Enable python GPU tests#7854
Conversation
|
By default onnx runtime python binding registers all the available EPs. The list of EPs is gathered from build flags. But, can they all work together peacefully? No. DML and CUDA are not. They can't be both enabled. So, what should we do? |
|
Waiting #7987 |
|
|
||
|
|
||
| - task: PythonScript@0 | ||
| displayName: 'Regenerate cmake config with STATIC_ANALYSIS=ON ' |
There was a problem hiding this comment.
This part is temporarily removed because there is a bug in the latest VC++ release.
There was a problem hiding this comment.
nit: if we want to eventually add this back (after the bug fix is available) you can turn this step off by using condition False ... this way once the fix is available you can simply enable it again...
| //GPU failures | ||
| "^test_batchnorm_epsilon_training_mode_cuda", | ||
| "^test_batchnorm_example_training_mode_cuda", | ||
| "^test_convtranspose_autopad_same_cpu", |
There was a problem hiding this comment.
nit: move this above //gpu failures
| "^test_batchnorm_epsilon_training_mode_cuda", | ||
| "^test_batchnorm_example_training_mode_cuda", | ||
| "^test_convtranspose_autopad_same_cpu", | ||
| "^test_sub_uint8_cuda", |
There was a problem hiding this comment.
all the tests starting line 68 to 74 are not supported on both cuda and cpu so changing them to include * should work too like -> "^test_sub_uint8_*"
|
All the comments are nit... no fundamental issues hence approved the PR. |
|
Thank you for the comments. I will address them a bit later. |
3 participants
Description:
Enable python GPU tests for CUDA and ROCM EPs.
Motivation and Context
From https://github.com/onnx/onnx/blob/master/onnx/backend/test/runner/__init__.py#L230 you can see it defines two backends: "CPU" and "CUDA". But somehow our python extension names the CUDA backend as "GPU". Because "GPU" is not in ONNX test list, it doesn't get run.