[RL] fix update weight for FusedMoE with EP #8676
Conversation
There was a problem hiding this comment.
Summary of Changes
Hello @zhuzilin, 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 introduces a critical memory optimization for the FusedMoE layer, particularly relevant when using Expert Parallelism (EP). My changes ensure that GPU memory is not prematurely allocated during model initialization, which is essential for efficient resource management in environments like colocated Reinforcement Learning.
Highlights
- Memory Optimization: I've modified the initialization of
self.expert_map_cputo explicitly create it on the CPU, preventing it from defaulting to a GPU device during model setup. - Lazy GPU Tensor Creation: I've moved the creation and transfer of
self.expert_map_gpufrom the__init__method to theforwardmethod. This ensures the GPU tensor is only allocated when it's actually required for computation, significantly reducing the initial GPU memory footprint. - Support for Colocated RL: These changes specifically address memory management requirements in scenarios like colocated Reinforcement Learning (RL), where it's crucial to release GPU occupation and avoid creating GPU tensors that cannot be loaded via a weight loader.
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 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 or fill out our survey 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 addresses a critical memory management issue for FusedMoE with Expert Parallelism. The changes are logical, well-targeted, and align perfectly with the stated goal of avoiding premature GPU memory allocation. One minor suggestion is provided to refactor the lazy initialization logic for improved clarity.
| topk_output = topk_output._replace( | ||
| topk_ids=self.expert_map_gpu[topk_output.topk_ids] | ||
| ) | ||
| if not self.enable_flashinfer_cutlass_moe: |
There was a problem hiding this comment.
This line does not follow the original logic. How about this?
| if not self.enable_flashinfer_cutlass_moe: | |
| if self.moe_ep_size > 1 and not self.enable_flashinfer_cutlass_moe: |
Motivation
During colocated RL, we will use
/release_memory_occupationto release all GPU allocations within model init, and we should not create GPU tensors that cannot be loaded through weight_loader.This PR moves the creation of
self.expert_map_gputoforwardand make sureself.expert_map_cpuis created on CPU (which was by default created on GPU).Thank you for your time on reviewing this PR :)
Modifications
Accuracy Test
Benchmark & Profiling
Checklist