Bug fix for trainer_state saving

[chenyushuo]

Description

As the title says. Fix for #399 .

Checklist

Please check the following items before code is ready to be reviewed.

• Code has passed all tests
• Docstrings have been added/updated in Google Style
• Documentation has been updated
• Code is ready for review

[gemini-code-assist]

Summary of Changes

Hello @chenyushuo, 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 resolves a bug related to the incomplete saving of the trainer's state, specifically addressing how the sampling strategy's progress is persisted across training sessions. The changes introduce a more robust state management system for sample strategies, ensuring that training can be accurately resumed from a saved checkpoint. This involves adding dedicated state serialization methods to sampling strategies and updating the trainer's save/load logic to utilize these new capabilities, alongside improved test coverage.

Highlights

• Refactored Trainer State Saving: The mechanism for saving and loading the trainer's state has been updated to explicitly handle the state of the sampling strategy, ensuring proper resumption of training.
• Introduced Sample Strategy State Management: Abstract state_dict and load_state_dict methods were added to the base SampleStrategy class, with concrete implementations for DefaultSampleStrategy and MixSampleStrategy to manage their internal buffer states.
• Enhanced Trainer Test Coverage: New test cases were added to trainer_test.py to verify the correct persistence and restoration of the trainer's metadata, including the sample strategy's state, after resuming operations.
• Configuration Flexibility: Cluster node and GPU settings in config.yaml can now be dynamically overridden via environment variables, providing more flexible deployment options.

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 by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

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 pull request 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

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

1. 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. ↩

[gemini-code-assist]

Code Review

This pull request introduces a solid fix for saving trainer state by delegating state management to the SampleStrategy. This is a more robust approach than calculating an index. The implementation correctly adds state_dict and load_state_dict methods to the sample strategy hierarchy and updates the trainer to use them, including a backward-compatibility path for older checkpoints. The accompanying test changes verify the new save and resume functionality. I have a couple of minor suggestions: one to correct a typo and another to ensure test artifacts are cleaned up properly.

[chenyushuo]

/unittest-module-trainer

[github-actions]

Summary

Tests 📝	Passed ✅	Failed ❌	Skipped ⏭️	Other ❓	Flaky 🍂	Duration ⏱️
21	19	0	2	0	0	39m 41s

Skipped

Tests	Status
tests/trainer/trainer_test.py::TestMultiModalGRPO::test_trainer	skipped ⏭️
tests/trainer/trainer_test.py::TestMultiModalSFT::test_trainer	skipped ⏭️

Tests

Test Name	Status	Flaky	Duration
tests/trainer/trainer_test.py::TestTrainerCountdown_0_fsdp::test_trainer	✅		3m 9s
tests/trainer/trainer_test.py::TestTrainerCountdown_1_megatron::test_trainer	✅		3m 49s
tests/trainer/trainer_test.py::TestStepAheadAsyncRL::test_trainer	✅		1m 21s
tests/trainer/trainer_test.py::TestTrainerGSM8K_0_fsdp::test_trainer	✅		1m 23s
tests/trainer/trainer_test.py::TestTrainerGSM8K_1_fsdp2::test_trainer	✅		1m 25s
tests/trainer/trainer_test.py::TestTrainerGSM8K_2_fsdp::test_trainer	✅		1m 25s
tests/trainer/trainer_test.py::TestTrainerGSM8K_3_fsdp2::test_trainer	✅		1m 34s
tests/trainer/trainer_test.py::TestTrainerSFTWarmupGSM8K::test_trainer	✅		2m 31s
tests/trainer/trainer_test.py::TestTrainerDPO::test_trainer	✅		1m 5s
tests/trainer/trainer_test.py::TestTrainerSFT::test_trainer	✅		1m 1s
tests/trainer/trainer_test.py::TestTrainerToolsSFT::test_trainer_tools	✅		1m 1s
tests/trainer/trainer_test.py::TestFullyAsyncMode_0_fsdp::test_fully_async_mode	✅		1m 58s
tests/trainer/trainer_test.py::TestFullyAsyncMode_1_fsdp::test_fully_async_mode	✅		1m 58s
tests/trainer/trainer_test.py::TestFullyAsyncMode_2_megatron::test_fully_async_mode	✅		2m 28s
tests/trainer/trainer_test.py::TestTrainerCheckpointSave_0_fsdp::test_trainer	✅		2m 17s
tests/trainer/trainer_test.py::TestTrainerCheckpointSave_1_megatron::test_trainer	✅		4m 7s
tests/trainer/trainer_test.py::TestTrainerMIX::test_trainer	✅		2m 31s
tests/trainer/trainer_test.py::TestMultiModalGRPO::test_trainer	⏭️		610ms
tests/trainer/trainer_test.py::TestMultiModalSFT::test_trainer	⏭️		807ms
tests/trainer/trainer_test.py::TestTrainerLoRA::test_trainer	✅		2m 53s
tests/trainer/trainer_test.py::TestOverRollout::test_trainer	✅		1m 17s

Github Test Reporter by CTRF 💚