[torch.compile] A simple solution to recursively compile loaded model: using phi3-small as an example
It's often hard to apply torch.compile directly to a model because many limitations. However, there are only a few key nn.Modules we really need to compile to get max speed. This PR presents a short solution to solve the problem. Phi3-small got 15% latency improvement with this.
This PR contains 3 major functions:
- register_module_to_compile: it registers which torch.nn.Module to compile when scanning child modules from top to bottom (e.g., from Phi3 to MLP, layernorm).
- compile_child_modules: it is called after model is loaded to visit child nn.Module's from top (e.g., GPT2) to bottom (e.g., MLP). If a nn.Module is compiled, its child nn.Module's won't be compiled again.
- unregister_module_to_compile: it undoes register_module_to_compile.
- To enable this feature set COMPILE_CHILD_NN_MODULES=1.
👋 Hi! Thank you for contributing to the vLLM project.
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 do one of these:
- Add
readylabel to the PR - Enable auto-merge.
🚀
@youkaichao, did you get a chance to review? @comaniac, it'd nice if you can help reviewing. At least, let me know if I should continue writing test. This solution is extremely simple and effective.
Thank you.
This pull request has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this pull request should remain open. Thank you!
This pull request has merge conflicts that must be resolved before it can be merged. Please rebase the PR, @wschin.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork
This pull request has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this pull request should remain open. Thank you!
This pull request has been automatically closed due to inactivity. Please feel free to reopen if you intend to continue working on it. Thank you!