[doc] feat: add a doc for vllm+megatron training #3974
Conversation
techkang
requested review from
eric-haibin-lin and
zhaochenyang20
as code owners
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request adds a new documentation page for best practices on training large models with Verl, using DAPO on Qwen3-235B as an example. The documentation is comprehensive and provides valuable, detailed guidance on configuring various parameters. My review focuses on ensuring clarity and minimizing potential risks for users. I've identified a couple of points that could be improved: the model name used in the example might be confusing as it doesn't seem to be public, and a dependency link points to a personal repository, which could be a supply chain risk. Addressing these would enhance the quality and trustworthiness of the documentation.
docs/perf/best_practices.rst
Outdated
| * - | ||
| - ``actor_rollout_ref.actor.megatron.use_mbridge`` | ||
| - 是否启用 mbridge 格式转换 | ||
| - 使用 Megatron 训练时建议开启；需配合最新版 mbridge（详见 https://github.com/ISEEKYAN/mbridge） |
There was a problem hiding this comment.
The documentation for the mbridge dependency points to a personal GitHub repository (https://github.com/ISEEKYAN/mbridge). For a production-oriented framework like verl, relying on personal repositories for dependencies can be a supply chain security and maintenance risk. It is recommended to link to an official, organization-owned repository or the project's page on PyPI to ensure stability and trust.
### What does this PR do? As described in title, add a doc for vllm+megatron training.
### What does this PR do? As described in title, add a doc for vllm+megatron training.
2 participants
What does this PR do?
As described in title, add a doc for vllm+megatron training.