[DP/EP] PPLX<>Triton Debug #20957
[DP/EP] PPLX<>Triton Debug #20957
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 🚀 |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @robertgshaw2-redhat, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
This pull request focuses on improving the debugging and performance optimization capabilities for parallel execution, particularly for Mixture of Experts (MoE) models leveraging PPLX kernels within vLLM. It introduces new hardware-specific MoE configurations, adds profiling to key kernel operations, and provides utility scripts for easier testing.
Highlights
- PPLX Kernel Profiling: I've added detailed timing logs for the
dispatchandcombineoperations within the PPLX (Perplexity) kernels used in fused Mixture of Experts (MoE) layers. This will help in debugging and performance analysis of these critical parallel operations. - New MoE Configurations: I've introduced a set of new JSON configuration files for fused MoE layers, specifically tailored for NVIDIA B200 and H100 GPUs. These configurations define optimal block sizes, warps, and stages for various expert and neuron counts, supporting fp8_w8a8 data types.
- Engine Debugging Visibility: I've enhanced the logging in the vLLM engine's core loop to explicitly indicate when a dummy batch is being executed. This provides better visibility into the engine's internal state during parallel processing.
- Smoke Test Recipes: I've added a new
Justfilecontaining convenient recipes for launching vLLM in data parallel/expert parallel and tensor parallel modes, as well as for runninglm_evalandbenchmark_serving.py. This streamlines common testing and benchmarking workflows.
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This pull request adds a Justfile for smoke testing and several JSON configuration files for MoE kernels on different hardware. It also includes some debug logging for performance analysis in pplx_prepare_finalize.py and core.py.
My review focuses on ensuring the code is maintainable and portable. I've pointed out a hardcoded path in the Justfile and suggested removing or adjusting the debug logging before merging into a main branch. The changes are otherwise consistent with the debugging purpose of this PR.
| @@ -0,0 +1,36 @@ | |||
| # set this on your machine | |||
| vllm-directory := "/home/rshaw/vllm/" | |||
There was a problem hiding this comment.
The vllm-directory is hardcoded to a user-specific absolute path. This makes the script not portable and will fail for other developers. It's better to use a relative path to the project root.
vllm-directory := "../.."
| start = time.perf_counter() | ||
| self.a2a.combine(out_tokens=output, | ||
| indices=topk_ids.view(dtype=torch.uint32), | ||
| weights=topk_weights, | ||
| expert_y=fused_expert_output, | ||
| bound_m=bound_m) | ||
| end = time.perf_counter() | ||
| logger.info("combine took %.3f ms", (end - start) * 1000) |
There was a problem hiding this comment.
Similar to the dispatch call, this performance logging for combine seems to be for debugging. It should be removed or changed to DEBUG level before merging to a main branch to avoid log spam and potential performance impact.
self.a2a.combine(out_tokens=output,
indices=topk_ids.view(dtype=torch.uint32),
weights=topk_weights,
expert_y=fused_expert_output,
bound_m=bound_m)
vllm/v1/engine/core.py
Outdated
| @@ -946,6 +946,7 @@ def run_busy_loop(self): | |||
|
|
|||
| # We are in a running state and so must execute a dummy pass | |||
| # if the model didn't execute any ready requests. | |||
| logger.info("Executing dummy batch for wave %d.", self.current_wave) | |||
There was a problem hiding this comment.
This log message is useful for debugging, but using logger.info can make the logs noisy during normal operation. It would be better to use logger.debug for this kind of diagnostic information.
| logger.info("Executing dummy batch for wave %d.", self.current_wave) | |
| logger.debug("Executing dummy batch for wave %d.", self.current_wave) |
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.Purpose
Test Plan
Test Result
(Optional) Documentation Update