TensorRT produces wrong results for onnx model with only Or and Cast on GPU 3080

[coffezhou]

Description

For the following simple onnx model,

the results Y produced by onnxruntime are as follows:

ONNXRuntime:
 [[[[ True  True  True  True]
   [ True  True  True  True]
   [ True  True  True  True]
   [ True  True  True  True]]

  [[ True  True  True  True]
   [ True  True  True  True]
   [ True  True  True  True]
   [ True  True  True  True]]

  [[ True  True  True  True]
   [ True  True  True  True]
   [ True  True  True  True]
   [ True  True  True  True]]]]

However, when I run it using tensorrt, the results Y are as follows:

TensorRT:
 [[[[ True  True  True  True]
   [False False False False]
   [False False False False]
   [False False False False]]

  [[ True  True  True  True]
   [False False False False]
   [ True False False False]
   [False False False False]]

  [[ True  True  True  True]
   [False False False False]
   [ True  True  True  True]
   [False False False False]]]]

64.6% elements are mismatched.
I have verified that the results CAST_Y for both onnxruntime and tensorrt are identical.
This issue is simliiar to 4511. Both issues only contain the Or and Cast operators.

Environment

TensorRT Version: 10.12.0.36

NVIDIA GPU: GeForce RTX 3080

NVIDIA Driver Version: 535.183.01

CUDA Version: 12.2

CUDNN Version: none

Operating System: ubuntu 20.04

Python Version (if applicable): 3.12.9

Relevant Files

Model link:

Steps To Reproduce

This issue can be reproduced by the following code with the model in the attachment.

from typing import Dict, List, Literal, Optional
import sys
import os

import numpy as np
import onnx
import onnxruntime

import tensorrt as trt
import pycuda.driver as cuda
import pycuda.autoinit

import argparse
import pickle


def test():
    onnx_model = onnx.load('333.onnx')
    
    with open("inputs.pkl", "rb") as fp:
        inputs = pickle.load(fp)

    try:
        ort_session = onnxruntime.InferenceSession(
            onnx_model.SerializeToString(), providers=["CPUExecutionProvider"]
        )
        ort_output = ort_session.run([], inputs)
    except Exception as e:
        print(e)
        print("This model cannot be executed by onnxruntime!")
        sys.exit(1)
    
    print("ONNXRuntime:\n", ort_output[1])
    
    #--------------------------------------------------------
        
    trt_logger = trt.Logger(trt.Logger.WARNING)
    trt.init_libnvinfer_plugins(trt_logger, '')
    builder = trt.Builder(trt_logger)
    network = builder.create_network(flags=1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))

    parser = trt.OnnxParser(network, trt_logger)
    with open('333.onnx', 'rb') as model_file:
        if not parser.parse(model_file.read()):
            for error in range(parser.num_errors):
                print(parser.get_error(error))
            sys.exit(1)
    
    config = builder.create_builder_config()
    serialized_engine = builder.build_serialized_network(network, config)
    
    if serialized_engine == None:
        sys.exit(1)
    
    with open("engine.trt", "wb") as f:
        f.write(serialized_engine)
        
    with open("engine.trt", "rb") as f, trt.Runtime(trt_logger) as runtime:
        engine = runtime.deserialize_cuda_engine(f.read())
    
        
    context = engine.create_execution_context()

    inputs_trt, outputs_trt, bindings = [], [], []
    stream = cuda.Stream()
    input_name = []
    output_shape_dtype = []
    #------------------------------------------------------------
    for binding in engine:
        size = trt.volume(engine.get_tensor_shape(binding))
        dtype = trt.nptype(engine.get_tensor_dtype(binding))
        host_mem = cuda.pagelocked_empty(size, dtype)
        device_mem = cuda.mem_alloc(host_mem.nbytes)
        bindings.append({'name':binding, 'address':int(device_mem)})
        
        if engine.get_tensor_mode(binding) == trt.TensorIOMode.INPUT:
            inputs_trt.append({'host': host_mem, 'device': device_mem})
            input_name.append(binding)
        else:
            outputs_trt.append({'host': host_mem, 'device': device_mem})
            output_shape = engine.get_tensor_shape(binding)
            output_shape_dtype.append({'shape':output_shape, 'dtype':dtype})

    for i, input_mem in enumerate(inputs_trt):
        inp = np.ravel(inputs[input_name[i]])
        np.copyto(input_mem['host'], inp)
        cuda.memcpy_htod_async(input_mem['device'], input_mem['host'], stream)

    for bind in bindings:
        name = bind['name']
        addr = bind['address']
        context.set_tensor_address(name, addr)
    
    context.execute_async_v3(stream_handle=stream.handle)
    
    trt_output = []
    for i, output_mem in enumerate(outputs_trt):
        cuda.memcpy_dtoh_async(output_mem['host'], output_mem['device'], stream)
        out_shape = output_shape_dtype[i]['shape']
        out = output_mem['host'].reshape(out_shape)
        trt_output.append(out)

    stream.synchronize()
    
    print("TensorRT: \n", trt_output[1])
    
    assert len(ort_output) == len(trt_output), "Unequal number of outputs"
    
    np.testing.assert_allclose(trt_output[0], ort_output[0], rtol=0.1, atol=0.1) #CAST_Y, GOOD
    np.testing.assert_allclose(trt_output[1], ort_output[1], rtol=0.1, atol=0.1) #Y, BAD
        

    
if __name__ == "__main__":
    test()
    

testcase.zip

Commands or scripts:

Have you tried the latest release?: yes

Can this model run on other frameworks? For example run ONNX model with ONNXRuntime (polygraphy run <model.onnx> --onnxrt): the mode can be executed by onnxruntime.

[kevinch-nv]

Some clarification questions:

• what type is input X?
• when running in polygraphy (i.e. polygraphy run --trt --onnxrt model.onnx) is the accuracy also different?