549120)))];
            tensor<fp16, [1152]> vpm_encoder_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414203392)))];
            tensor<fp16, [1, 1024, 1152]> linear_79_cast_fp16 = linear(bias = vpm_encoder_layers_13_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_13_self_attn_k_proj_weight_to_fp16, x = hidden_states_79_cast_fp16)[name = tensor<string, []>("linear_79_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(414205760)))];
            tensor<fp16, [1152]> vpm_encoder_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(416860032)))];
            tensor<fp16, [1, 1024, 1152]> linear_80_cast_fp16 = linear(bias = vpm_encoder_layers_13_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_13_self_attn_v_proj_weight_to_fp16, x = hidden_states_79_cast_fp16)[name = tensor<string, []>("linear_80_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = tensor<string, []>("op_925"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_78_cast_fp16)[name = tensor<string, []>("op_926_cast_fp16")];
            tensor<int32, [4]> var_928 = const()[name = tensor<string, []>("op_928"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_929_cast_fp16 = reshape(shape = var_928, x = linear_79_cast_fp16)[name = tensor<string, []>("op_929_cast_fp16")];
            tensor<int32, [4]> var_931 = const()[name = tensor<string, []>("op_931"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_932_cast_fp16 = reshape(shape = var_931, x = linear_80_cast_fp16)[name = tensor<string, []>("op_932_cast_fp16")];
            tensor<int32, [4]> value_states_55_perm_0 = const()[name = tensor<string, []>("value_states_55_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_935_transpose_x_0 = const()[name = tensor<string, []>("op_935_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_935_transpose_y_0 = const()[name = tensor<string, []>("op_935_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_110_perm_0 = const()[name = tensor<string, []>("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_929_cast_fp16)[name = tensor<string, []>("transpose_195")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_109 = transpose(perm = transpose_109_perm_0, x = var_926_cast_fp16)[name = tensor<string, []>("transpose_196")];
            tensor<fp16, [1, 16, 1024, 1024]> var_935_cast_fp16 = matmul(transpose_x = var_935_transpose_x_0, transpose_y = var_935_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor<string, []>("op_935_cast_fp16")];
            tensor<fp16, []> var_936_to_fp16 = const()[name = tensor<string, []>("op_936_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_53_cast_fp16 = mul(x = var_935_cast_fp16, y = var_936_to_fp16)[name = tensor<string, []>("attn_weights_53_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_938_cast_fp16 = softmax(axis = var_18, x = attn_weights_53_cast_fp16)[name = tensor<string, []>("op_938_cast_fp16")];
            tensor<bool, []> attn_output_53_transpose_x_0 = const()[name = tensor<string, []>("attn_output_53_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_53_transpose_y_0 = const()[name = tensor<string, []>("attn_output_53_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_55_cast_fp16 = transpose(perm = value_states_55_perm_0, x = var_932_cast_fp16)[name = tensor<string, []>("transpose_197")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = var_938_cast_fp16, y = value_states_55_cast_fp16)[name = tensor<string, []>("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_942_perm_0 = const()[name = tensor<string, []>("op_942_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_944 = const()[name = tensor<string, []>("op_944"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_942_cast_fp16 = transpose(perm = var_942_perm_0, x = attn_output_53_cast_fp16)[name = tensor<string, []>("transpose_194")];
            tensor<fp16, [1, 1024, 1152]> input_189_cast_fp16 = reshape(shape = var_944, x = var_942_cast_fp16)[name = tensor<string, []>("input_189_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_13_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(416862400)))];
            tensor<fp16, [1152]> vpm_encoder_layers_13_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419516672)))];
            tensor<fp16, [1, 1024, 1152]> linear_81_cast_fp16 = linear(bias = vpm_encoder_layers_13_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_13_self_attn_out_proj_weight_to_fp16, x = input_189_cast_fp16)[name = tensor<string, []>("linear_81_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_191_cast_fp16 = add(x = input_185_cast_fp16, y = linear_81_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
            tensor<int32, [1]> input_193_axes_0 = const()[name = tensor<string, []>("input_193_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_13_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419519040)))];
            tensor<fp16, [1152]> vpm_encoder_layers_13_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419521408)))];
            tensor<fp16, [1, 1024, 1152]> input_193_cast_fp16 = layer_norm(axes = input_193_axes_0, beta = vpm_encoder_layers_13_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_13_layer_norm2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor<string, []>("input_193_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_13_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419523776)))];
            tensor<fp16, [4304]> vpm_encoder_layers_13_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429440256)))];
            tensor<fp16, [1, 1024, 4304]> linear_82_cast_fp16 = linear(bias = vpm_encoder_layers_13_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_13_mlp_fc1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor<string, []>("linear_82_cast_fp16")];
            tensor<string, []> input_197_mode_0 = const()[name = tensor<string, []>("input_197_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_197_cast_fp16 = gelu(mode = input_197_mode_0, x = linear_82_cast_fp16)[name = tensor<string, []>("input_197_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_13_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429448960)))];
            tensor<fp16, [1152]> vpm_encoder_layers_13_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_13_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439365440)))];
            tensor<fp16, [1, 1024, 1152]> linear_83_cast_fp16 = linear(bias = vpm_encoder_layers_13_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_13_mlp_fc2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor<string, []>("linear_83_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_199_cast_fp16 = add(x = input_191_cast_fp16, y = linear_83_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
            tensor<int32, [1]> hidden_states_85_axes_0 = const()[name = tensor<string, []>("hidden_states_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_14_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439367808)))];
            tensor<fp16, [1152]> vpm_encoder_layers_14_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439370176)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_85_cast_fp16 = layer_norm(axes = hidden_states_85_axes_0, beta = vpm_encoder_layers_14_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_14_layer_norm1_weight_to_fp16, x = input_199_cast_fp16)[name = tensor<string, []>("hidden_states_85_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439372544)))];
            tensor<fp16, [1152]> vpm_encoder_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442026816)))];
            tensor<fp16, [1, 1024, 1152]> linear_84_cast_fp16 = linear(bias = vpm_encoder_layers_14_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_14_self_attn_q_proj_weight_to_fp16, x = hidden_states_85_cast_fp16)[name = tensor<string, []>("linear_84_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442029184)))];
            tensor<fp16, [1152]> vpm_encoder_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444683456)))];
            tensor<fp16, [1, 1024, 1152]> linear_85_cast_fp16 = linear(bias = vpm_encoder_layers_14_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_14_self_attn_k_proj_weight_to_fp16, x = hidden_states_85_cast_fp16)[name = tensor<string, []>("linear_85_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444685824)))];
            tensor<fp16, [1152]> vpm_encoder_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447340096)))];
            tensor<fp16, [1, 1024, 1152]> linear_86_cast_fp16 = linear(bias = vpm_encoder_layers_14_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_14_self_attn_v_proj_weight_to_fp16, x = hidden_states_85_cast_fp16)[name = tensor<string, []>("linear_86_cast_fp16")];
            tensor<int32, [4]> var_987 = const()[name = tensor<string, []>("op_987"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_988_cast_fp16 = reshape(shape = var_987, x = linear_84_cast_fp16)[name = tensor<string, []>("op_988_cast_fp16")];
            tensor<int32, [4]> var_990 = const()[name = tensor<string, []>("op_990"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_991_cast_fp16 = reshape(shape = var_990, x = linear_85_cast_fp16)[name = tensor<string, []>("op_991_cast_fp16")];
            tensor<int32, [4]> var_993 = const()[name = tensor<string, []>("op_993"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_994_cast_fp16 = reshape(shape = var_993, x = linear_86_cast_fp16)[name = tensor<string, []>("op_994_cast_fp16")];
            tensor<int32, [4]> value_states_59_perm_0 = const()[name = tensor<string, []>("value_states_59_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_997_transpose_x_0 = const()[name = tensor<string, []>("op_997_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_997_transpose_y_0 = const()[name = tensor<string, []>("op_997_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_112_perm_0 = const()[name = tensor<string, []>("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_991_cast_fp16)[name = tensor<string, []>("transpose_191")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_111 = transpose(perm = transpose_111_perm_0, x = var_988_cast_fp16)[name = tensor<string, []>("transpose_192")];
            tensor<fp16, [1, 16, 1024, 1024]> var_997_cast_fp16 = matmul(transpose_x = var_997_transpose_x_0, transpose_y = var_997_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor<string, []>("op_997_cast_fp16")];
            tensor<fp16, []> var_998_to_fp16 = const()[name = tensor<string, []>("op_998_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_57_cast_fp16 = mul(x = var_997_cast_fp16, y = var_998_to_fp16)[name = tensor<string, []>("attn_weights_57_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1000_cast_fp16 = softmax(axis = var_18, x = attn_weights_57_cast_fp16)[name = tensor<string, []>("op_1000_cast_fp16")];
            tensor<bool, []> attn_output_57_transpose_x_0 = const()[name = tensor<string, []>("attn_output_57_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_57_transpose_y_0 = const()[name = tensor<string, []>("attn_output_57_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_59_cast_fp16 = transpose(perm = value_states_59_perm_0, x = var_994_cast_fp16)[name = tensor<string, []>("transpose_193")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = var_1000_cast_fp16, y = value_states_59_cast_fp16)[name = tensor<string, []>("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_1004_perm_0 = const()[name = tensor<string, []>("op_1004_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1006 = const()[name = tensor<string, []>("op_1006"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1004_cast_fp16 = transpose(perm = var_1004_perm_0, x = attn_output_57_cast_fp16)[name = tensor<string, []>("transpose_190")];
            tensor<fp16, [1, 1024, 1152]> input_203_cast_fp16 = reshape(shape = var_1006, x = var_1004_cast_fp16)[name = tensor<string, []>("input_203_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_14_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447342464)))];
            tensor<fp16, [1152]> vpm_encoder_layers_14_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449996736)))];
            tensor<fp16, [1, 1024, 1152]> linear_87_cast_fp16 = linear(bias = vpm_encoder_layers_14_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_14_self_attn_out_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor<string, []>("linear_87_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_87_cast_fp16)[name = tensor<string, []>("input_205_cast_fp16")];
            tensor<int32, [1]> input_207_axes_0 = const()[name = tensor<string, []>("input_207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_14_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449999104)))];
            tensor<fp16, [1152]> vpm_encoder_layers_14_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450001472)))];
            tensor<fp16, [1, 1024, 1152]> input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = vpm_encoder_layers_14_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_14_layer_norm2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_14_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450003840)))];
            tensor<fp16, [4304]> vpm_encoder_layers_14_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459920320)))];
            tensor<fp16, [1, 1024, 4304]> linear_88_cast_fp16 = linear(bias = vpm_encoder_layers_14_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_14_mlp_fc1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor<string, []>("linear_88_cast_fp16")];
            tensor<string, []> input_211_mode_0 = const()[name = tensor<string, []>("input_211_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_88_cast_fp16)[name = tensor<string, []>("input_211_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_14_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459929024)))];
            tensor<fp16, [1152]> vpm_encoder_layers_14_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_14_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(469845504)))];
            tensor<fp16, [1, 1024, 1152]> linear_89_cast_fp16 = linear(bias = vpm_encoder_layers_14_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_14_mlp_fc2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor<string, []>("linear_89_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_213_cast_fp16 = add(x = input_205_cast_fp16, y = linear_89_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
            tensor<int32, [1]> hidden_states_91_axes_0 = const()[name = tensor<string, []>("hidden_states_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_15_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(469847872)))];
            tensor<fp16, [1152]> vpm_encoder_layers_15_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(469850240)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_91_cast_fp16 = layer_norm(axes = hidden_states_91_axes_0, beta = vpm_encoder_layers_15_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_15_layer_norm1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor<string, []>("hidden_states_91_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(469852608)))];
            tensor<fp16, [1152]> vpm_encoder_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(472506880)))];
            tensor<fp16, [1, 1024, 1152]> linear_90_cast_fp16 = linear(bias = vpm_encoder_layers_15_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_15_self_attn_q_proj_weight_to_fp16, x = hidden_states_91_cast_fp16)[name = tensor<string, []>("linear_90_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(472509248)))];
            tensor<fp16, [1152]> vpm_encoder_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(475163520)))];
            tensor<fp16, [1, 1024, 1152]> linear_91_cast_fp16 = linear(bias = vpm_encoder_layers_15_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_15_self_attn_k_proj_weight_to_fp16, x = hidden_states_91_cast_fp16)[name = tensor<string, []>("linear_91_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(475165888)))];
            tensor<fp16, [1152]> vpm_encoder_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477820160)))];
            tensor<fp16, [1, 1024, 1152]> linear_92_cast_fp16 = linear(bias = vpm_encoder_layers_15_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_15_self_attn_v_proj_weight_to_fp16, x = hidden_states_91_cast_fp16)[name = tensor<string, []>("linear_92_cast_fp16")];
            tensor<int32, [4]> var_1049 = const()[name = tensor<string, []>("op_1049"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1050_cast_fp16 = reshape(shape = var_1049, x = linear_90_cast_fp16)[name = tensor<string, []>("op_1050_cast_fp16")];
            tensor<int32, [4]> var_1052 = const()[name = tensor<string, []>("op_1052"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1053_cast_fp16 = reshape(shape = var_1052, x = linear_91_cast_fp16)[name = tensor<string, []>("op_1053_cast_fp16")];
            tensor<int32, [4]> var_1055 = const()[name = tensor<string, []>("op_1055"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1056_cast_fp16 = reshape(shape = var_1055, x = linear_92_cast_fp16)[name = tensor<string, []>("op_1056_cast_fp16")];
            tensor<int32, [4]> value_states_63_perm_0 = const()[name = tensor<string, []>("value_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1059_transpose_x_0 = const()[name = tensor<string, []>("op_1059_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1059_transpose_y_0 = const()[name = tensor<string, []>("op_1059_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_113_perm_0 = const()[name = tensor<string, []>("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_114_perm_0 = const()[name = tensor<string, []>("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_1053_cast_fp16)[name = tensor<string, []>("transpose_187")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_113 = transpose(perm = transpose_113_perm_0, x = var_1050_cast_fp16)[name = tensor<string, []>("transpose_188")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1059_cast_fp16 = matmul(transpose_x = var_1059_transpose_x_0, transpose_y = var_1059_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor<string, []>("op_1059_cast_fp16")];
            tensor<fp16, []> var_1060_to_fp16 = const()[name = tensor<string, []>("op_1060_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_61_cast_fp16 = mul(x = var_1059_cast_fp16, y = var_1060_to_fp16)[name = tensor<string, []>("attn_weights_61_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1062_cast_fp16 = softmax(axis = var_18, x = attn_weights_61_cast_fp16)[name = tensor<string, []>("op_1062_cast_fp16")];
            tensor<bool, []> attn_output_61_transpose_x_0 = const()[name = tensor<string, []>("attn_output_61_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_61_transpose_y_0 = const()[name = tensor<string, []>("attn_output_61_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_63_cast_fp16 = transpose(perm = value_states_63_perm_0, x = var_1056_cast_fp16)[name = tensor<string, []>("transpose_189")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = var_1062_cast_fp16, y = value_states_63_cast_fp16)[name = tensor<string, []>("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_1066_perm_0 = const()[name = tensor<string, []>("op_1066_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1068 = const()[name = tensor<string, []>("op_1068"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1066_cast_fp16 = transpose(perm = var_1066_perm_0, x = attn_output_61_cast_fp16)[name = tensor<string, []>("transpose_186")];
            tensor<fp16, [1, 1024, 1152]> input_217_cast_fp16 = reshape(shape = var_1068, x = var_1066_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_15_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477822528)))];
            tensor<fp16, [1152]> vpm_encoder_layers_15_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480476800)))];
            tensor<fp16, [1, 1024, 1152]> linear_93_cast_fp16 = linear(bias = vpm_encoder_layers_15_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_15_self_attn_out_proj_weight_to_fp16, x = input_217_cast_fp16)[name = tensor<string, []>("linear_93_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_93_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
            tensor<int32, [1]> input_221_axes_0 = const()[name = tensor<string, []>("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_15_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480479168)))];
            tensor<fp16, [1152]> vpm_encoder_layers_15_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480481536)))];
            tensor<fp16, [1, 1024, 1152]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = vpm_encoder_layers_15_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_15_layer_norm2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor<string, []>("input_221_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_15_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480483904)))];
            tensor<fp16, [4304]> vpm_encoder_layers_15_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490400384)))];
            tensor<fp16, [1, 1024, 4304]> linear_94_cast_fp16 = linear(bias = vpm_encoder_layers_15_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_15_mlp_fc1_weight_to_fp16, x = input_221_cast_fp16)[name = tensor<string, []>("linear_94_cast_fp16")];
            tensor<string, []> input_225_mode_0 = const()[name = tensor<string, []>("input_225_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_225_cast_fp16 = gelu(mode = input_225_mode_0, x = linear_94_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_15_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490409088)))];
            tensor<fp16, [1152]> vpm_encoder_layers_15_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_15_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500325568)))];
            tensor<fp16, [1, 1024, 1152]> linear_95_cast_fp16 = linear(bias = vpm_encoder_layers_15_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_15_mlp_fc2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("linear_95_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_227_cast_fp16 = add(x = input_219_cast_fp16, y = linear_95_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
            tensor<int32, [1]> hidden_states_97_axes_0 = const()[name = tensor<string, []>("hidden_states_97_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_16_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500327936)))];
            tensor<fp16, [1152]> vpm_encoder_layers_16_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500330304)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_97_cast_fp16 = layer_norm(axes = hidden_states_97_axes_0, beta = vpm_encoder_layers_16_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_16_layer_norm1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor<string, []>("hidden_states_97_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500332672)))];
            tensor<fp16, [1152]> vpm_encoder_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502986944)))];
            tensor<fp16, [1, 1024, 1152]> linear_96_cast_fp16 = linear(bias = vpm_encoder_layers_16_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_16_self_attn_q_proj_weight_to_fp16, x = hidden_states_97_cast_fp16)[name = tensor<string, []>("linear_96_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502989312)))];
            tensor<fp16, [1152]> vpm_encoder_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(505643584)))];
            tensor<fp16, [1, 1024, 1152]> linear_97_cast_fp16 = linear(bias = vpm_encoder_layers_16_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_16_self_attn_k_proj_weight_to_fp16, x = hidden_states_97_cast_fp16)[name = tensor<string, []>("linear_97_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(505645952)))];
            tensor<fp16, [1152]> vpm_encoder_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508300224)))];
            tensor<fp16, [1, 1024, 1152]> linear_98_cast_fp16 = linear(bias = vpm_encoder_layers_16_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_16_self_attn_v_proj_weight_to_fp16, x = hidden_states_97_cast_fp16)[name = tensor<string, []>("linear_98_cast_fp16")];
            tensor<int32, [4]> var_1111 = const()[name = tensor<string, []>("op_1111"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1112_cast_fp16 = reshape(shape = var_1111, x = linear_96_cast_fp16)[name = tensor<string, []>("op_1112_cast_fp16")];
            tensor<int32, [4]> var_1114 = const()[name = tensor<string, []>("op_1114"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1115_cast_fp16 = reshape(shape = var_1114, x = linear_97_cast_fp16)[name = tensor<string, []>("op_1115_cast_fp16")];
            tensor<int32, [4]> var_1117 = const()[name = tensor<string, []>("op_1117"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1118_cast_fp16 = reshape(shape = var_1117, x = linear_98_cast_fp16)[name = tensor<string, []>("op_1118_cast_fp16")];
            tensor<int32, [4]> value_states_67_perm_0 = const()[name = tensor<string, []>("value_states_67_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1121_transpose_x_0 = const()[name = tensor<string, []>("op_1121_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1121_transpose_y_0 = const()[name = tensor<string, []>("op_1121_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_116_perm_0 = const()[name = tensor<string, []>("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_1115_cast_fp16)[name = tensor<string, []>("transpose_183")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_115 = transpose(perm = transpose_115_perm_0, x = var_1112_cast_fp16)[name = tensor<string, []>("transpose_184")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1121_cast_fp16 = matmul(transpose_x = var_1121_transpose_x_0, transpose_y = var_1121_transpose_y_0, x = transpose_115, y = transpose_116)[name = tensor<string, []>("op_1121_cast_fp16")];
            tensor<fp16, []> var_1122_to_fp16 = const()[name = tensor<string, []>("op_1122_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_65_cast_fp16 = mul(x = var_1121_cast_fp16, y = var_1122_to_fp16)[name = tensor<string, []>("attn_weights_65_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1124_cast_fp16 = softmax(axis = var_18, x = attn_weights_65_cast_fp16)[name = tensor<string, []>("op_1124_cast_fp16")];
            tensor<bool, []> attn_output_65_transpose_x_0 = const()[name = tensor<string, []>("attn_output_65_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_65_transpose_y_0 = const()[name = tensor<string, []>("attn_output_65_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_67_cast_fp16 = transpose(perm = value_states_67_perm_0, x = var_1118_cast_fp16)[name = tensor<string, []>("transpose_185")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = var_1124_cast_fp16, y = value_states_67_cast_fp16)[name = tensor<string, []>("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_1128_perm_0 = const()[name = tensor<string, []>("op_1128_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1130 = const()[name = tensor<string, []>("op_1130"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1128_cast_fp16 = transpose(perm = var_1128_perm_0, x = attn_output_65_cast_fp16)[name = tensor<string, []>("transpose_182")];
            tensor<fp16, [1, 1024, 1152]> input_231_cast_fp16 = reshape(shape = var_1130, x = var_1128_cast_fp16)[name = tensor<string, []>("input_231_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_16_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508302592)))];
            tensor<fp16, [1152]> vpm_encoder_layers_16_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510956864)))];
            tensor<fp16, [1, 1024, 1152]> linear_99_cast_fp16 = linear(bias = vpm_encoder_layers_16_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_16_self_attn_out_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor<string, []>("linear_99_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_233_cast_fp16 = add(x = input_227_cast_fp16, y = linear_99_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
            tensor<int32, [1]> input_235_axes_0 = const()[name = tensor<string, []>("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_16_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510959232)))];
            tensor<fp16, [1152]> vpm_encoder_layers_16_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510961600)))];
            tensor<fp16, [1, 1024, 1152]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = vpm_encoder_layers_16_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_16_layer_norm2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor<string, []>("input_235_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_16_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510963968)))];
            tensor<fp16, [4304]> vpm_encoder_layers_16_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520880448)))];
            tensor<fp16, [1, 1024, 4304]> linear_100_cast_fp16 = linear(bias = vpm_encoder_layers_16_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_16_mlp_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor<string, []>("linear_100_cast_fp16")];
            tensor<string, []> input_239_mode_0 = const()[name = tensor<string, []>("input_239_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_100_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_16_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520889152)))];
            tensor<fp16, [1152]> vpm_encoder_layers_16_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_16_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530805632)))];
            tensor<fp16, [1, 1024, 1152]> linear_101_cast_fp16 = linear(bias = vpm_encoder_layers_16_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_16_mlp_fc2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor<string, []>("linear_101_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_241_cast_fp16 = add(x = input_233_cast_fp16, y = linear_101_cast_fp16)[name = tensor<string, []>("input_241_cast_fp16")];
            tensor<int32, [1]> hidden_states_103_axes_0 = const()[name = tensor<string, []>("hidden_states_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_17_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530808000)))];
            tensor<fp16, [1152]> vpm_encoder_layers_17_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530810368)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_103_cast_fp16 = layer_norm(axes = hidden_states_103_axes_0, beta = vpm_encoder_layers_17_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_17_layer_norm1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor<string, []>("hidden_states_103_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530812736)))];
            tensor<fp16, [1152]> vpm_encoder_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533467008)))];
            tensor<fp16, [1, 1024, 1152]> linear_102_cast_fp16 = linear(bias = vpm_encoder_layers_17_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_17_self_attn_q_proj_weight_to_fp16, x = hidden_states_103_cast_fp16)[name = tensor<string, []>("linear_102_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533469376)))];
            tensor<fp16, [1152]> vpm_encoder_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(536123648)))];
            tensor<fp16, [1, 1024, 1152]> linear_103_cast_fp16 = linear(bias = vpm_encoder_layers_17_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_17_self_attn_k_proj_weight_to_fp16, x = hidden_states_103_cast_fp16)[name = tensor<string, []>("linear_103_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(536126016)))];
            tensor<fp16, [1152]> vpm_encoder_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538780288)))];
            tensor<fp16, [1, 1024, 1152]> linear_104_cast_fp16 = linear(bias = vpm_encoder_layers_17_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_17_self_attn_v_proj_weight_to_fp16, x = hidden_states_103_cast_fp16)[name = tensor<string, []>("linear_104_cast_fp16")];
            tensor<int32, [4]> var_1173 = const()[name = tensor<string, []>("op_1173"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1174_cast_fp16 = reshape(shape = var_1173, x = linear_102_cast_fp16)[name = tensor<string, []>("op_1174_cast_fp16")];
            tensor<int32, [4]> var_1176 = const()[name = tensor<string, []>("op_1176"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1177_cast_fp16 = reshape(shape = var_1176, x = linear_103_cast_fp16)[name = tensor<string, []>("op_1177_cast_fp16")];
            tensor<int32, [4]> var_1179 = const()[name = tensor<string, []>("op_1179"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1180_cast_fp16 = reshape(shape = var_1179, x = linear_104_cast_fp16)[name = tensor<string, []>("op_1180_cast_fp16")];
            tensor<int32, [4]> value_states_71_perm_0 = const()[name = tensor<string, []>("value_states_71_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1183_transpose_x_0 = const()[name = tensor<string, []>("op_1183_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1183_transpose_y_0 = const()[name = tensor<string, []>("op_1183_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_118_perm_0 = const()[name = tensor<string, []>("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_1177_cast_fp16)[name = tensor<string, []>("transpose_179")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_117 = transpose(perm = transpose_117_perm_0, x = var_1174_cast_fp16)[name = tensor<string, []>("transpose_180")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1183_cast_fp16 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = transpose_117, y = transpose_118)[name = tensor<string, []>("op_1183_cast_fp16")];
            tensor<fp16, []> var_1184_to_fp16 = const()[name = tensor<string, []>("op_1184_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_69_cast_fp16 = mul(x = var_1183_cast_fp16, y = var_1184_to_fp16)[name = tensor<string, []>("attn_weights_69_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1186_cast_fp16 = softmax(axis = var_18, x = attn_weights_69_cast_fp16)[name = tensor<string, []>("op_1186_cast_fp16")];
            tensor<bool, []> attn_output_69_transpose_x_0 = const()[name = tensor<string, []>("attn_output_69_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_69_transpose_y_0 = const()[name = tensor<string, []>("attn_output_69_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_71_cast_fp16 = transpose(perm = value_states_71_perm_0, x = var_1180_cast_fp16)[name = tensor<string, []>("transpose_181")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = var_1186_cast_fp16, y = value_states_71_cast_fp16)[name = tensor<string, []>("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_1190_perm_0 = const()[name = tensor<string, []>("op_1190_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1192 = const()[name = tensor<string, []>("op_1192"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1190_cast_fp16 = transpose(perm = var_1190_perm_0, x = attn_output_69_cast_fp16)[name = tensor<string, []>("transpose_178")];
            tensor<fp16, [1, 1024, 1152]> input_245_cast_fp16 = reshape(shape = var_1192, x = var_1190_cast_fp16)[name = tensor<string, []>("input_245_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_17_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538782656)))];
            tensor<fp16, [1152]> vpm_encoder_layers_17_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(541436928)))];
            tensor<fp16, [1, 1024, 1152]> linear_105_cast_fp16 = linear(bias = vpm_encoder_layers_17_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_17_self_attn_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor<string, []>("linear_105_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_105_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
            tensor<int32, [1]> input_249_axes_0 = const()[name = tensor<string, []>("input_249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_17_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(541439296)))];
            tensor<fp16, [1152]> vpm_encoder_layers_17_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(541441664)))];
            tensor<fp16, [1, 1024, 1152]> input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = vpm_encoder_layers_17_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_17_layer_norm2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("input_249_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_17_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(541444032)))];
            tensor<fp16, [4304]> vpm_encoder_layers_17_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551360512)))];
            tensor<fp16, [1, 1024, 4304]> linear_106_cast_fp16 = linear(bias = vpm_encoder_layers_17_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_17_mlp_fc1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor<string, []>("linear_106_cast_fp16")];
            tensor<string, []> input_253_mode_0 = const()[name = tensor<string, []>("input_253_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_106_cast_fp16)[name = tensor<string, []>("input_253_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_17_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551369216)))];
            tensor<fp16, [1152]> vpm_encoder_layers_17_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_17_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(561285696)))];
            tensor<fp16, [1, 1024, 1152]> linear_107_cast_fp16 = linear(bias = vpm_encoder_layers_17_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_17_mlp_fc2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor<string, []>("linear_107_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_255_cast_fp16 = add(x = input_247_cast_fp16, y = linear_107_cast_fp16)[name = tensor<string, []>("input_255_cast_fp16")];
            tensor<int32, [1]> hidden_states_109_axes_0 = const()[name = tensor<string, []>("hidden_states_109_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_18_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(561288064)))];
            tensor<fp16, [1152]> vpm_encoder_layers_18_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(561290432)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_109_cast_fp16 = layer_norm(axes = hidden_states_109_axes_0, beta = vpm_encoder_layers_18_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_18_layer_norm1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor<string, []>("hidden_states_109_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(561292800)))];
            tensor<fp16, [1152]> vpm_encoder_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563947072)))];
            tensor<fp16, [1, 1024, 1152]> linear_108_cast_fp16 = linear(bias = vpm_encoder_layers_18_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_18_self_attn_q_proj_weight_to_fp16, x = hidden_states_109_cast_fp16)[name = tensor<string, []>("linear_108_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563949440)))];
            tensor<fp16, [1152]> vpm_encoder_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566603712)))];
            tensor<fp16, [1, 1024, 1152]> linear_109_cast_fp16 = linear(bias = vpm_encoder_layers_18_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_18_self_attn_k_proj_weight_to_fp16, x = hidden_states_109_cast_fp16)[name = tensor<string, []>("linear_109_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566606080)))];
            tensor<fp16, [1152]> vpm_encoder_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(569260352)))];
            tensor<fp16, [1, 1024, 1152]> linear_110_cast_fp16 = linear(bias = vpm_encoder_layers_18_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_18_self_attn_v_proj_weight_to_fp16, x = hidden_states_109_cast_fp16)[name = tensor<string, []>("linear_110_cast_fp16")];
            tensor<int32, [4]> var_1235 = const()[name = tensor<string, []>("op_1235"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1236_cast_fp16 = reshape(shape = var_1235, x = linear_108_cast_fp16)[name = tensor<string, []>("op_1236_cast_fp16")];
            tensor<int32, [4]> var_1238 = const()[name = tensor<string, []>("op_1238"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1239_cast_fp16 = reshape(shape = var_1238, x = linear_109_cast_fp16)[name = tensor<string, []>("op_1239_cast_fp16")];
            tensor<int32, [4]> var_1241 = const()[name = tensor<string, []>("op_1241"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1242_cast_fp16 = reshape(shape = var_1241, x = linear_110_cast_fp16)[name = tensor<string, []>("op_1242_cast_fp16")];
            tensor<int32, [4]> value_states_75_perm_0 = const()[name = tensor<string, []>("value_states_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1245_transpose_x_0 = const()[name = tensor<string, []>("op_1245_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1245_transpose_y_0 = const()[name = tensor<string, []>("op_1245_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_120_perm_0 = const()[name = tensor<string, []>("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_120 = transpose(perm = transpose_120_perm_0, x = var_1239_cast_fp16)[name = tensor<string, []>("transpose_175")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_119 = transpose(perm = transpose_119_perm_0, x = var_1236_cast_fp16)[name = tensor<string, []>("transpose_176")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1245_cast_fp16 = matmul(transpose_x = var_1245_transpose_x_0, transpose_y = var_1245_transpose_y_0, x = transpose_119, y = transpose_120)[name = tensor<string, []>("op_1245_cast_fp16")];
            tensor<fp16, []> var_1246_to_fp16 = const()[name = tensor<string, []>("op_1246_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_73_cast_fp16 = mul(x = var_1245_cast_fp16, y = var_1246_to_fp16)[name = tensor<string, []>("attn_weights_73_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1248_cast_fp16 = softmax(axis = var_18, x = attn_weights_73_cast_fp16)[name = tensor<string, []>("op_1248_cast_fp16")];
            tensor<bool, []> attn_output_73_transpose_x_0 = const()[name = tensor<string, []>("attn_output_73_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_73_transpose_y_0 = const()[name = tensor<string, []>("attn_output_73_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_75_cast_fp16 = transpose(perm = value_states_75_perm_0, x = var_1242_cast_fp16)[name = tensor<string, []>("transpose_177")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = var_1248_cast_fp16, y = value_states_75_cast_fp16)[name = tensor<string, []>("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_1252_perm_0 = const()[name = tensor<string, []>("op_1252_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1254 = const()[name = tensor<string, []>("op_1254"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1252_cast_fp16 = transpose(perm = var_1252_perm_0, x = attn_output_73_cast_fp16)[name = tensor<string, []>("transpose_174")];
            tensor<fp16, [1, 1024, 1152]> input_259_cast_fp16 = reshape(shape = var_1254, x = var_1252_cast_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_18_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(569262720)))];
            tensor<fp16, [1152]> vpm_encoder_layers_18_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571916992)))];
            tensor<fp16, [1, 1024, 1152]> linear_111_cast_fp16 = linear(bias = vpm_encoder_layers_18_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_18_self_attn_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor<string, []>("linear_111_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_111_cast_fp16)[name = tensor<string, []>("input_261_cast_fp16")];
            tensor<int32, [1]> input_263_axes_0 = const()[name = tensor<string, []>("input_263_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_18_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571919360)))];
            tensor<fp16, [1152]> vpm_encoder_layers_18_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571921728)))];
            tensor<fp16, [1, 1024, 1152]> input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = vpm_encoder_layers_18_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_18_layer_norm2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor<string, []>("input_263_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_18_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571924096)))];
            tensor<fp16, [4304]> vpm_encoder_layers_18_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581840576)))];
            tensor<fp16, [1, 1024, 4304]> linear_112_cast_fp16 = linear(bias = vpm_encoder_layers_18_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_18_mlp_fc1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor<string, []>("linear_112_cast_fp16")];
            tensor<string, []> input_267_mode_0 = const()[name = tensor<string, []>("input_267_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_267_cast_fp16 = gelu(mode = input_267_mode_0, x = linear_112_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_18_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581849280)))];
            tensor<fp16, [1152]> vpm_encoder_layers_18_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_18_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591765760)))];
            tensor<fp16, [1, 1024, 1152]> linear_113_cast_fp16 = linear(bias = vpm_encoder_layers_18_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_18_mlp_fc2_weight_to_fp16, x = input_267_cast_fp16)[name = tensor<string, []>("linear_113_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_269_cast_fp16 = add(x = input_261_cast_fp16, y = linear_113_cast_fp16)[name = tensor<string, []>("input_269_cast_fp16")];
            tensor<int32, [1]> hidden_states_115_axes_0 = const()[name = tensor<string, []>("hidden_states_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_19_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591768128)))];
            tensor<fp16, [1152]> vpm_encoder_layers_19_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591770496)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_115_cast_fp16 = layer_norm(axes = hidden_states_115_axes_0, beta = vpm_encoder_layers_19_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_19_layer_norm1_weight_to_fp16, x = input_269_cast_fp16)[name = tensor<string, []>("hidden_states_115_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591772864)))];
            tensor<fp16, [1152]> vpm_encoder_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(594427136)))];
            tensor<fp16, [1, 1024, 1152]> linear_114_cast_fp16 = linear(bias = vpm_encoder_layers_19_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_19_self_attn_q_proj_weight_to_fp16, x = hidden_states_115_cast_fp16)[name = tensor<string, []>("linear_114_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(594429504)))];
            tensor<fp16, [1152]> vpm_encoder_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(597083776)))];
            tensor<fp16, [1, 1024, 1152]> linear_115_cast_fp16 = linear(bias = vpm_encoder_layers_19_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_19_self_attn_k_proj_weight_to_fp16, x = hidden_states_115_cast_fp16)[name = tensor<string, []>("linear_115_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(597086144)))];
            tensor<fp16, [1152]> vpm_encoder_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(599740416)))];
            tensor<fp16, [1, 1024, 1152]> linear_116_cast_fp16 = linear(bias = vpm_encoder_layers_19_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_19_self_attn_v_proj_weight_to_fp16, x = hidden_states_115_cast_fp16)[name = tensor<string, []>("linear_116_cast_fp16")];
            tensor<int32, [4]> var_1297 = const()[name = tensor<string, []>("op_1297"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1298_cast_fp16 = reshape(shape = var_1297, x = linear_114_cast_fp16)[name = tensor<string, []>("op_1298_cast_fp16")];
            tensor<int32, [4]> var_1300 = const()[name = tensor<string, []>("op_1300"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1301_cast_fp16 = reshape(shape = var_1300, x = linear_115_cast_fp16)[name = tensor<string, []>("op_1301_cast_fp16")];
            tensor<int32, [4]> var_1303 = const()[name = tensor<string, []>("op_1303"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1304_cast_fp16 = reshape(shape = var_1303, x = linear_116_cast_fp16)[name = tensor<string, []>("op_1304_cast_fp16")];
            tensor<int32, [4]> value_states_79_perm_0 = const()[name = tensor<string, []>("value_states_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1307_transpose_x_0 = const()[name = tensor<string, []>("op_1307_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1307_transpose_y_0 = const()[name = tensor<string, []>("op_1307_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_121_perm_0 = const()[name = tensor<string, []>("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_122_perm_0 = const()[name = tensor<string, []>("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_122 = transpose(perm = transpose_122_perm_0, x = var_1301_cast_fp16)[name = tensor<string, []>("transpose_171")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_121 = transpose(perm = transpose_121_perm_0, x = var_1298_cast_fp16)[name = tensor<string, []>("transpose_172")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1307_cast_fp16 = matmul(transpose_x = var_1307_transpose_x_0, transpose_y = var_1307_transpose_y_0, x = transpose_121, y = transpose_122)[name = tensor<string, []>("op_1307_cast_fp16")];
            tensor<fp16, []> var_1308_to_fp16 = const()[name = tensor<string, []>("op_1308_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_77_cast_fp16 = mul(x = var_1307_cast_fp16, y = var_1308_to_fp16)[name = tensor<string, []>("attn_weights_77_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1310_cast_fp16 = softmax(axis = var_18, x = attn_weights_77_cast_fp16)[name = tensor<string, []>("op_1310_cast_fp16")];
            tensor<bool, []> attn_output_77_transpose_x_0 = const()[name = tensor<string, []>("attn_output_77_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_77_transpose_y_0 = const()[name = tensor<string, []>("attn_output_77_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_79_cast_fp16 = transpose(perm = value_states_79_perm_0, x = var_1304_cast_fp16)[name = tensor<string, []>("transpose_173")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = var_1310_cast_fp16, y = value_states_79_cast_fp16)[name = tensor<string, []>("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_1314_perm_0 = const()[name = tensor<string, []>("op_1314_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1316 = const()[name = tensor<string, []>("op_1316"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1314_cast_fp16 = transpose(perm = var_1314_perm_0, x = attn_output_77_cast_fp16)[name = tensor<string, []>("transpose_170")];
            tensor<fp16, [1, 1024, 1152]> input_273_cast_fp16 = reshape(shape = var_1316, x = var_1314_cast_fp16)[name = tensor<string, []>("input_273_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_19_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(599742784)))];
            tensor<fp16, [1152]> vpm_encoder_layers_19_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(602397056)))];
            tensor<fp16, [1, 1024, 1152]> linear_117_cast_fp16 = linear(bias = vpm_encoder_layers_19_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_19_self_attn_out_proj_weight_to_fp16, x = input_273_cast_fp16)[name = tensor<string, []>("linear_117_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_275_cast_fp16 = add(x = input_269_cast_fp16, y = linear_117_cast_fp16)[name = tensor<string, []>("input_275_cast_fp16")];
            tensor<int32, [1]> input_277_axes_0 = const()[name = tensor<string, []>("input_277_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_19_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(602399424)))];
            tensor<fp16, [1152]> vpm_encoder_layers_19_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(602401792)))];
            tensor<fp16, [1, 1024, 1152]> input_277_cast_fp16 = layer_norm(axes = input_277_axes_0, beta = vpm_encoder_layers_19_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_19_layer_norm2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_19_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(602404160)))];
            tensor<fp16, [4304]> vpm_encoder_layers_19_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(612320640)))];
            tensor<fp16, [1, 1024, 4304]> linear_118_cast_fp16 = linear(bias = vpm_encoder_layers_19_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_19_mlp_fc1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor<string, []>("linear_118_cast_fp16")];
            tensor<string, []> input_281_mode_0 = const()[name = tensor<string, []>("input_281_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_281_cast_fp16 = gelu(mode = input_281_mode_0, x = linear_118_cast_fp16)[name = tensor<string, []>("input_281_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_19_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(612329344)))];
            tensor<fp16, [1152]> vpm_encoder_layers_19_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_19_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622245824)))];
            tensor<fp16, [1, 1024, 1152]> linear_119_cast_fp16 = linear(bias = vpm_encoder_layers_19_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_19_mlp_fc2_weight_to_fp16, x = input_281_cast_fp16)[name = tensor<string, []>("linear_119_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_283_cast_fp16 = add(x = input_275_cast_fp16, y = linear_119_cast_fp16)[name = tensor<string, []>("input_283_cast_fp16")];
            tensor<int32, [1]> hidden_states_121_axes_0 = const()[name = tensor<string, []>("hidden_states_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_20_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622248192)))];
            tensor<fp16, [1152]> vpm_encoder_layers_20_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622250560)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_121_cast_fp16 = layer_norm(axes = hidden_states_121_axes_0, beta = vpm_encoder_layers_20_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_20_layer_norm1_weight_to_fp16, x = input_283_cast_fp16)[name = tensor<string, []>("hidden_states_121_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(622252928)))];
            tensor<fp16, [1152]> vpm_encoder_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(624907200)))];
            tensor<fp16, [1, 1024, 1152]> linear_120_cast_fp16 = linear(bias = vpm_encoder_layers_20_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_20_self_attn_q_proj_weight_to_fp16, x = hidden_states_121_cast_fp16)[name = tensor<string, []>("linear_120_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(624909568)))];
            tensor<fp16, [1152]> vpm_encoder_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(627563840)))];
            tensor<fp16, [1, 1024, 1152]> linear_121_cast_fp16 = linear(bias = vpm_encoder_layers_20_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_20_self_attn_k_proj_weight_to_fp16, x = hidden_states_121_cast_fp16)[name = tensor<string, []>("linear_121_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(627566208)))];
            tensor<fp16, [1152]> vpm_encoder_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(630220480)))];
            tensor<fp16, [1, 1024, 1152]> linear_122_cast_fp16 = linear(bias = vpm_encoder_layers_20_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_20_self_attn_v_proj_weight_to_fp16, x = hidden_states_121_cast_fp16)[name = tensor<string, []>("linear_122_cast_fp16")];
            tensor<int32, [4]> var_1359 = const()[name = tensor<string, []>("op_1359"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1360_cast_fp16 = reshape(shape = var_1359, x = linear_120_cast_fp16)[name = tensor<string, []>("op_1360_cast_fp16")];
            tensor<int32, [4]> var_1362 = const()[name = tensor<string, []>("op_1362"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1363_cast_fp16 = reshape(shape = var_1362, x = linear_121_cast_fp16)[name = tensor<string, []>("op_1363_cast_fp16")];
            tensor<int32, [4]> var_1365 = const()[name = tensor<string, []>("op_1365"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1366_cast_fp16 = reshape(shape = var_1365, x = linear_122_cast_fp16)[name = tensor<string, []>("op_1366_cast_fp16")];
            tensor<int32, [4]> value_states_83_perm_0 = const()[name = tensor<string, []>("value_states_83_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1369_transpose_x_0 = const()[name = tensor<string, []>("op_1369_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1369_transpose_y_0 = const()[name = tensor<string, []>("op_1369_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_124_perm_0 = const()[name = tensor<string, []>("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_124 = transpose(perm = transpose_124_perm_0, x = var_1363_cast_fp16)[name = tensor<string, []>("transpose_167")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_123 = transpose(perm = transpose_123_perm_0, x = var_1360_cast_fp16)[name = tensor<string, []>("transpose_168")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1369_cast_fp16 = matmul(transpose_x = var_1369_transpose_x_0, transpose_y = var_1369_transpose_y_0, x = transpose_123, y = transpose_124)[name = tensor<string, []>("op_1369_cast_fp16")];
            tensor<fp16, []> var_1370_to_fp16 = const()[name = tensor<string, []>("op_1370_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_81_cast_fp16 = mul(x = var_1369_cast_fp16, y = var_1370_to_fp16)[name = tensor<string, []>("attn_weights_81_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1372_cast_fp16 = softmax(axis = var_18, x = attn_weights_81_cast_fp16)[name = tensor<string, []>("op_1372_cast_fp16")];
            tensor<bool, []> attn_output_81_transpose_x_0 = const()[name = tensor<string, []>("attn_output_81_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_81_transpose_y_0 = const()[name = tensor<string, []>("attn_output_81_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_83_cast_fp16 = transpose(perm = value_states_83_perm_0, x = var_1366_cast_fp16)[name = tensor<string, []>("transpose_169")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = var_1372_cast_fp16, y = value_states_83_cast_fp16)[name = tensor<string, []>("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_1376_perm_0 = const()[name = tensor<string, []>("op_1376_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1378 = const()[name = tensor<string, []>("op_1378"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1376_cast_fp16 = transpose(perm = var_1376_perm_0, x = attn_output_81_cast_fp16)[name = tensor<string, []>("transpose_166")];
            tensor<fp16, [1, 1024, 1152]> input_287_cast_fp16 = reshape(shape = var_1378, x = var_1376_cast_fp16)[name = tensor<string, []>("input_287_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_20_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(630222848)))];
            tensor<fp16, [1152]> vpm_encoder_layers_20_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(632877120)))];
            tensor<fp16, [1, 1024, 1152]> linear_123_cast_fp16 = linear(bias = vpm_encoder_layers_20_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_20_self_attn_out_proj_weight_to_fp16, x = input_287_cast_fp16)[name = tensor<string, []>("linear_123_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_289_cast_fp16 = add(x = input_283_cast_fp16, y = linear_123_cast_fp16)[name = tensor<string, []>("input_289_cast_fp16")];
            tensor<int32, [1]> input_291_axes_0 = const()[name = tensor<string, []>("input_291_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_20_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(632879488)))];
            tensor<fp16, [1152]> vpm_encoder_layers_20_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(632881856)))];
            tensor<fp16, [1, 1024, 1152]> input_291_cast_fp16 = layer_norm(axes = input_291_axes_0, beta = vpm_encoder_layers_20_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_20_layer_norm2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor<string, []>("input_291_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_20_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(632884224)))];
            tensor<fp16, [4304]> vpm_encoder_layers_20_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(642800704)))];
            tensor<fp16, [1, 1024, 4304]> linear_124_cast_fp16 = linear(bias = vpm_encoder_layers_20_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_20_mlp_fc1_weight_to_fp16, x = input_291_cast_fp16)[name = tensor<string, []>("linear_124_cast_fp16")];
            tensor<string, []> input_295_mode_0 = const()[name = tensor<string, []>("input_295_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_295_cast_fp16 = gelu(mode = input_295_mode_0, x = linear_124_cast_fp16)[name = tensor<string, []>("input_295_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_20_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(642809408)))];
            tensor<fp16, [1152]> vpm_encoder_layers_20_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_20_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652725888)))];
            tensor<fp16, [1, 1024, 1152]> linear_125_cast_fp16 = linear(bias = vpm_encoder_layers_20_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_20_mlp_fc2_weight_to_fp16, x = input_295_cast_fp16)[name = tensor<string, []>("linear_125_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_297_cast_fp16 = add(x = input_289_cast_fp16, y = linear_125_cast_fp16)[name = tensor<string, []>("input_297_cast_fp16")];
            tensor<int32, [1]> hidden_states_127_axes_0 = const()[name = tensor<string, []>("hidden_states_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_21_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652728256)))];
            tensor<fp16, [1152]> vpm_encoder_layers_21_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652730624)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_127_cast_fp16 = layer_norm(axes = hidden_states_127_axes_0, beta = vpm_encoder_layers_21_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_21_layer_norm1_weight_to_fp16, x = input_297_cast_fp16)[name = tensor<string, []>("hidden_states_127_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652732992)))];
            tensor<fp16, [1152]> vpm_encoder_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655387264)))];
            tensor<fp16, [1, 1024, 1152]> linear_126_cast_fp16 = linear(bias = vpm_encoder_layers_21_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_21_self_attn_q_proj_weight_to_fp16, x = hidden_states_127_cast_fp16)[name = tensor<string, []>("linear_126_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655389632)))];
            tensor<fp16, [1152]> vpm_encoder_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(658043904)))];
            tensor<fp16, [1, 1024, 1152]> linear_127_cast_fp16 = linear(bias = vpm_encoder_layers_21_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_21_self_attn_k_proj_weight_to_fp16, x = hidden_states_127_cast_fp16)[name = tensor<string, []>("linear_127_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(658046272)))];
            tensor<fp16, [1152]> vpm_encoder_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(660700544)))];
            tensor<fp16, [1, 1024, 1152]> linear_128_cast_fp16 = linear(bias = vpm_encoder_layers_21_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_21_self_attn_v_proj_weight_to_fp16, x = hidden_states_127_cast_fp16)[name = tensor<string, []>("linear_128_cast_fp16")];
            tensor<int32, [4]> var_1421 = const()[name = tensor<string, []>("op_1421"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1422_cast_fp16 = reshape(shape = var_1421, x = linear_126_cast_fp16)[name = tensor<string, []>("op_1422_cast_fp16")];
            tensor<int32, [4]> var_1424 = const()[name = tensor<string, []>("op_1424"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1425_cast_fp16 = reshape(shape = var_1424, x = linear_127_cast_fp16)[name = tensor<string, []>("op_1425_cast_fp16")];
            tensor<int32, [4]> var_1427 = const()[name = tensor<string, []>("op_1427"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1428_cast_fp16 = reshape(shape = var_1427, x = linear_128_cast_fp16)[name = tensor<string, []>("op_1428_cast_fp16")];
            tensor<int32, [4]> value_states_87_perm_0 = const()[name = tensor<string, []>("value_states_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1431_transpose_x_0 = const()[name = tensor<string, []>("op_1431_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1431_transpose_y_0 = const()[name = tensor<string, []>("op_1431_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_126_perm_0 = const()[name = tensor<string, []>("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_126 = transpose(perm = transpose_126_perm_0, x = var_1425_cast_fp16)[name = tensor<string, []>("transpose_163")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_125 = transpose(perm = transpose_125_perm_0, x = var_1422_cast_fp16)[name = tensor<string, []>("transpose_164")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1431_cast_fp16 = matmul(transpose_x = var_1431_transpose_x_0, transpose_y = var_1431_transpose_y_0, x = transpose_125, y = transpose_126)[name = tensor<string, []>("op_1431_cast_fp16")];
            tensor<fp16, []> var_1432_to_fp16 = const()[name = tensor<string, []>("op_1432_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_85_cast_fp16 = mul(x = var_1431_cast_fp16, y = var_1432_to_fp16)[name = tensor<string, []>("attn_weights_85_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1434_cast_fp16 = softmax(axis = var_18, x = attn_weights_85_cast_fp16)[name = tensor<string, []>("op_1434_cast_fp16")];
            tensor<bool, []> attn_output_85_transpose_x_0 = const()[name = tensor<string, []>("attn_output_85_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_85_transpose_y_0 = const()[name = tensor<string, []>("attn_output_85_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_87_cast_fp16 = transpose(perm = value_states_87_perm_0, x = var_1428_cast_fp16)[name = tensor<string, []>("transpose_165")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = var_1434_cast_fp16, y = value_states_87_cast_fp16)[name = tensor<string, []>("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_1438_perm_0 = const()[name = tensor<string, []>("op_1438_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1440 = const()[name = tensor<string, []>("op_1440"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1438_cast_fp16 = transpose(perm = var_1438_perm_0, x = attn_output_85_cast_fp16)[name = tensor<string, []>("transpose_162")];
            tensor<fp16, [1, 1024, 1152]> input_301_cast_fp16 = reshape(shape = var_1440, x = var_1438_cast_fp16)[name = tensor<string, []>("input_301_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_21_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(660702912)))];
            tensor<fp16, [1152]> vpm_encoder_layers_21_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(663357184)))];
            tensor<fp16, [1, 1024, 1152]> linear_129_cast_fp16 = linear(bias = vpm_encoder_layers_21_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_21_self_attn_out_proj_weight_to_fp16, x = input_301_cast_fp16)[name = tensor<string, []>("linear_129_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_303_cast_fp16 = add(x = input_297_cast_fp16, y = linear_129_cast_fp16)[name = tensor<string, []>("input_303_cast_fp16")];
            tensor<int32, [1]> input_305_axes_0 = const()[name = tensor<string, []>("input_305_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_21_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(663359552)))];
            tensor<fp16, [1152]> vpm_encoder_layers_21_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(663361920)))];
            tensor<fp16, [1, 1024, 1152]> input_305_cast_fp16 = layer_norm(axes = input_305_axes_0, beta = vpm_encoder_layers_21_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_21_layer_norm2_weight_to_fp16, x = input_303_cast_fp16)[name = tensor<string, []>("input_305_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_21_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(663364288)))];
            tensor<fp16, [4304]> vpm_encoder_layers_21_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(673280768)))];
            tensor<fp16, [1, 1024, 4304]> linear_130_cast_fp16 = linear(bias = vpm_encoder_layers_21_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_21_mlp_fc1_weight_to_fp16, x = input_305_cast_fp16)[name = tensor<string, []>("linear_130_cast_fp16")];
            tensor<string, []> input_309_mode_0 = const()[name = tensor<string, []>("input_309_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_309_cast_fp16 = gelu(mode = input_309_mode_0, x = linear_130_cast_fp16)[name = tensor<string, []>("input_309_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_21_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(673289472)))];
            tensor<fp16, [1152]> vpm_encoder_layers_21_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_21_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(683205952)))];
            tensor<fp16, [1, 1024, 1152]> linear_131_cast_fp16 = linear(bias = vpm_encoder_layers_21_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_21_mlp_fc2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor<string, []>("linear_131_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_311_cast_fp16 = add(x = input_303_cast_fp16, y = linear_131_cast_fp16)[name = tensor<string, []>("input_311_cast_fp16")];
            tensor<int32, [1]> hidden_states_133_axes_0 = const()[name = tensor<string, []>("hidden_states_133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_22_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(683208320)))];
            tensor<fp16, [1152]> vpm_encoder_layers_22_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(683210688)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_133_cast_fp16 = layer_norm(axes = hidden_states_133_axes_0, beta = vpm_encoder_layers_22_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_22_layer_norm1_weight_to_fp16, x = input_311_cast_fp16)[name = tensor<string, []>("hidden_states_133_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(683213056)))];
            tensor<fp16, [1152]> vpm_encoder_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(685867328)))];
            tensor<fp16, [1, 1024, 1152]> linear_132_cast_fp16 = linear(bias = vpm_encoder_layers_22_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_22_self_attn_q_proj_weight_to_fp16, x = hidden_states_133_cast_fp16)[name = tensor<string, []>("linear_132_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(685869696)))];
            tensor<fp16, [1152]> vpm_encoder_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(688523968)))];
            tensor<fp16, [1, 1024, 1152]> linear_133_cast_fp16 = linear(bias = vpm_encoder_layers_22_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_22_self_attn_k_proj_weight_to_fp16, x = hidden_states_133_cast_fp16)[name = tensor<string, []>("linear_133_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(688526336)))];
            tensor<fp16, [1152]> vpm_encoder_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(691180608)))];
            tensor<fp16, [1, 1024, 1152]> linear_134_cast_fp16 = linear(bias = vpm_encoder_layers_22_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_22_self_attn_v_proj_weight_to_fp16, x = hidden_states_133_cast_fp16)[name = tensor<string, []>("linear_134_cast_fp16")];
            tensor<int32, [4]> var_1483 = const()[name = tensor<string, []>("op_1483"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1484_cast_fp16 = reshape(shape = var_1483, x = linear_132_cast_fp16)[name = tensor<string, []>("op_1484_cast_fp16")];
            tensor<int32, [4]> var_1486 = const()[name = tensor<string, []>("op_1486"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1487_cast_fp16 = reshape(shape = var_1486, x = linear_133_cast_fp16)[name = tensor<string, []>("op_1487_cast_fp16")];
            tensor<int32, [4]> var_1489 = const()[name = tensor<string, []>("op_1489"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1490_cast_fp16 = reshape(shape = var_1489, x = linear_134_cast_fp16)[name = tensor<string, []>("op_1490_cast_fp16")];
            tensor<int32, [4]> value_states_91_perm_0 = const()[name = tensor<string, []>("value_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1493_transpose_x_0 = const()[name = tensor<string, []>("op_1493_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1493_transpose_y_0 = const()[name = tensor<string, []>("op_1493_transpose_y_0"), val = tensor<bool, []>(false)];
            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, [4]> transpose_128_perm_0 = const()[name = tensor<string, []>("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_128 = transpose(perm = transpose_128_perm_0, x = var_1487_cast_fp16)[name = tensor<string, []>("transpose_159")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_127 = transpose(perm = transpose_127_perm_0, x = var_1484_cast_fp16)[name = tensor<string, []>("transpose_160")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1493_cast_fp16 = matmul(transpose_x = var_1493_transpose_x_0, transpose_y = var_1493_transpose_y_0, x = transpose_127, y = transpose_128)[name = tensor<string, []>("op_1493_cast_fp16")];
            tensor<fp16, []> var_1494_to_fp16 = const()[name = tensor<string, []>("op_1494_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_89_cast_fp16 = mul(x = var_1493_cast_fp16, y = var_1494_to_fp16)[name = tensor<string, []>("attn_weights_89_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1496_cast_fp16 = softmax(axis = var_18, x = attn_weights_89_cast_fp16)[name = tensor<string, []>("op_1496_cast_fp16")];
            tensor<bool, []> attn_output_89_transpose_x_0 = const()[name = tensor<string, []>("attn_output_89_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_89_transpose_y_0 = const()[name = tensor<string, []>("attn_output_89_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_91_cast_fp16 = transpose(perm = value_states_91_perm_0, x = var_1490_cast_fp16)[name = tensor<string, []>("transpose_161")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = var_1496_cast_fp16, y = value_states_91_cast_fp16)[name = tensor<string, []>("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_1500_perm_0 = const()[name = tensor<string, []>("op_1500_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1502 = const()[name = tensor<string, []>("op_1502"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1500_cast_fp16 = transpose(perm = var_1500_perm_0, x = attn_output_89_cast_fp16)[name = tensor<string, []>("transpose_158")];
            tensor<fp16, [1, 1024, 1152]> input_315_cast_fp16 = reshape(shape = var_1502, x = var_1500_cast_fp16)[name = tensor<string, []>("input_315_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_22_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(691182976)))];
            tensor<fp16, [1152]> vpm_encoder_layers_22_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693837248)))];
            tensor<fp16, [1, 1024, 1152]> linear_135_cast_fp16 = linear(bias = vpm_encoder_layers_22_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_22_self_attn_out_proj_weight_to_fp16, x = input_315_cast_fp16)[name = tensor<string, []>("linear_135_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_317_cast_fp16 = add(x = input_311_cast_fp16, y = linear_135_cast_fp16)[name = tensor<string, []>("input_317_cast_fp16")];
            tensor<int32, [1]> input_319_axes_0 = const()[name = tensor<string, []>("input_319_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_22_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693839616)))];
            tensor<fp16, [1152]> vpm_encoder_layers_22_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693841984)))];
            tensor<fp16, [1, 1024, 1152]> input_319_cast_fp16 = layer_norm(axes = input_319_axes_0, beta = vpm_encoder_layers_22_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_22_layer_norm2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_22_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(693844352)))];
            tensor<fp16, [4304]> vpm_encoder_layers_22_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(703760832)))];
            tensor<fp16, [1, 1024, 4304]> linear_136_cast_fp16 = linear(bias = vpm_encoder_layers_22_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_22_mlp_fc1_weight_to_fp16, x = input_319_cast_fp16)[name = tensor<string, []>("linear_136_cast_fp16")];
            tensor<string, []> input_323_mode_0 = const()[name = tensor<string, []>("input_323_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_323_cast_fp16 = gelu(mode = input_323_mode_0, x = linear_136_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_22_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(703769536)))];
            tensor<fp16, [1152]> vpm_encoder_layers_22_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_22_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713686016)))];
            tensor<fp16, [1, 1024, 1152]> linear_137_cast_fp16 = linear(bias = vpm_encoder_layers_22_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_22_mlp_fc2_weight_to_fp16, x = input_323_cast_fp16)[name = tensor<string, []>("linear_137_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_325_cast_fp16 = add(x = input_317_cast_fp16, y = linear_137_cast_fp16)[name = tensor<string, []>("input_325_cast_fp16")];
            tensor<int32, [1]> hidden_states_139_axes_0 = const()[name = tensor<string, []>("hidden_states_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_23_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713688384)))];
            tensor<fp16, [1152]> vpm_encoder_layers_23_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713690752)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_139_cast_fp16 = layer_norm(axes = hidden_states_139_axes_0, beta = vpm_encoder_layers_23_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_23_layer_norm1_weight_to_fp16, x = input_325_cast_fp16)[name = tensor<string, []>("hidden_states_139_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(713693120)))];
            tensor<fp16, [1152]> vpm_encoder_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(716347392)))];
            tensor<fp16, [1, 1024, 1152]> linear_138_cast_fp16 = linear(bias = vpm_encoder_layers_23_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_23_self_attn_q_proj_weight_to_fp16, x = hidden_states_139_cast_fp16)[name = tensor<string, []>("linear_138_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(716349760)))];
            tensor<fp16, [1152]> vpm_encoder_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(719004032)))];
            tensor<fp16, [1, 1024, 1152]> linear_139_cast_fp16 = linear(bias = vpm_encoder_layers_23_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_23_self_attn_k_proj_weight_to_fp16, x = hidden_states_139_cast_fp16)[name = tensor<string, []>("linear_139_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(719006400)))];
            tensor<fp16, [1152]> vpm_encoder_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721660672)))];
            tensor<fp16, [1, 1024, 1152]> linear_140_cast_fp16 = linear(bias = vpm_encoder_layers_23_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_23_self_attn_v_proj_weight_to_fp16, x = hidden_states_139_cast_fp16)[name = tensor<string, []>("linear_140_cast_fp16")];
            tensor<int32, [4]> var_1545 = const()[name = tensor<string, []>("op_1545"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1546_cast_fp16 = reshape(shape = var_1545, x = linear_138_cast_fp16)[name = tensor<string, []>("op_1546_cast_fp16")];
            tensor<int32, [4]> var_1548 = const()[name = tensor<string, []>("op_1548"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1549_cast_fp16 = reshape(shape = var_1548, x = linear_139_cast_fp16)[name = tensor<string, []>("op_1549_cast_fp16")];
            tensor<int32, [4]> var_1551 = const()[name = tensor<string, []>("op_1551"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1552_cast_fp16 = reshape(shape = var_1551, x = linear_140_cast_fp16)[name = tensor<string, []>("op_1552_cast_fp16")];
            tensor<int32, [4]> value_states_95_perm_0 = const()[name = tensor<string, []>("value_states_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1555_transpose_x_0 = const()[name = tensor<string, []>("op_1555_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1555_transpose_y_0 = const()[name = tensor<string, []>("op_1555_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_129_perm_0 = const()[name = tensor<string, []>("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_130_perm_0 = const()[name = tensor<string, []>("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_130 = transpose(perm = transpose_130_perm_0, x = var_1549_cast_fp16)[name = tensor<string, []>("transpose_155")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_129 = transpose(perm = transpose_129_perm_0, x = var_1546_cast_fp16)[name = tensor<string, []>("transpose_156")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1555_cast_fp16 = matmul(transpose_x = var_1555_transpose_x_0, transpose_y = var_1555_transpose_y_0, x = transpose_129, y = transpose_130)[name = tensor<string, []>("op_1555_cast_fp16")];
            tensor<fp16, []> var_1556_to_fp16 = const()[name = tensor<string, []>("op_1556_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_93_cast_fp16 = mul(x = var_1555_cast_fp16, y = var_1556_to_fp16)[name = tensor<string, []>("attn_weights_93_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1558_cast_fp16 = softmax(axis = var_18, x = attn_weights_93_cast_fp16)[name = tensor<string, []>("op_1558_cast_fp16")];
            tensor<bool, []> attn_output_93_transpose_x_0 = const()[name = tensor<string, []>("attn_output_93_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_93_transpose_y_0 = const()[name = tensor<string, []>("attn_output_93_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_95_cast_fp16 = transpose(perm = value_states_95_perm_0, x = var_1552_cast_fp16)[name = tensor<string, []>("transpose_157")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = var_1558_cast_fp16, y = value_states_95_cast_fp16)[name = tensor<string, []>("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_1562_perm_0 = const()[name = tensor<string, []>("op_1562_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1564 = const()[name = tensor<string, []>("op_1564"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1562_cast_fp16 = transpose(perm = var_1562_perm_0, x = attn_output_93_cast_fp16)[name = tensor<string, []>("transpose_154")];
            tensor<fp16, [1, 1024, 1152]> input_329_cast_fp16 = reshape(shape = var_1564, x = var_1562_cast_fp16)[name = tensor<string, []>("input_329_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_23_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721663040)))];
            tensor<fp16, [1152]> vpm_encoder_layers_23_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(724317312)))];
            tensor<fp16, [1, 1024, 1152]> linear_141_cast_fp16 = linear(bias = vpm_encoder_layers_23_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_23_self_attn_out_proj_weight_to_fp16, x = input_329_cast_fp16)[name = tensor<string, []>("linear_141_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_331_cast_fp16 = add(x = input_325_cast_fp16, y = linear_141_cast_fp16)[name = tensor<string, []>("input_331_cast_fp16")];
            tensor<int32, [1]> input_333_axes_0 = const()[name = tensor<string, []>("input_333_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_23_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(724319680)))];
            tensor<fp16, [1152]> vpm_encoder_layers_23_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(724322048)))];
            tensor<fp16, [1, 1024, 1152]> input_333_cast_fp16 = layer_norm(axes = input_333_axes_0, beta = vpm_encoder_layers_23_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_23_layer_norm2_weight_to_fp16, x = input_331_cast_fp16)[name = tensor<string, []>("input_333_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_23_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(724324416)))];
            tensor<fp16, [4304]> vpm_encoder_layers_23_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(734240896)))];
            tensor<fp16, [1, 1024, 4304]> linear_142_cast_fp16 = linear(bias = vpm_encoder_layers_23_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_23_mlp_fc1_weight_to_fp16, x = input_333_cast_fp16)[name = tensor<string, []>("linear_142_cast_fp16")];
            tensor<string, []> input_337_mode_0 = const()[name = tensor<string, []>("input_337_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_337_cast_fp16 = gelu(mode = input_337_mode_0, x = linear_142_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_23_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(734249600)))];
            tensor<fp16, [1152]> vpm_encoder_layers_23_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_23_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(744166080)))];
            tensor<fp16, [1, 1024, 1152]> linear_143_cast_fp16 = linear(bias = vpm_encoder_layers_23_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_23_mlp_fc2_weight_to_fp16, x = input_337_cast_fp16)[name = tensor<string, []>("linear_143_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_339_cast_fp16 = add(x = input_331_cast_fp16, y = linear_143_cast_fp16)[name = tensor<string, []>("input_339_cast_fp16")];
            tensor<int32, [1]> hidden_states_145_axes_0 = const()[name = tensor<string, []>("hidden_states_145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_24_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(744168448)))];
            tensor<fp16, [1152]> vpm_encoder_layers_24_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(744170816)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_145_cast_fp16 = layer_norm(axes = hidden_states_145_axes_0, beta = vpm_encoder_layers_24_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_24_layer_norm1_weight_to_fp16, x = input_339_cast_fp16)[name = tensor<string, []>("hidden_states_145_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_24_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(744173184)))];
            tensor<fp16, [1152]> vpm_encoder_layers_24_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(746827456)))];
            tensor<fp16, [1, 1024, 1152]> linear_144_cast_fp16 = linear(bias = vpm_encoder_layers_24_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_24_self_attn_q_proj_weight_to_fp16, x = hidden_states_145_cast_fp16)[name = tensor<string, []>("linear_144_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_24_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(746829824)))];
            tensor<fp16, [1152]> vpm_encoder_layers_24_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(749484096)))];
            tensor<fp16, [1, 1024, 1152]> linear_145_cast_fp16 = linear(bias = vpm_encoder_layers_24_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_24_self_attn_k_proj_weight_to_fp16, x = hidden_states_145_cast_fp16)[name = tensor<string, []>("linear_145_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_24_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(749486464)))];
            tensor<fp16, [1152]> vpm_encoder_layers_24_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(752140736)))];
            tensor<fp16, [1, 1024, 1152]> linear_146_cast_fp16 = linear(bias = vpm_encoder_layers_24_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_24_self_attn_v_proj_weight_to_fp16, x = hidden_states_145_cast_fp16)[name = tensor<string, []>("linear_146_cast_fp16")];
            tensor<int32, [4]> var_1607 = const()[name = tensor<string, []>("op_1607"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1608_cast_fp16 = reshape(shape = var_1607, x = linear_144_cast_fp16)[name = tensor<string, []>("op_1608_cast_fp16")];
            tensor<int32, [4]> var_1610 = const()[name = tensor<string, []>("op_1610"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1611_cast_fp16 = reshape(shape = var_1610, x = linear_145_cast_fp16)[name = tensor<string, []>("op_1611_cast_fp16")];
            tensor<int32, [4]> var_1613 = const()[name = tensor<string, []>("op_1613"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1614_cast_fp16 = reshape(shape = var_1613, x = linear_146_cast_fp16)[name = tensor<string, []>("op_1614_cast_fp16")];
            tensor<int32, [4]> value_states_99_perm_0 = const()[name = tensor<string, []>("value_states_99_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1617_transpose_x_0 = const()[name = tensor<string, []>("op_1617_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1617_transpose_y_0 = const()[name = tensor<string, []>("op_1617_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_131_perm_0 = const()[name = tensor<string, []>("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_132_perm_0 = const()[name = tensor<string, []>("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_132 = transpose(perm = transpose_132_perm_0, x = var_1611_cast_fp16)[name = tensor<string, []>("transpose_151")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_131 = transpose(perm = transpose_131_perm_0, x = var_1608_cast_fp16)[name = tensor<string, []>("transpose_152")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1617_cast_fp16 = matmul(transpose_x = var_1617_transpose_x_0, transpose_y = var_1617_transpose_y_0, x = transpose_131, y = transpose_132)[name = tensor<string, []>("op_1617_cast_fp16")];
            tensor<fp16, []> var_1618_to_fp16 = const()[name = tensor<string, []>("op_1618_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_97_cast_fp16 = mul(x = var_1617_cast_fp16, y = var_1618_to_fp16)[name = tensor<string, []>("attn_weights_97_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1620_cast_fp16 = softmax(axis = var_18, x = attn_weights_97_cast_fp16)[name = tensor<string, []>("op_1620_cast_fp16")];
            tensor<bool, []> attn_output_97_transpose_x_0 = const()[name = tensor<string, []>("attn_output_97_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_97_transpose_y_0 = const()[name = tensor<string, []>("attn_output_97_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_99_cast_fp16 = transpose(perm = value_states_99_perm_0, x = var_1614_cast_fp16)[name = tensor<string, []>("transpose_153")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_97_cast_fp16 = matmul(transpose_x = attn_output_97_transpose_x_0, transpose_y = attn_output_97_transpose_y_0, x = var_1620_cast_fp16, y = value_states_99_cast_fp16)[name = tensor<string, []>("attn_output_97_cast_fp16")];
            tensor<int32, [4]> var_1624_perm_0 = const()[name = tensor<string, []>("op_1624_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1626 = const()[name = tensor<string, []>("op_1626"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1624_cast_fp16 = transpose(perm = var_1624_perm_0, x = attn_output_97_cast_fp16)[name = tensor<string, []>("transpose_150")];
            tensor<fp16, [1, 1024, 1152]> input_343_cast_fp16 = reshape(shape = var_1626, x = var_1624_cast_fp16)[name = tensor<string, []>("input_343_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_24_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(752143104)))];
            tensor<fp16, [1152]> vpm_encoder_layers_24_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754797376)))];
            tensor<fp16, [1, 1024, 1152]> linear_147_cast_fp16 = linear(bias = vpm_encoder_layers_24_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_24_self_attn_out_proj_weight_to_fp16, x = input_343_cast_fp16)[name = tensor<string, []>("linear_147_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_345_cast_fp16 = add(x = input_339_cast_fp16, y = linear_147_cast_fp16)[name = tensor<string, []>("input_345_cast_fp16")];
            tensor<int32, [1]> input_347_axes_0 = const()[name = tensor<string, []>("input_347_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_24_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754799744)))];
            tensor<fp16, [1152]> vpm_encoder_layers_24_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754802112)))];
            tensor<fp16, [1, 1024, 1152]> input_347_cast_fp16 = layer_norm(axes = input_347_axes_0, beta = vpm_encoder_layers_24_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_24_layer_norm2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor<string, []>("input_347_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_24_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(754804480)))];
            tensor<fp16, [4304]> vpm_encoder_layers_24_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(764720960)))];
            tensor<fp16, [1, 1024, 4304]> linear_148_cast_fp16 = linear(bias = vpm_encoder_layers_24_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_24_mlp_fc1_weight_to_fp16, x = input_347_cast_fp16)[name = tensor<string, []>("linear_148_cast_fp16")];
            tensor<string, []> input_351_mode_0 = const()[name = tensor<string, []>("input_351_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_351_cast_fp16 = gelu(mode = input_351_mode_0, x = linear_148_cast_fp16)[name = tensor<string, []>("input_351_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_24_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(764729664)))];
            tensor<fp16, [1152]> vpm_encoder_layers_24_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_24_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774646144)))];
            tensor<fp16, [1, 1024, 1152]> linear_149_cast_fp16 = linear(bias = vpm_encoder_layers_24_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_24_mlp_fc2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("linear_149_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_353_cast_fp16 = add(x = input_345_cast_fp16, y = linear_149_cast_fp16)[name = tensor<string, []>("input_353_cast_fp16")];
            tensor<int32, [1]> hidden_states_151_axes_0 = const()[name = tensor<string, []>("hidden_states_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_25_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774648512)))];
            tensor<fp16, [1152]> vpm_encoder_layers_25_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774650880)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_151_cast_fp16 = layer_norm(axes = hidden_states_151_axes_0, beta = vpm_encoder_layers_25_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_25_layer_norm1_weight_to_fp16, x = input_353_cast_fp16)[name = tensor<string, []>("hidden_states_151_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_25_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774653248)))];
            tensor<fp16, [1152]> vpm_encoder_layers_25_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(777307520)))];
            tensor<fp16, [1, 1024, 1152]> linear_150_cast_fp16 = linear(bias = vpm_encoder_layers_25_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_25_self_attn_q_proj_weight_to_fp16, x = hidden_states_151_cast_fp16)[name = tensor<string, []>("linear_150_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_25_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(777309888)))];
            tensor<fp16, [1152]> vpm_encoder_layers_25_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(779964160)))];
            tensor<fp16, [1, 1024, 1152]> linear_151_cast_fp16 = linear(bias = vpm_encoder_layers_25_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_25_self_attn_k_proj_weight_to_fp16, x = hidden_states_151_cast_fp16)[name = tensor<string, []>("linear_151_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_25_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(779966528)))];
            tensor<fp16, [1152]> vpm_encoder_layers_25_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(782620800)))];
            tensor<fp16, [1, 1024, 1152]> linear_152_cast_fp16 = linear(bias = vpm_encoder_layers_25_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_25_self_attn_v_proj_weight_to_fp16, x = hidden_states_151_cast_fp16)[name = tensor<string, []>("linear_152_cast_fp16")];
            tensor<int32, [4]> var_1669 = const()[name = tensor<string, []>("op_1669"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1670_cast_fp16 = reshape(shape = var_1669, x = linear_150_cast_fp16)[name = tensor<string, []>("op_1670_cast_fp16")];
            tensor<int32, [4]> var_1672 = const()[name = tensor<string, []>("op_1672"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1673_cast_fp16 = reshape(shape = var_1672, x = linear_151_cast_fp16)[name = tensor<string, []>("op_1673_cast_fp16")];
            tensor<int32, [4]> var_1675 = const()[name = tensor<string, []>("op_1675"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1676_cast_fp16 = reshape(shape = var_1675, x = linear_152_cast_fp16)[name = tensor<string, []>("op_1676_cast_fp16")];
            tensor<int32, [4]> value_states_103_perm_0 = const()[name = tensor<string, []>("value_states_103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1679_transpose_x_0 = const()[name = tensor<string, []>("op_1679_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1679_transpose_y_0 = const()[name = tensor<string, []>("op_1679_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_133_perm_0 = const()[name = tensor<string, []>("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_134_perm_0 = const()[name = tensor<string, []>("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_134 = transpose(perm = transpose_134_perm_0, x = var_1673_cast_fp16)[name = tensor<string, []>("transpose_147")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_133 = transpose(perm = transpose_133_perm_0, x = var_1670_cast_fp16)[name = tensor<string, []>("transpose_148")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1679_cast_fp16 = matmul(transpose_x = var_1679_transpose_x_0, transpose_y = var_1679_transpose_y_0, x = transpose_133, y = transpose_134)[name = tensor<string, []>("op_1679_cast_fp16")];
            tensor<fp16, []> var_1680_to_fp16 = const()[name = tensor<string, []>("op_1680_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_101_cast_fp16 = mul(x = var_1679_cast_fp16, y = var_1680_to_fp16)[name = tensor<string, []>("attn_weights_101_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1682_cast_fp16 = softmax(axis = var_18, x = attn_weights_101_cast_fp16)[name = tensor<string, []>("op_1682_cast_fp16")];
            tensor<bool, []> attn_output_101_transpose_x_0 = const()[name = tensor<string, []>("attn_output_101_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_101_transpose_y_0 = const()[name = tensor<string, []>("attn_output_101_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_103_cast_fp16 = transpose(perm = value_states_103_perm_0, x = var_1676_cast_fp16)[name = tensor<string, []>("transpose_149")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_101_cast_fp16 = matmul(transpose_x = attn_output_101_transpose_x_0, transpose_y = attn_output_101_transpose_y_0, x = var_1682_cast_fp16, y = value_states_103_cast_fp16)[name = tensor<string, []>("attn_output_101_cast_fp16")];
            tensor<int32, [4]> var_1686_perm_0 = const()[name = tensor<string, []>("op_1686_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1688 = const()[name = tensor<string, []>("op_1688"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1686_cast_fp16 = transpose(perm = var_1686_perm_0, x = attn_output_101_cast_fp16)[name = tensor<string, []>("transpose_146")];
            tensor<fp16, [1, 1024, 1152]> input_357_cast_fp16 = reshape(shape = var_1688, x = var_1686_cast_fp16)[name = tensor<string, []>("input_357_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_25_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(782623168)))];
            tensor<fp16, [1152]> vpm_encoder_layers_25_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(785277440)))];
            tensor<fp16, [1, 1024, 1152]> linear_153_cast_fp16 = linear(bias = vpm_encoder_layers_25_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_25_self_attn_out_proj_weight_to_fp16, x = input_357_cast_fp16)[name = tensor<string, []>("linear_153_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_359_cast_fp16 = add(x = input_353_cast_fp16, y = linear_153_cast_fp16)[name = tensor<string, []>("input_359_cast_fp16")];
            tensor<int32, [1]> input_361_axes_0 = const()[name = tensor<string, []>("input_361_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_25_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(785279808)))];
            tensor<fp16, [1152]> vpm_encoder_layers_25_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(785282176)))];
            tensor<fp16, [1, 1024, 1152]> input_361_cast_fp16 = layer_norm(axes = input_361_axes_0, beta = vpm_encoder_layers_25_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_25_layer_norm2_weight_to_fp16, x = input_359_cast_fp16)[name = tensor<string, []>("input_361_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_25_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(785284544)))];
            tensor<fp16, [4304]> vpm_encoder_layers_25_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795201024)))];
            tensor<fp16, [1, 1024, 4304]> linear_154_cast_fp16 = linear(bias = vpm_encoder_layers_25_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_25_mlp_fc1_weight_to_fp16, x = input_361_cast_fp16)[name = tensor<string, []>("linear_154_cast_fp16")];
            tensor<string, []> input_365_mode_0 = const()[name = tensor<string, []>("input_365_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_365_cast_fp16 = gelu(mode = input_365_mode_0, x = linear_154_cast_fp16)[name = tensor<string, []>("input_365_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_25_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(795209728)))];
            tensor<fp16, [1152]> vpm_encoder_layers_25_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_25_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805126208)))];
            tensor<fp16, [1, 1024, 1152]> linear_155_cast_fp16 = linear(bias = vpm_encoder_layers_25_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_25_mlp_fc2_weight_to_fp16, x = input_365_cast_fp16)[name = tensor<string, []>("linear_155_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_367_cast_fp16 = add(x = input_359_cast_fp16, y = linear_155_cast_fp16)[name = tensor<string, []>("input_367_cast_fp16")];
            tensor<int32, [1]> hidden_states_157_axes_0 = const()[name = tensor<string, []>("hidden_states_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_26_layer_norm1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_layer_norm1_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805128576)))];
            tensor<fp16, [1152]> vpm_encoder_layers_26_layer_norm1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_layer_norm1_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805130944)))];
            tensor<fp16, [1, 1024, 1152]> hidden_states_157_cast_fp16 = layer_norm(axes = hidden_states_157_axes_0, beta = vpm_encoder_layers_26_layer_norm1_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_26_layer_norm1_weight_to_fp16, x = input_367_cast_fp16)[name = tensor<string, []>("hidden_states_157_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_26_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(805133312)))];
            tensor<fp16, [1152]> vpm_encoder_layers_26_self_attn_q_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(807787584)))];
            tensor<fp16, [1, 1024, 1152]> linear_156_cast_fp16 = linear(bias = vpm_encoder_layers_26_self_attn_q_proj_bias_to_fp16, weight = vpm_encoder_layers_26_self_attn_q_proj_weight_to_fp16, x = hidden_states_157_cast_fp16)[name = tensor<string, []>("linear_156_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_26_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(807789952)))];
            tensor<fp16, [1152]> vpm_encoder_layers_26_self_attn_k_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(810444224)))];
            tensor<fp16, [1, 1024, 1152]> linear_157_cast_fp16 = linear(bias = vpm_encoder_layers_26_self_attn_k_proj_bias_to_fp16, weight = vpm_encoder_layers_26_self_attn_k_proj_weight_to_fp16, x = hidden_states_157_cast_fp16)[name = tensor<string, []>("linear_157_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_26_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(810446592)))];
            tensor<fp16, [1152]> vpm_encoder_layers_26_self_attn_v_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(813100864)))];
            tensor<fp16, [1, 1024, 1152]> linear_158_cast_fp16 = linear(bias = vpm_encoder_layers_26_self_attn_v_proj_bias_to_fp16, weight = vpm_encoder_layers_26_self_attn_v_proj_weight_to_fp16, x = hidden_states_157_cast_fp16)[name = tensor<string, []>("linear_158_cast_fp16")];
            tensor<int32, [4]> var_1731 = const()[name = tensor<string, []>("op_1731"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1732_cast_fp16 = reshape(shape = var_1731, x = linear_156_cast_fp16)[name = tensor<string, []>("op_1732_cast_fp16")];
            tensor<int32, [4]> var_1734 = const()[name = tensor<string, []>("op_1734"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1735_cast_fp16 = reshape(shape = var_1734, x = linear_157_cast_fp16)[name = tensor<string, []>("op_1735_cast_fp16")];
            tensor<int32, [4]> var_1737 = const()[name = tensor<string, []>("op_1737"), val = tensor<int32, [4]>([1, 1024, 16, 72])];
            tensor<fp16, [1, 1024, 16, 72]> var_1738_cast_fp16 = reshape(shape = var_1737, x = linear_158_cast_fp16)[name = tensor<string, []>("op_1738_cast_fp16")];
            tensor<int32, [4]> value_states_perm_0 = const()[name = tensor<string, []>("value_states_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<bool, []> var_1741_transpose_x_0 = const()[name = tensor<string, []>("op_1741_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1741_transpose_y_0 = const()[name = tensor<string, []>("op_1741_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_135_perm_0 = const()[name = tensor<string, []>("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_136_perm_0 = const()[name = tensor<string, []>("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 16, 72, 1024]> transpose_136 = transpose(perm = transpose_136_perm_0, x = var_1735_cast_fp16)[name = tensor<string, []>("transpose_143")];
            tensor<fp16, [1, 16, 1024, 72]> transpose_135 = transpose(perm = transpose_135_perm_0, x = var_1732_cast_fp16)[name = tensor<string, []>("transpose_144")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1741_cast_fp16 = matmul(transpose_x = var_1741_transpose_x_0, transpose_y = var_1741_transpose_y_0, x = transpose_135, y = transpose_136)[name = tensor<string, []>("op_1741_cast_fp16")];
            tensor<fp16, []> var_1742_to_fp16 = const()[name = tensor<string, []>("op_1742_to_fp16"), val = tensor<fp16, []>(0x1.e2cp-4)];
            tensor<fp16, [1, 16, 1024, 1024]> attn_weights_105_cast_fp16 = mul(x = var_1741_cast_fp16, y = var_1742_to_fp16)[name = tensor<string, []>("attn_weights_105_cast_fp16")];
            tensor<fp16, [1, 16, 1024, 1024]> var_1744_cast_fp16 = softmax(axis = var_18, x = attn_weights_105_cast_fp16)[name = tensor<string, []>("op_1744_cast_fp16")];
            tensor<bool, []> attn_output_105_transpose_x_0 = const()[name = tensor<string, []>("attn_output_105_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_105_transpose_y_0 = const()[name = tensor<string, []>("attn_output_105_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 16, 1024, 72]> value_states_cast_fp16 = transpose(perm = value_states_perm_0, x = var_1738_cast_fp16)[name = tensor<string, []>("transpose_145")];
            tensor<fp16, [1, 16, 1024, 72]> attn_output_105_cast_fp16 = matmul(transpose_x = attn_output_105_transpose_x_0, transpose_y = attn_output_105_transpose_y_0, x = var_1744_cast_fp16, y = value_states_cast_fp16)[name = tensor<string, []>("attn_output_105_cast_fp16")];
            tensor<int32, [4]> var_1748_perm_0 = const()[name = tensor<string, []>("op_1748_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1750 = const()[name = tensor<string, []>("op_1750"), val = tensor<int32, [3]>([1, 1024, 1152])];
            tensor<fp16, [1, 1024, 16, 72]> var_1748_cast_fp16 = transpose(perm = var_1748_perm_0, x = attn_output_105_cast_fp16)[name = tensor<string, []>("transpose_142")];
            tensor<fp16, [1, 1024, 1152]> input_371_cast_fp16 = reshape(shape = var_1750, x = var_1748_cast_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
            tensor<fp16, [1152, 1152]> vpm_encoder_layers_26_self_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_self_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [1152, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(813103232)))];
            tensor<fp16, [1152]> vpm_encoder_layers_26_self_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815757504)))];
            tensor<fp16, [1, 1024, 1152]> linear_159_cast_fp16 = linear(bias = vpm_encoder_layers_26_self_attn_out_proj_bias_to_fp16, weight = vpm_encoder_layers_26_self_attn_out_proj_weight_to_fp16, x = input_371_cast_fp16)[name = tensor<string, []>("linear_159_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_373_cast_fp16 = add(x = input_367_cast_fp16, y = linear_159_cast_fp16)[name = tensor<string, []>("input_373_cast_fp16")];
            tensor<int32, [1]> input_375_axes_0 = const()[name = tensor<string, []>("input_375_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_encoder_layers_26_layer_norm2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_layer_norm2_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815759872)))];
            tensor<fp16, [1152]> vpm_encoder_layers_26_layer_norm2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_layer_norm2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815762240)))];
            tensor<fp16, [1, 1024, 1152]> input_375_cast_fp16 = layer_norm(axes = input_375_axes_0, beta = vpm_encoder_layers_26_layer_norm2_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_encoder_layers_26_layer_norm2_weight_to_fp16, x = input_373_cast_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
            tensor<fp16, [4304, 1152]> vpm_encoder_layers_26_mlp_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_mlp_fc1_weight_to_fp16"), val = tensor<fp16, [4304, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(815764608)))];
            tensor<fp16, [4304]> vpm_encoder_layers_26_mlp_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_mlp_fc1_bias_to_fp16"), val = tensor<fp16, [4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(825681088)))];
            tensor<fp16, [1, 1024, 4304]> linear_160_cast_fp16 = linear(bias = vpm_encoder_layers_26_mlp_fc1_bias_to_fp16, weight = vpm_encoder_layers_26_mlp_fc1_weight_to_fp16, x = input_375_cast_fp16)[name = tensor<string, []>("linear_160_cast_fp16")];
            tensor<string, []> input_379_mode_0 = const()[name = tensor<string, []>("input_379_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1024, 4304]> input_379_cast_fp16 = gelu(mode = input_379_mode_0, x = linear_160_cast_fp16)[name = tensor<string, []>("input_379_cast_fp16")];
            tensor<fp16, [1152, 4304]> vpm_encoder_layers_26_mlp_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_mlp_fc2_weight_to_fp16"), val = tensor<fp16, [1152, 4304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(825689792)))];
            tensor<fp16, [1152]> vpm_encoder_layers_26_mlp_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vpm_encoder_layers_26_mlp_fc2_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(835606272)))];
            tensor<fp16, [1, 1024, 1152]> linear_161_cast_fp16 = linear(bias = vpm_encoder_layers_26_mlp_fc2_bias_to_fp16, weight = vpm_encoder_layers_26_mlp_fc2_weight_to_fp16, x = input_379_cast_fp16)[name = tensor<string, []>("linear_161_cast_fp16")];
            tensor<fp16, [1, 1024, 1152]> input_381_cast_fp16 = add(x = input_373_cast_fp16, y = linear_161_cast_fp16)[name = tensor<string, []>("input_381_cast_fp16")];
            tensor<int32, [1]> x_1_axes_0 = const()[name = tensor<string, []>("x_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1152]> vpm_post_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("vpm_post_layernorm_weight_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(835608640)))];
            tensor<fp16, [1152]> vpm_post_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("vpm_post_layernorm_bias_to_fp16"), val = tensor<fp16, [1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(835611008)))];
            tensor<fp16, [1, 1024, 1152]> x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = vpm_post_layernorm_bias_to_fp16, epsilon = var_12_to_fp16, gamma = vpm_post_layernorm_weight_to_fp16, x = input_381_cast_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
            tensor<int32, []> var_1777 = const()[name = tensor<string, []>("op_1777"), val = tensor<int32, []>(-1)];
            tensor<fp16, [2560, 1152]> resampler_kv_proj_weight_to_fp16 = const()[name = tensor<string, []>("resampler_kv_proj_weight_to_fp16"), val = tensor<fp16, [2560, 1152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(835613376)))];
            tensor<fp16, [2560]> linear_162_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_162_bias_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(841511680)))];
            tensor<fp16, [1, 1024, 2560]> linear_162_cast_fp16 = linear(bias = linear_162_bias_0_to_fp16, weight = resampler_kv_proj_weight_to_fp16, x = x_1_cast_fp16)[name = tensor<string, []>("linear_162_cast_fp16")];
            tensor<int32, [1]> var_1798_axes_0 = const()[name = tensor<string, []>("op_1798_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2560]> resampler_ln_kv_weight_to_fp16 = const()[name = tensor<string, []>("resampler_ln_kv_weight_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(841516864)))];
            tensor<fp16, [2560]> resampler_ln_kv_bias_to_fp16 = const()[name = tensor<string, []>("resampler_ln_kv_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(841522048)))];
            tensor<fp16, []> var_1781_to_fp16 = const()[name = tensor<string, []>("op_1781_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1024, 2560]> var_1798_cast_fp16 = layer_norm(axes = var_1798_axes_0, beta = resampler_ln_kv_bias_to_fp16, epsilon = var_1781_to_fp16, gamma = resampler_ln_kv_weight_to_fp16, x = linear_162_cast_fp16)[name = tensor<string, []>("op_1798_cast_fp16")];
            tensor<int32, [3]> var_1799 = const()[name = tensor<string, []>("op_1799"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<string, []> pos_embed_to_fp16_dtype_0 = const()[name = tensor<string, []>("pos_embed_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1024, 1, 2560]> pos_embed_to_fp16 = cast(dtype = pos_embed_to_fp16_dtype_0, x = pos_embed)[name = tensor<string, []>("cast_149")];
            tensor<fp16, [1024, 1, 2560]> x_3_cast_fp16 = transpose(perm = var_1799, x = var_1798_cast_fp16)[name = tensor<string, []>("transpose_141")];
            tensor<fp16, [1024, 1, 2560]> key_cast_fp16 = add(x = x_3_cast_fp16, y = pos_embed_to_fp16)[name = tensor<string, []>("key_cast_fp16")];
            tensor<fp16, [2560, 2560]> var_1822_1_to_fp16 = const()[name = tensor<string, []>("op_1822_1_to_fp16"), val = tensor<fp16, [2560, 2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(841527232)))];
            tensor<fp16, [2560]> var_1825_1_to_fp16 = const()[name = tensor<string, []>("op_1825_1_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(854634496)))];
            tensor<fp16, [1024, 1, 2560]> linear_164_cast_fp16 = linear(bias = var_1825_1_to_fp16, weight = var_1822_1_to_fp16, x = key_cast_fp16)[name = tensor<string, []>("linear_164_cast_fp16")];
            tensor<fp16, [2560, 2560]> var_1822_2_to_fp16 = const()[name = tensor<string, []>("op_1822_2_to_fp16"), val = tensor<fp16, [2560, 2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(854639680)))];
            tensor<fp16, [2560]> var_1825_2_to_fp16 = const()[name = tensor<string, []>("op_1825_2_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(867746944)))];
            tensor<fp16, [1024, 1, 2560]> linear_165_cast_fp16 = linear(bias = var_1825_2_to_fp16, weight = var_1822_2_to_fp16, x = x_3_cast_fp16)[name = tensor<string, []>("linear_165_cast_fp16")];
            tensor<int32, [3]> var_1841 = const()[name = tensor<string, []>("op_1841"), val = tensor<int32, [3]>([1024, 20, 128])];
            tensor<fp16, [1024, 20, 128]> var_1842_cast_fp16 = reshape(shape = var_1841, x = linear_164_cast_fp16)[name = tensor<string, []>("op_1842_cast_fp16")];
            tensor<int32, [3]> transpose_28_perm_0 = const()[name = tensor<string, []>("transpose_28_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [3]> var_1848 = const()[name = tensor<string, []>("op_1848"), val = tensor<int32, [3]>([1024, 20, 128])];
            tensor<fp16, [1024, 20, 128]> var_1849_cast_fp16 = reshape(shape = var_1848, x = linear_165_cast_fp16)[name = tensor<string, []>("op_1849_cast_fp16")];
            tensor<int32, [3]> v_perm_0 = const()[name = tensor<string, []>("v_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<bool, []> input_387_transpose_x_0 = const()[name = tensor<string, []>("input_387_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> input_387_transpose_y_0 = const()[name = tensor<string, []>("input_387_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [20, 64, 128]> q_scaled_to_fp16 = const()[name = tensor<string, []>("q_scaled_to_fp16"), val = tensor<fp16, [20, 64, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(867752128)))];
            tensor<fp16, [20, 128, 1024]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = var_1842_cast_fp16)[name = tensor<string, []>("transpose_140")];
            tensor<fp16, [20, 64, 1024]> input_387_cast_fp16 = matmul(transpose_x = input_387_transpose_x_0, transpose_y = input_387_transpose_y_0, x = q_scaled_to_fp16, y = transpose_28_cast_fp16)[name = tensor<string, []>("input_387_cast_fp16")];
            tensor<fp16, [20, 64, 1024]> attn_output_weights_1_cast_fp16 = softmax(axis = var_1777, x = input_387_cast_fp16)[name = tensor<string, []>("attn_output_weights_1_cast_fp16")];
            tensor<bool, []> attn_output_109_transpose_x_0 = const()[name = tensor<string, []>("attn_output_109_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_109_transpose_y_0 = const()[name = tensor<string, []>("attn_output_109_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [20, 1024, 128]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_1849_cast_fp16)[name = tensor<string, []>("transpose_139")];
            tensor<fp16, [20, 64, 128]> attn_output_109_cast_fp16 = matmul(transpose_x = attn_output_109_transpose_x_0, transpose_y = attn_output_109_transpose_y_0, x = attn_output_weights_1_cast_fp16, y = v_cast_fp16)[name = tensor<string, []>("attn_output_109_cast_fp16")];
            tensor<int32, [3]> var_1857_perm_0 = const()[name = tensor<string, []>("op_1857_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [2]> var_1861 = const()[name = tensor<string, []>("op_1861"), val = tensor<int32, [2]>([64, 2560])];
            tensor<fp16, [64, 20, 128]> var_1857_cast_fp16 = transpose(perm = var_1857_perm_0, x = attn_output_109_cast_fp16)[name = tensor<string, []>("transpose_138")];
            tensor<fp16, [64, 2560]> input_389_cast_fp16 = reshape(shape = var_1861, x = var_1857_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
            tensor<fp16, [2560, 2560]> resampler_attn_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("resampler_attn_out_proj_weight_to_fp16"), val = tensor<fp16, [2560, 2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(868079872)))];
            tensor<fp16, [2560]> resampler_attn_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("resampler_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(881187136)))];
            tensor<fp16, [64, 2560]> linear_166_cast_fp16 = linear(bias = resampler_attn_out_proj_bias_to_fp16, weight = resampler_attn_out_proj_weight_to_fp16, x = input_389_cast_fp16)[name = tensor<string, []>("linear_166_cast_fp16")];
            tensor<int32, [3]> var_1867 = const()[name = tensor<string, []>("op_1867"), val = tensor<int32, [3]>([64, 1, 2560])];
            tensor<fp16, [64, 1, 2560]> out_cast_fp16 = reshape(shape = var_1867, x = linear_166_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
            tensor<int32, [3]> var_1869 = const()[name = tensor<string, []>("op_1869"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<int32, [1]> x_axes_0 = const()[name = tensor<string, []>("x_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2560]> resampler_ln_post_weight_to_fp16 = const()[name = tensor<string, []>("resampler_ln_post_weight_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(881192320)))];
            tensor<fp16, [2560]> resampler_ln_post_bias_to_fp16 = const()[name = tensor<string, []>("resampler_ln_post_bias_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(881197504)))];
            tensor<fp16, [1, 64, 2560]> input_cast_fp16 = transpose(perm = var_1869, x = out_cast_fp16)[name = tensor<string, []>("transpose_137")];
            tensor<fp16, [1, 64, 2560]> x_cast_fp16 = layer_norm(axes = x_axes_0, beta = resampler_ln_post_bias_to_fp16, epsilon = var_1781_to_fp16, gamma = resampler_ln_post_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
            tensor<fp16, [2560, 2560]> transpose_0_to_fp16 = const()[name = tensor<string, []>("transpose_0_to_fp16"), val = tensor<fp16, [2560, 2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(881202688)))];
            tensor<fp16, [1, 64, 2560]> var_1875_cast_fp16 = linear(bias = linear_162_bias_0_to_fp16, weight = transpose_0_to_fp16, x = x_cast_fp16)[name = tensor<string, []>("op_1875_cast_fp16")];
            tensor<string, []> var_1875_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1875_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 64, 2560]> output = cast(dtype = var_1875_cast_fp16_to_fp32_dtype_0, x = var_1875_cast_fp16)[name = tensor<string, []>("cast_148")];
        } -> (output);
}