the output of onnx model is different from model inferenced by TensorRT

Description

I attempted to compile a Hugging Face model (the Hugging Face model link is: https://huggingface.co/OpenGVLab/InternViT-6B-448px-V1-5, which includes both the model architecture code and model files) using TensorRT (TRT) to improve inference speed. The steps I followed are hf -> onnx -> trt.

I performed inference on the same image using Hugging Face (hf), ONNX, and TRT engine. I found that the inference results from hf and ONNX were consistent, but the TRT engine's result was different from the former two.

I would like to know why the ONNX results are correct, but the inference results from the engine compiled with trtexec are wrong. Why is this happening?

The conversion code from hf to ONNX is:

import torch
from transformers import AutoModel
from typing import List,Dict
import time
import os
import warnings


class InternVL2VisionWrapper(torch.nn.Module):

    def __init__(self, model, ):
        super().__init__()
        self.model = model

    def forward(self, pixel_values):
        features = self.model(pixel_values).last_hidden_state
        return features


model_path="/media/star/disk2/pretrained_model/InternViT/InternViT-6B-448px-V1-5"
output_dir="/media/star/8T/tmp/InternViT-6B-448px-V1-5/onnx"
os.makedirs(f'{output_dir}', exist_ok=True)
model = AutoModel.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True,
    trust_remote_code=True).cuda().eval()

device = torch.device("cuda") if torch.cuda.is_available() else "cpu"
wrapper = InternVL2VisionWrapper(model=model.to(device))
dummy_image = torch.empty(1,
                          3,
                          448,
                          448,
                          dtype=torch.float16,
                          device=device)  # dummy image

def export_visual_wrapper_onnx(visual_wrapper,
                               input,
                               output_dir,
                               input_names=['input'],
                               dynamic_axes={'input': {0: 'batch'}, 'output': {0: 'batch'}}

                               ):

    torch.onnx.export(visual_wrapper,
                      input,
                      f'{output_dir}/visual_encoder.onnx',
                      opset_version=17,
                      input_names=input_names,
                      output_names=['output'],
                      dynamic_axes=dynamic_axes)

export_visual_wrapper_onnx(wrapper, dummy_image, output_dir)

The conversion code from ONNX to TRT engine is:

MODEL_NAME="InternVL2-40B"
OUTPUT_MODEL_NAME="InternVL2_40B"
onnx_process_version="onnx_v1"
max_batch_size=24
onnx_dtype="float16"
trt_dtype="best"

/usr/src/tensorrt/bin/trtexec \
--onnx=/data/eas/visual_engine/a100/InternViT-6B-448px-V1-5/onnx/visual_encoder.onnx \
--saveEngine=/data/eas/visual_engine/a100/InternViT-6B-448px-V1-5/visual_encoder.trtexec.${trt_dtype}.maxBatchSize${max_batch_size}.engine \
--minShapes=input:1x3x448x448 \
--optShapes=input:8x3x448x448 \
--maxShapes=input:24x3x448x448 \
--best

The inference code for hf is:

import torch
import time
from transformers import AutoModel, CLIPImageProcessor
import numpy as np

model_path="/media/star/disk2/pretrained_model/InternViT/InternViT-6B-448px-V1-5"
torch_dtype=torch.float16
model = AutoModel.from_pretrained(
    model_path,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
    trust_remote_code=True).cuda().eval()

pixel_values = torch.tensor(np.load("/media/star/8T/tmp/internvn2_40b_image2_patch1.npy")).to(torch_dtype).cuda()
print(f"pixel_values.shape={pixel_values.shape}")

pixel_values=torch.concat([pixel_values], dim=0)

start_time=time.time()
outputs = model(pixel_values)
print(f"run time is {time.time()-start_time} seconds")
print(f"outputs.last_hidden_state={outputs.last_hidden_state}")
print(f"outputs.last_hidden_state.shape={outputs.last_hidden_state.shape}")

The inference code for ONNX is:

import os
import torch
import numpy as np
import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode
from PIL import Image
import time
import onnxruntime
from transformers import CLIPImageProcessor



if __name__ == '__main__':
    img_sess_options = onnxruntime.SessionOptions()
    img_run_options = onnxruntime.RunOptions()
    img_run_options.log_severity_level = 2

    img_onnx_model_path = "/media/star/8T/tmp/InternViT-6B-448px-V1-5/onnx/visual_encoder.onnx"

    pixel_values = torch.tensor(np.load("/media/star/8T/tmp/internvn2_40b_image2_patch1.npy")).to(torch.float16)

    print(f"pixel_values.shape={pixel_values.shape}")

    img_session = onnxruntime.InferenceSession(img_onnx_model_path,
                                            sess_options=img_sess_options,
                                            providers=["CUDAExecutionProvider"])
    onnx_image_features = img_session.run(["output"], {"input": pixel_values.cpu().numpy()})[
        0]
    onnx_image_features = torch.tensor(onnx_image_features).cuda()
    print(f"onnx_image_features.shape={onnx_image_features.shape}")
    print(f"onnx_image_features={onnx_image_features}")

The inference code for TRT engine is:

import tensorrt_llm
from tensorrt_llm import logger
from tensorrt_llm._utils import str_dtype_to_trt
from tensorrt_llm.runtime import Session, TensorInfo
import os
import torch
import tensorrt as trt
import numpy as np
import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode
from PIL import Image
import time
from transformers import AutoModel,AutoTokenizer
import math


def trt_dtype_to_torch(dtype):
    if dtype == trt.float16:
        return torch.float16
    elif dtype == trt.float32:
        return torch.float32
    elif dtype == trt.int32:
        return torch.int32
    elif dtype == trt.bfloat16:
        return torch.bfloat16
    else:
        raise TypeError("%s is not supported" % dtype)

vision_precision="float16"
if vision_precision=="float16":
    torch_dtype = torch.float16
else:
    torch_dtype = torch.float

start_time=time.time()

image=torch.tensor(np.load("/data/eas/ndarray/internvn2_40b_image2_patch1.npy")).to(torch_dtype).cuda()
print(f"load image time is {time.time()-start_time} seconds")

# print(f"image pixels={image}")

print(f"origin image.shape={image.shape}")
image=torch.concat([image]*1, dim=0)
print(f"new image.shape={image.shape}")
iter_num=1

#=======================================================================================================================
attention_mask=None

stream = torch.cuda.Stream(torch.cuda.current_device())
torch.cuda.set_stream(stream)

vision_encoder_path="/data/eas/visual_engine/a100/InternViT-6B-448px-V1-5/visual_encoder.trtexec.best.maxBatchSize24.engine"
logger.info(f'Loading engine from {vision_encoder_path}')


with open(vision_encoder_path, 'rb') as f:
    engine_buffer = f.read()
logger.info(f'Creating session from engine {vision_encoder_path}')
visual_encoder_session = Session.from_serialized_engine(
    engine_buffer)


visual_features = {'input':image.to(tensorrt_llm._utils.str_dtype_to_torch(vision_precision))}
if attention_mask is not None:
    visual_features['attention_mask'] = attention_mask
tensor_info = [TensorInfo('input', str_dtype_to_trt(vision_precision),image.shape)]

if attention_mask is not None:
    tensor_info.append(TensorInfo('attention_mask', trt.DataType.INT32,attention_mask.shape))

visual_output_info = visual_encoder_session.infer_shapes(tensor_info)
visual_outputs = {
    t.name: torch.empty(tuple(t.shape),
                        dtype=trt_dtype_to_torch(t.dtype),
                        device=image.device)
    for t in visual_output_info
}




start_time=time.time()
for _ in range(iter_num):
    ok = visual_encoder_session.run(visual_features, visual_outputs, stream.cuda_stream)
assert ok, "Runtime execution failed for vision encoder session"
stream.synchronize()
print(f"tensorrt,single image run time is {round((time.time()-start_time)/(iter_num),2)} seconds")


outputs_trt = visual_outputs['output']
outputs_trt=torch.squeeze(outputs_trt)
print(f"outputs_trt.shape={outputs_trt.shape}")
print(f"outputs_trt={outputs_trt}")

The inference results are as follows:

InternViT-6B-448px-V1-5
hf
outputs.last_hidden_state=tensor([[[  1.0576,  -4.4062,   1.1816,  ...,   0.4963,   0.5752,   0.4436],
         [  3.6680,   4.8086,   4.7578,  ..., -14.2969,   6.4336, -12.0312],
         [  3.9355,   4.4805,   4.4922,  ..., -14.7031,   6.2812, -11.2266],
         ...,
         [ -2.5684,  -2.8164,   5.3242,  ...,  -7.0508,   0.2556,  -6.5859],
         [ -6.5156,  -6.5859,   9.9531,  ...,  -4.0938,  -4.5703, -14.6719],
         [ -6.2383,  -6.2930,  10.0391,  ...,  -3.8965,  -4.2891, -15.1016]]],
       device='cuda:0', dtype=torch.float16, grad_fn=<AddBackward0>)
outputs.last_hidden_state.shape=torch.Size([1, 1025, 3200])

onnx
onnx_image_features=tensor([[[  1.0508,  -4.3672,   1.1572,  ...,   0.4902,   0.5601,   0.4417],
         [  3.6680,   4.8086,   4.7539,  ..., -14.2969,   6.4180, -12.0391],
         [  3.9316,   4.4805,   4.4805,  ..., -14.6953,   6.2656, -11.2188],
         ...,
         [ -2.5840,  -2.8320,   5.3555,  ...,  -7.0664,   0.2539,  -6.6055],
         [ -6.5195,  -6.5898,   9.9375,  ...,  -4.0977,  -4.5781, -14.6641],
         [ -6.2461,  -6.2891,  10.0234,  ...,  -3.9062,  -4.2930, -15.0938]]],
       device='cuda:0', dtype=torch.float16)


trt_exec
outputs_trt=tensor([[  1.6309,  -3.6855,   2.2578,  ...,   0.0575,   1.1006,  -0.1124],
        [  4.1562,   3.1484,   5.3594,  ..., -10.8594,   6.0859, -10.5625],
        [  4.8984,   2.9824,   5.1719,  ..., -11.0938,   6.1602,  -9.5312],
        ...,
        [ -0.2856,  -3.4414,   5.4531,  ...,  -3.5000,   0.0735,  -5.1602],
        [ -3.1660,  -5.7734,   8.8047,  ...,  -4.5742,  -3.1387, -10.2969],
        [ -3.0371,  -5.5156,   8.8984,  ...,  -4.5469,  -3.0234, -10.8906]],
       device='cuda:0', dtype=torch.float16)

Environment

TensorRT Version:v100500

NVIDIA GPU:A100

NVIDIA Driver Version:535.54.03

CUDA Version:12.2

CUDNN Version:8920

Operating System: docker mirror: nvidia_cuda_12.4.0-devel-ubuntu22.04

Python Version (if applicable):3.10.12

Tensorflow Version (if applicable):

PyTorch Version (if applicable):2.2.2+cu121

Baremetal or Container (if so, version):docker

Relevant Files

Model link: https://huggingface.co/OpenGVLab/InternViT-6B-448px-V1-5
internvn2_40b_image2_patch1.npy: internvn2_40b_image2_patch1.zip
onnx file link: https://drive.google.com/file/d/1lnEmuQ4cNzf8YA7ddznqUnYsz-W5y5aJ/view?usp=sharing

Start with the linked InternViT-6B-448px-V1-5 model, the provided .npy input, and the ONNX file, then reproduce the three paths shown: Hugging Face, ONNX Runtime, and the trtexec-built engine. Compare the outputs and the TensorRT engine configuration around the trtexec entry point; done means identifying why the TensorRT result diverges or documenting the required reproducible fix.