Tensorrt8.5 inference different with origin onnx model


i have trained a model with pytorch and export it to onnx format model, this onnx model predicts correct value. but when i convert it to tensorrt engine, it gives wrong inference.


TensorRT Version:
GPU Type: NVIDIA GeForce RTX 3090
Nvidia Driver Version: 470.129.06
CUDA Version: cuda_11.8.r11.8
CUDNN Version: 8
Operating System + Version: Ubuntu 20.04.5 LTS
Python Version (if applicable): 3.8.10
TensorFlow Version (if applicable):
PyTorch Version (if applicable): 1.12.1
Baremetal or Container (if container which image + tag):

Hi @liwenjudetiankong ,
Can you please share the onnx model and reproducible script with us, so that we can try reproducing.

the model is here: Tensorrt8.5 inference different with origin onnx model


Could you please share with us the issue repro script as well to try from our end for better debugging.

Thank you.

from transformers import BertTokenizerFast

import os 
import os
import json 
import time 
import threading
from typing import Tuple, List 
import numpy as np 

    import pycuda.driver as cuda 
    import tensorrt as trt
    TRT_LOGGER = trt.Logger(trt.Logger.VERBOSE)
except ImportError:
    import onnxruntime

tokenizer = BertTokenizerFast.from_pretrained("hfl/chinese-macbert-base")
def preprocess_data(text="this is a sad thing", is_trt=False):
    texts = [text for _ in range(5)]
    context = tokenizer(texts, padding="max_length", return_tensors='pt',max_length=128, truncation=True, return_offsets_mapping=True)
    input_ids = context['input_ids'].detach().cpu().numpy()
    attention_mask = context['attention_mask'].detach().cpu().numpy()
    token_type_ids = context['token_type_ids'].detach().cpu().numpy()
    if is_trt:
        return  [input_ids, attention_mask, token_type_ids]
        return  {"input_ids": input_ids, "attention_mask": attention_mask, "token_type_ids": token_type_ids}

class TrtModel:
    def __init__(self, model_name="detector_corrector", model_dir=".", cached_engine=True, max_batch_size=5) -> None:
        self.cfx = cuda.Device(0).make_context()
        self.model_dir = model_dir
        self.model_name = model_name
        self.max_batch_size = max_batch_size
        self.catched_engine = cached_engine
        self.engine = self.load_model(
            os.path.join(model_dir, model_name + ".onnx"))
        self.input_binding_idxs, self.output_binding_idxs = self.get_binding_idxs()
        self.input_names = [self.engine.get_binding_name(
            binding_idx) for binding_idx in self.input_binding_idxs]
        self.output_names = [self.engine.get_binding_name(
            binding_idx) for binding_idx in self.output_binding_idxs]

    def __del__(self):

    def load_model(self, model_path):
        return self.load_engine(model_path)

    def get_context(self):
        return self.engine.create_execution_context()

    def get_stream(self):
        return cuda.Stream()

    def predict(self, host_inputs):
        context = self.get_context()
        stream = self.get_stream()
        device_inputs = [cuda.mem_alloc(h_input.nbytes)
                         for h_input in host_inputs]
        for h_input, d_input in zip(host_inputs, device_inputs):
            cuda.memcpy_htod_async(d_input, h_input, stream)

        host_outputs, device_outputs = self.gen_output_buffer(
            host_inputs, context)
        bindings = device_inputs + device_outputs
        exe_res = context.execute_async_v2(
            bindings=bindings, stream_handle=stream.handle)
        if not exe_res:
            print(f"{self.__class__.__name__} execute_async_v2 error")
        for h_output, d_output in zip(host_outputs, device_outputs):
            cuda.memcpy_dtoh_async(h_output, d_output, stream)


        for b in bindings:

        return host_outputs

    def gen_output_buffer(self, host_inputs: List[np.ndarray], context):

        assert context.all_binding_shapes_specified

        host_outputs = []
        device_outputs = []
        for binding_index in self.output_binding_idxs:
            output_shape = context.get_binding_shape(binding_index)
            # Allocate buffers to hold output results after copying back to host
            buffer = np.empty(output_shape, dtype=np.float32)
            # Allocate output buffers on device

        return host_outputs, device_outputs

    def get_binding_idxs(self):
        # Separate input and output binding indices for convenience
        input_binding_idxs = []
        output_binding_idxs = []
        for binding_index in range(0, self.engine.num_bindings):
            if self.engine.binding_is_input(binding_index):

        return input_binding_idxs, output_binding_idxs

    def load_engine(self, onnx_file_path):
        runtime = trt.Runtime(TRT_LOGGER)
        cached_engine_path = os.path.join(
            self.model_dir, self.model_name + ".engine")
        if self.catched_engine and os.path.exists(cached_engine_path):
            with open(cached_engine_path, "rb") as f:
                serialized_engine = f.read()
                engine = runtime.deserialize_cuda_engine(serialized_engine)
                print(f"load engine from cache: {cached_engine_path} sucessfully")
                return engine

        EXPLICIT_BATCH = 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)

        with trt.Builder(TRT_LOGGER) as builder, builder.create_network(EXPLICIT_BATCH) as network, trt.OnnxParser(network, TRT_LOGGER) as parser:
            builder.max_batch_size = self.max_batch_size
            with open(onnx_file_path, 'rb') as model:
                if not parser.parse(model.read()):
                    for error in range(parser.num_errors):
            config = builder.create_builder_config()
            config.max_workspace_size = 1 << 30
            # config.set_flag(trt.BuilderFlag.FP16)
            serialized_engine = builder.build_serialized_network(
                network, config=config)
            engine = runtime.deserialize_cuda_engine(serialized_engine)
            print("build engine from {}  sucessfully".format(onnx_file_path))
            if self.catched_engine:
                with open(cached_engine_path, "wb") as f:
                    print(f"cached engine to: {cached_engine_path}")

        return engine

class OnnxModel:
    def __init__(self, model_name="detector_corrector", model_dir="."):
        if not model_name.endswith(".onnx"):
            model_name = model_name + ".onnx"
        model_path = os.path.join(model_dir, model_name)
        print(f"onnx model path is {model_path}")
        self.ort_session = self.load_model(model_path)
    def load_model(self, model_path):
        providers = ['CUDAExecutionProvider']  # onnxruntime
        # sess_options = onnxruntime.SessionOptions()
        # sess_options.intra_op_num_threads = 10
        # sess_options.inter_op_num_threads = 10
        print(f"onnxruntime get device {onnxruntime.get_device()} available providers {onnxruntime.get_available_providers()}")
        ort_session = onnxruntime.InferenceSession(
            model_path, providers=providers)
        print(f"onnxruntime session providers {ort_session.get_providers()}")

        return ort_session

    def predict(self, inputs):
        ort_outs = self.ort_session.run(None, inputs)
        return ort_outs

trt_model = TrtModel()
onnx_model = OnnxModel()

trt_output = trt_model.predict(preprocess_data(is_trt=True))

onnx_output = onnx_model.predict(preprocess_data(is_trt=False))

trt_detector_logits, trt_corrector_logits = trt_output
onnx_detector_logits, onnx_corrector_logits = onnx_output

import numpy as np
assert np.allclose(trt_detector_logits, onnx_detector_logits)
assert np.allclose(trt_corrector_logits, onnx_corrector_logits)

if you run this code, np.allclose result always False


Sorry for the delayed response. We could reproduce the issue.
Please allow us some time to work on this.

Thank you.

Ok,I am still waiting.