When p=1, GlobalLpPool produces difference results for TensorRT and ONNXRuntime

Description

For the following simple onnx model,

the results 'output' are different when running this model using tensorrt and onnxruntime. The results of onnxruntime are as follows:

ONNXRuntime:
 [array([[[[1.]],

        [[1.]],

        [[1.]]]], dtype=float32), 
array([[[[-1.]],

        [[-1.]],

        [[-1.]]]], dtype=float32)]

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

TensorRT: 
 [array([[[[-1.]],

        [[-1.]],

        [[-1.]]]], dtype=float32), 
array([[[[-1.]],

        [[-1.]],

        [[-1.]]]], dtype=float32)]

In the above results, the first array is the values of 'output' and the second array is the values of ‘min_output'. From the above results, we can see that the results ‘min_output' are identical for both tensorrt and onnxruntime, while the results 'output' are different.

The reason for different 'output' may be the different calculation method in tensorrt and onnxruntime.
According to the formula of GlobalLpPool:

when p=1, the formula is as follows:

I think that onnxruntime utilizes the above formula, while tensorrt removes the Abs operation, which causes this issue.

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

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('44.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)
    
    #--------------------------------------------------------
        
    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('44.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)
    
    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) # 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.

Start by running the supplied Python reproducer with 44.onnx and inputs.pkl from testcase.zip, then compare the GlobalLpPool p=1 outputs from TensorRT and ONNXRuntime. Trace the TensorRT handling of this operator and use the script's assertion to verify completion; done means the supplied case produces matching output values.