Conversation
Add dumper CLI arguments (--dumper-enable, --dumper-dir, per-phase config), dumper_utils.py for SGLang/Megatron dumper integration, model.py hooks for forward-only and forward-backward phases, rollout env var plumbing, source patcher wiring in training actors, and basic e2e test.
Add _maybe_apply_dumper_overrides to disable heartbeats, force single rollout, and disable eval/save when --dumper-enable is set.
Summary of ChangesHello, 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 mechanism to dynamically adjust runtime arguments based on whether a 'dumper mode' is active. The primary goal is to streamline the configuration for dumper operations by automatically overriding certain parameters, such as disabling health checks and controlling rollout behavior, to prevent conflicts and ensure consistent execution within this specialized mode. This change is part of a larger series of enhancements related to dumper and megatron runner functionality. Highlights
Using Gemini Code AssistThe 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 by creating a comment using either
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 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here. Footnotes
|
There was a problem hiding this comment.
Code Review
This pull request introduces logic to override certain arguments when dumper mode is enabled, along with corresponding unit tests. The changes are logical and well-tested. My review identifies a minor issue with a log message that could be misleading and provides a suggestion to improve its clarity and accuracy.
| logger.info("Dumper mode: forced num_rollout=%d, disabled eval and save", args.num_rollout) | ||
| args.num_rollout = (args.start_rollout_id or 0) + 1 |
There was a problem hiding this comment.
The log message on line 1950 is misleading because it logs the value of args.num_rollout before it is updated on the next line. This can cause confusion during debugging, as the logged value will not match the value actually used. To ensure the log reflects the correct value, the assignment should happen before the logging statement.
| logger.info("Dumper mode: forced num_rollout=%d, disabled eval and save", args.num_rollout) | |
| args.num_rollout = (args.start_rollout_id or 0) + 1 | |
| args.num_rollout = (args.start_rollout_id or 0) + 1 | |
| logger.info("Dumper mode: forced num_rollout=%d, disabled eval and save", args.num_rollout) |
No description provided.