[WIP][EPLB] Enable Llama4 EPLB #20792
Conversation
Signed-off-by: Brayden Zhong <b8zhong@uwaterloo.ca>
Signed-off-by: Brayden Zhong <b8zhong@uwaterloo.ca>
|
👋 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 @b8zhong, 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 is a work-in-progress effort to integrate Expert Parallel Load Balancing (EPLB) capabilities into the Llama4 model within vLLM. My changes involve extending the model's architecture to support EPLB configuration, implementing a standardized interface for MoE management, and ensuring proper propagation of EPLB settings throughout the model's layers.
Highlights
- Enable Expert Parallel Load Balancing (EPLB) for Llama4: I've introduced an
enable_eplbboolean parameter across various Llama4 model components (Llama4MoE,Llama4DecoderLayer,Llama4Model) to control the activation of Expert Parallel Load Balancing. This flag is propagated from the top-level model configuration down to individual MoE layers. - MixtureOfExperts Interface Implementation: The
Llama4ForCausalLMclass now inherits from a newMixtureOfExpertsinterface. This involves implementing several properties (expert_weights,num_moe_layers,num_logical_experts, etc.) and methods (set_eplb_state,get_expert_weights) to provide a standardized way to query and manage the state of Mixture-of-Experts layers for EPLB. - MoE Layer Tracking and Management: I've added logic within
Llama4ModelandLlama4ForCausalLMto identify and track all Mixture-of-Experts (MoE) layers. This tracking is crucial for theMixtureOfExpertsinterface to correctly manage and expose information about the model's expert configuration and state for load balancing. - Dependency Updates and Model Layer Creation Override: I've added new imports for the
MixtureOfExpertsinterface andPPMissingLayerfrom internal vLLM modules, indicating integration with existing MoE and pipeline parallelism infrastructure. Additionally, themake_layersmethod inLlama4Modelhas been overridden to ensure theenable_eplbparameter is correctly passed to the decoder layers during their creation.
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 introduces changes to enable EPLB for the Llama4 model. The changes involve propagating an enable_eplb flag through the model's layers and implementing the MixtureOfExperts interface. There are opportunities for code improvement to enhance maintainability and correctness.
| self.enable_eplb = vllm_config.parallel_config.enable_eplb if hasattr( | ||
| vllm_config.parallel_config, 'enable_eplb') else False |
There was a problem hiding this comment.
Consider using getattr with a default value directly in the assignment to simplify the code and avoid the hasattr check.
This makes the code more concise and readable.
self.enable_eplb = vllm_config.parallel_config.enable_eplb| for i, layer in enumerate(self.layers): | ||
| if isinstance(layer, layer_type): | ||
| is_moe_layer = (config.interleave_moe_layer_step > 0 | ||
| and (i + 1) % config.interleave_moe_layer_step | ||
| == 0) |
There was a problem hiding this comment.
The logic to determine if a layer is a MoE layer can be simplified by extracting it into a separate function for better readability and maintainability. This also avoids duplicating the logic.
def is_moe_layer(config, i):
return config.interleave_moe_layer_step > 0 and (i + 1) % config.interleave_moe_layer_step == 0
for i, layer in enumerate(self.layers):
if isinstance(layer, layer_type) and is_moe_layer(config, i):| self.moe_layers: list[FusedMoE] = [] | ||
| config = vllm_config.model_config.hf_config | ||
| for i, layer in enumerate(self.model.layers): | ||
| if isinstance(layer, Llama4DecoderLayer): | ||
| is_moe_layer = (config.interleave_moe_layer_step > 0 | ||
| and (i + 1) % config.interleave_moe_layer_step | ||
| == 0) | ||
| if is_moe_layer and hasattr(layer, 'feed_forward') and hasattr( | ||
| layer.feed_forward, 'experts'): | ||
| self.moe_layers.append(layer.feed_forward.experts) |
There was a problem hiding this comment.
The logic to identify MoE layers is duplicated from Llama4Model.__init__. This could be refactored into a helper function to improve maintainability and reduce code duplication.
For example, you could define a helper function at the module level:
def _is_moe_layer(config, layer_idx: int) -> bool:
return (config.interleave_moe_layer_step > 0 and
(layer_idx + 1) % config.interleave_moe_layer_step == 0)Then, you can use this helper in both places.
|
This pull request has merge conflicts that must be resolved before it can be |
Purpose
As a part of #20468, support EPLB for Llama4
Test Plan
WIP,
Test Result
WIP
(Optional) Documentation Update
WIP