[TPU] Avoid Triton Import

[robertgshaw2-redhat]

SUMMARY:

• delay triton import until needed
• breaks tpu backend

[github-actions]

👋 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 fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

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 ready label to the PR or enable auto-merge.

🚀

[houseroad]

Looks good.

Btw, shall we consider to have a common interface of layers, and implementations for different backend are put in the different files.

[robertgshaw2-redhat]

Looks good.

Btw, shall we consider to have a common interface of layers, and implementations for different backend are put in the different files.

yes, I can catch you up @houseroad on what our plan is offline in a separate thread. We have been planning a refactor led by Lucas but its just been hard to prioritize it with all the models coming out. If your team has capacity to help with it, we can collaborate

[mgoin]

Makes sense, unfortunately

[yaochengji]

LGTM, thanks!

[wangxiyuan]

+1, break other backend as well.

[MengqingCao]

Thanks for this fix! BTW, I'm doing smilar things in pr #15099.
Do you think adding a defensive check before importing trion, and adding dummy decorators for triton is a good way to avoid triton breaks like this? Please let me know your suggestions, thx!

[robertgshaw2-redhat]

Thanks for this fix! BTW, I'm doing smilar things in pr #15099. Do you think adding a defensive check before importing trion, and adding dummy decorators for triton is a good way to avoid triton breaks like this? Please let me know your suggestions, thx!

This would be a much better solution. Our code is riddled with delayed imports for this reason.

[robertgshaw2-redhat]

Thanks for this fix! BTW, I'm doing smilar things in pr #15099. Do you think adding a defensive check before importing Triton, and adding dummy decorators for triton is a good way to avoid triton breaks like this? Please let me know your suggestions, thx!

Thanks @MengqingCao - will take a look

[yuanfz98]

Hello @robertgshaw2-redhat ,
While you moved imports into function process_weights_after_loading, the UnquantizedFusedMoEMethod will still be used in FusedMoe

vllm/vllm/model_executor/layers/fused_moe/layer.py

Line 497 in ec7da6f

if quant_config is None:

and Non-CUDA backends will run into triton import.

[MengqingCao]

Hello @robertgshaw2-redhat , While you moved imports into function process_weights_after_loading, the UnquantizedFusedMoEMethod will still be used in FusedMoe

vllm/vllm/model_executor/layers/fused_moe/layer.py

Line 497 in ec7da6f

if quant_config is None:

and Non-CUDA backends will run into triton import.

Could you try with #15099, which is a pr addressing the triton import issue on NON-CUDA devices. Feel free to review this pr or feed back your result, thanks!