Inconsistent Results for Output v1_0 After ONNX Runtime Optimization (Flaky Test)

[Thrsu]

Describe the issue

I encountered an issue where the results from the optimized ONNX model are inconsistent with the original unoptimized model. Specifically, the output v1_0 is inconsistent, and the discrepancy occurs intermittently (flaky test). This issue arises after applying different optimization levels (opt_level=0, 1, 2, 99), though it does not occur consistently in every test.

The following error is reported when comparing the results from the optimized and original models:

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

Mismatched elements: 5 / 5 (100%)
Max absolute difference: 0.79532735
Max relative difference: 1.
 x: array([6.910885e-310, 6.910885e-310, 6.910885e-310, 6.910885e-310,
           6.910885e-310])
 y: array([0.795327, 0.75308 , 0.59723 , 0.711406, 0.667502])

Could anyone tell me why the v1_0 output is inconsistent after applying optimizations? Specifically, I'd like to understand the cause of this intermittent discrepancy and whether there are optimizations that could be adjusted to improve the consistency of the results.

To reproduce

1. Download the ONNX model.
2. Run the below script:

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

model_path = "inconsis3.onnx"
optimized_model_path = f"./opt.onnx"
sess_options = ort.SessionOptions()
sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
this_provider_list = ort.get_available_providers()

original_session = ort.InferenceSession(model_path, sess_options, providers=this_provider_list)
input_data = {"v0_0": np.random.rand(5).astype(np.float64)}
output_names = [output.name for output in original_session.get_outputs()]
original_result = original_session.run(output_names, input_data)

optimized_model = optimizer.optimize_model(model_path, opt_level=1, use_gpu=True)
optimized_model.save_model_to_file(optimized_model_path)
optimized_session = ort.InferenceSession(optimized_model_path, providers=this_provider_list)
optimized_model = onnx.load(optimized_model_path)
optimized_result = optimized_session.run(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)

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