[Nightly] Enhance XPU test workflows #1723
Conversation
| @@ -85,6 +85,17 @@ rm -rf ./tmp | |||
| bash third_party/torch-xpu-ops/.github/scripts/rpath.sh ${WORKSPACE}/pytorch/dist/torch*.whl | |||
| python -m pip install --force-reinstall tmp/torch*.whl | |||
|
|
|||
| # Build torchvision torchaudio | |||
There was a problem hiding this comment.
Please use a env var or parameter control this behavior, by default off unless the var/parameter has been set to on.
Including triton build also. For pinned triton commit, we can use make triton directly under pytorch root dir. For customized triton commit, we can build by ourselves or leverage the scripts directly https://github.com/pytorch/pytorch/blob/main/.github/scripts/build_triton_wheel.py and refer usage https://github.com/chuanqi129/pytorch/blob/fix_triton_version_split/.github/workflows/build-triton-wheel.yml#L158-L160. Before this step, we need replace the pined triton xpu commit file content to customized one.
| @@ -60,6 +60,9 @@ jobs: | |||
| HF_HUB_ETAG_TIMEOUT: 120 | |||
| HF_HUB_DOWNLOAD_TIMEOUT: 120 | |||
| steps: | |||
| - name: Cleanup Workspace | |||
There was a problem hiding this comment.
Why we need this change? If this test failed, it should caused by other jobs don't clean workspace at the end of the workflow
| - name: Checkout torch-xpu-ops | ||
| uses: actions/checkout@v4 | ||
| - name: Prepare Stock Pytorch | ||
| - name: Cleanup Workspace |
There was a problem hiding this comment.
Same here. And suggest to use a github action or define a github action directly for the workspace cleanup, such as https://github.com/pytorch/pytorch/blob/main/.github/workflows/_xpu-test.yml#L338 or https://github.com/pytorch/pytorch/blob/main/.github/workflows/_xpu-test.yml#L83
| @@ -0,0 +1,87 @@ | |||
| #!/bin/bash | |||
|
|
|||
| set -xe -o pipefail | |||
There was a problem hiding this comment.
I feel like this script is over design somehow. Add some suggestions
- Try to define new github actions in stead of bash script directly, they both can reuse in different workflows.
- Split this script as multiple parts. For example, source code preparation and requirements setup, env information collection and display etc. There is a good example https://github.com/intel/torch-xpu-ops/tree/main/.github/actions/print-environment
- Can we remove conda dependency?
| else | ||
| pip install torch torchvision torchaudio --pre --index-url https://download.pytorch.org/whl/nightly/xpu | ||
| TORCH_COMMIT_ID=$(python -c 'import torch; print(torch.version.git_version)') | ||
| PYTORCH_VERSION="main" |
There was a problem hiding this comment.
Seem the main has ambiguity here, if we don't test main branch from source code build
| python pytorch/torch/utils/collect_env.py | ||
| rm -rf /tmp/torchinductor_* | ||
| rm -rf ~/.triton/cache | ||
| ./.github/scripts/setup_test_env.sh \ |
There was a problem hiding this comment.
Commented in the scripts file directly
| elif [ "${PYTORCH_VERSION}" == "nightly" ];then | ||
| python -m pip install torch torchvision torchaudio --pre --index-url https://download.pytorch.org/whl/nightly/xpu | ||
| else | ||
| python -m pip install ${WORKSPACE}/torch*.whl |
There was a problem hiding this comment.
Will this step include torchvision and torchaudio installation?
Uh oh!
There was an error while loading. Please reload this page.