-
-
Notifications
You must be signed in to change notification settings - Fork 5.7k
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
[Feature] Pluggable platform-specific scheduler #13161
base: main
Are you sure you want to change the base?
Conversation
Signed-off-by: Yannick Schnider <[email protected]>
👋 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 🚀 |
While we need more discussions about this feature in v1, I think it's ok for v0 to have it. Could you add a unit test with a dummy scheduler to 1) test the functionality, and 2) be an example? |
Signed-off-by: Yannick Schnider <[email protected]>
e0beefe
to
835e97b
Compare
Thanks for your feedback @comaniac ! I have added a test which validates the functionality. |
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.
It would be good to avoid a duplicated dummy platform in the plugin tests. Can we do either of the following:
- Do not use a dummy platform in the dummy scheduler plugin. This also demonstrates that the pluggable scheduler is compatible with existing platforms.
- If (1) is not feasible, we could consider extending dummy platform plugin to also have a dummy scheduler, so that it becomes
vllm_add_dummy_platform_and_scheduler
.
This PR enables different platforms to plugin their (hardware) specific scheduler class in vLLM version V0 and therefore addresses the scheduler part of this RFC #11162. A pluggable scheduler is needed to add support for the IBM Spyre accelerator #9652.
Note that this feature is under development for V1 but missing in the current V0.
Changes
scheduler_cls
to classParallelConfig
invllm/config.py
scheduler_cls
invllm/engine/llm_engine.py
Remarks
scheduler_cls
is initialized with "vllm.core.scheduler.Scheduler", the default behaviour does not change. However, the change allows to point to a (hardware) specific scheduler class in platforms.worker_cls
to have minimal changes (Otherwise I would have to overwrite "auto" for all platforms using the default scheduler).