fix broken test vllm:test_kernels - test_attention_selector.py::test_flash_attn #17873
Conversation
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels. Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can either: Add 🚀 |
|
This pull request was exported from Phabricator. Differential Revision: D74440549 |
|
Could you fix the lint issue? |
|
DCO failure can be ignored, but pre-commit is something we need to fix :-) |
|
Can you merge from main to fix pre-commit? |
tracelogfb
force-pushed
the
export-D74440549
branch
from
06f8c49 to
25585aa
Compare
…flash_attn (vllm-project#17873) Summary: Pull Request resolved: vllm-project#17873 xformers calls torch.cuda.get_device_capability("cuda"). Our monkey patching version of get_device_capability didn't accept string argument. Reviewed By: houseroad Differential Revision: D74440549
|
This pull request was exported from Phabricator. Differential Revision: D74440549 |
tracelogfb
force-pushed
the
export-D74440549
branch
from
25585aa to
7d18a2f
Compare
…flash_attn (vllm-project#17873) Co-authored-by: Stephen Chen <tracelog@meta.com> Signed-off-by: Mu Huai <tianbowen.tbw@antgroup.com>
…flash_attn (vllm-project#17873) Co-authored-by: Stephen Chen <tracelog@meta.com>
…flash_attn (vllm-project#17873) Co-authored-by: Stephen Chen <tracelog@meta.com> Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com>
…y::test_flash_attn (vllm-project#17873)" This reverts commit bed409e0ce8ec3f2fec70d1cd9ffb029d80b16f4. Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com>
…flash_attn (vllm-project#17873) Co-authored-by: Stephen Chen <tracelog@meta.com> Signed-off-by: Yuqi Zhang <yuqizhang@google.com>
4 participants
Summary: xformers calls torch.cuda.get_device_capability("cuda"). Our monkey patching version of get_device_capability didn't accept string argument.
Differential Revision: D74440549