-
-
Couldn't load subscription status.
- Fork 10.9k
[torch.compile] Fix tests for torch==2.9 inductor partition #26116
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[torch.compile] Fix tests for torch==2.9 inductor partition #26116
Conversation
|
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Code Review
This pull request updates the project to use PyTorch 2.9, which involves updating numerous dependency files and CI configurations. The changes are generally well-aligned with this goal, including necessary workarounds for compatibility. My main feedback is regarding a temporary comment left in a test script, which should be cleaned up before merging to maintain code clarity.
| echo 'import os; os.system("touch /tmp/changed.file")' >> vllm/__init__.py | ||
|
|
||
| VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL=1 VLLM_USE_PRECOMPILED=1 pip3 install -vvv -e . | ||
| # TESTING, TO BE REMOVED |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The comment # TESTING, TO BE REMOVED suggests this is a temporary change that should not be merged. If the following --extra-index-url is required for the CI to pass with PyTorch 2.9, please remove this comment to avoid confusion and prevent this from being accidentally reverted in the future. If this is indeed a temporary change for debugging, it should be reverted before merging.
66ab1b0 to
3285244
Compare
1c0a8d9 to
8171537
Compare
|
Documentation preview: https://vllm--26116.org.readthedocs.build/en/26116/ |
|
This pull request has merge conflicts that must be resolved before it can be |
…ests that would require it to be False by default Signed-off-by: ProExpertProg <[email protected]>
Signed-off-by: Luka Govedič <[email protected]>
Signed-off-by: Luka Govedič <[email protected]>
Signed-off-by: Luka Govedič <[email protected]>
Signed-off-by: ProExpertProg <[email protected]>
Signed-off-by: ProExpertProg <[email protected]>
Signed-off-by: ProExpertProg <[email protected]>
Signed-off-by: ProExpertProg <[email protected]>
Signed-off-by: ProExpertProg <[email protected]>
daf918e to
7499476
Compare
…ject#26116) Signed-off-by: ProExpertProg <[email protected]> Signed-off-by: Luka Govedič <[email protected]> Signed-off-by: Jonah Bernard <[email protected]>
…ject#26116) Signed-off-by: ProExpertProg <[email protected]> Signed-off-by: Luka Govedič <[email protected]> Signed-off-by: bbartels <[email protected]>
…ject#26116) Signed-off-by: ProExpertProg <[email protected]> Signed-off-by: Luka Govedič <[email protected]>
…ject#26116) Signed-off-by: ProExpertProg <[email protected]> Signed-off-by: Luka Govedič <[email protected]>
…ject#26116) Signed-off-by: ProExpertProg <[email protected]> Signed-off-by: Luka Govedič <[email protected]> Signed-off-by: xuebwang-amd <[email protected]>
…ject#26116) Signed-off-by: ProExpertProg <[email protected]> Signed-off-by: Luka Govedič <[email protected]> Signed-off-by: xuebwang-amd <[email protected]>
…ject#26116) Signed-off-by: ProExpertProg <[email protected]> Signed-off-by: Luka Govedič <[email protected]> Signed-off-by: 0xrushi <[email protected]>
…ject#26116) Signed-off-by: ProExpertProg <[email protected]> Signed-off-by: Luka Govedič <[email protected]> Signed-off-by: 0xrushi <[email protected]>
Purpose
Fix compilation tests for Inductor graph partitioning. Also remove a noisy warning and remove
cudagraph_unsafetags from attention ops to allow proper behavior with empty splitting ops.To work this still requires #26735 and a spawn workaround. But for current main this should not change any behavior.
Remaining issues:
2.9 issues:
Test Plan
Tested with torch 2.9 in all of CI with #26738. Tested all tests locally.
Test Result
test_multiple_graphs.pyfailstest_toy_llama.pydisables caching because otherwise AotAutograd cache hits from non-partitioned to partitioned graph.