Apply FSDP2 to offline training

[j1young]

Motivation

apply FSDP2 to the offline training pipeline.

On H100x4 GPUs, the training throughput shows an improvement of ~1.4× compared to the previous baseline.

The following command was used for testing

torchrun \
    --standalone \
    --nproc_per_node 4 \
    ./scripts/train_eagle3_offline.py \
    --target-model-path lmsys/gpt-oss-120b-bf16 \
    --draft-model-config ./configs/gpt-oss-120B-eagle3.json \
    --train-data-path ${TRAIN_DATA_PATH} \
    --train-hidden-states-path ${TRAIN_HS_PATH} \
    --output-dir ${OUTPUT_DIR} \
    --draft-global-batch-size 16 \
    --draft-micro-batch-size 1 \
    --num-epochs 10 \
    --learning-rate 1e-4 \
    --max-length 2048 \
    --chat-template gpt-oss

Modifications

Related Issues

Accuracy Test

Benchmark & Profiling

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://sgl-fru7574.slack.com/archives/C09784E3EN6 to discuss your PR.

[gemini-code-assist]

Summary of Changes

Hello @j1young, 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 significantly upgrades the offline training infrastructure by integrating PyTorch's FSDP2. The core changes involve migrating to the latest FSDP API for improved distributed training efficiency and carefully adapting the state saving mechanisms to properly handle distributed tensors, ensuring robust checkpointing and model persistence.

Highlights

• Performance Improvement: This pull request applies FSDP2 to the offline training pipeline, resulting in an approximate 1.4x improvement in training throughput on H100x4 GPUs compared to the previous baseline.
• FSDP2 Integration: The FSDP implementation has been updated from the older FullyShardedDataParallel API to the newer fully_shard function, utilizing MixedPrecisionPolicy for mixed-precision training.
• Distributed Device Mesh: A new get_dp_device_mesh function has been introduced and integrated to manage and retrieve the distributed data parallel device mesh, which is crucial for the new FSDP2 API.
• State Saving Enhancements: The logic for saving model and optimizer states has been refined to correctly handle DTensor objects, ensuring they are converted to full tensors before saving to prevent data loss or corruption in distributed environments.

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 migrates the offline training pipeline to use FSDP2, which results in a significant performance improvement. The changes correctly adapt the code to the new fully_shard API, including updates to model wrapping, checkpointing logic for sharded model and optimizer states, and necessary additions to distributed utilities. My review focuses on improving the maintainability of the new checkpointing logic. I've suggested refactoring a couple of areas to use more concise, Pythonic constructs like dictionary comprehensions, which will make the code easier to read and maintain.