The CPU is running normally, but the GPU running results are inconsistent

[FFchopon]

Describe the issue

Loading and optimizing the model using CUDA may result in inconsistent outputs after optimization. In contrast, performing the optimization on the CPU produces consistent results.

• Actual Behavior:

AssertionError: 
Not equal to tolerance rtol=0.001, atol=0.001

Mismatched elements: 2 / 138 (1.45%)
Max absolute difference: 32
Max relative difference: inf
 x: array([[32, 14,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
         0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
         0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0],...
 y: array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0],...

• Expected Behavior:
  The optimized model should produce identical results for all outputs when compared to the original model, within the specified tolerance.

To reproduce

1. Download the model

2. Run the following script:

import onnx
import onnxruntime as ort
from onnxruntime.transformers import optimizer
import numpy as np

model_path = "9256.onnx"
optimized_model_path = f"./opt.onnx"
input_data = {
    "v8_0": np.array([[[[0.5576], [0.4236]]]], dtype=np.float16),
    "v7_0": np.array([[[[0.1953]]], [[[0.94]]], [[[0.807]]]], dtype=np.float16),
}

sess_options = ort.SessionOptions()
sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
original_session = ort.InferenceSession(model_path, sess_options, providers=["CUDAExecutionProvider"])
original_output_names = [output.name for output in original_session.get_outputs()]
original_result = original_session.run(original_output_names, input_data)
original_result2 = original_session.run(original_output_names, input_data)
for r1, r2 in zip(original_result, original_result2):
    np.testing.assert_allclose(r1, r2, rtol=1e-3, atol=1e-3)

optimized_model = optimizer.optimize_model(model_path, opt_level=99)
optimized_model.save_model_to_file(optimized_model_path)
optimized_session = ort.InferenceSession(optimized_model_path, providers=["CUDAExecutionProvider"])
optimized_output_names = [output.name for output in optimized_session.get_outputs()]
optimized_result = optimized_session.run(optimized_output_names, input_data)
for r1, r2 in zip(original_result, optimized_result):
    np.testing.assert_allclose(r1, r2, atol=1e-3, rtol=1e-3)

notice:

1. providers=["CUDAExecutionProvider"] -> inconsistent outputs
2. providers=["CPUExecutionProvider"] -> run well

Urgency

No response

Platform

Linux

OS Version

Ubuntu 20.04

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

5c1b7cc

ONNX Runtime API

Python

Architecture

X64

Execution Provider

CUDA

Execution Provider Library Version

No response