[diffusion] feat: support efficient sequence shard

[nono-Sang]

Motivation

For video models, when SGLang enables SP (Sequence Parallelism), it uses frame-level partitioning. The number of frames needs to be adjusted to meet the divisibility condition, and the number of padding tokens is pad_frames * latent_H * latent_W. However, we can directly partition along the frames * latent_H * latent_W dimension, significantly reducing the number of padding tokens and improving computational efficiency.

Modifications

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello @nono-Sang, 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 more efficient sequence parallelism strategy for video models, moving from frame-level partitioning to a comprehensive frames * H * W dimension sharding. This enhancement aims to drastically reduce the amount of padding required and boost computational efficiency during model inference. The changes involve updating configuration parameters, refining how latent data is sharded, and adapting the core wanvideo model's forward pass to properly manage the new sharding approach and its associated positional embeddings.

Highlights

• Efficient Sequence Sharding: Introduced a new sp_shard_mode configuration option to enable sharding along the frames * H * W dimension for video models, significantly reducing padding tokens and improving computational efficiency compared to frame-level partitioning.
• Dynamic Rotary Positional Embedding (RoPE) Adjustment: Implemented conditional logic within the wanvideo model's forward pass to correctly calculate Rotary Positional Embeddings (RoPE) for both the traditional 'frame' sharding and the new 'sequence' sharding modes, ensuring accurate positional encoding for distributed sequences.
• Integrated Sharding and Gathering Logic: Modified the wanvideo model to handle padding, sharding of hidden states across SP ranks, and subsequent re-gathering using sequence_model_parallel_all_gather after processing, ensuring seamless distributed inference.
• Frame Adjustment Disablement: Disabled the adjust_frames mechanism when the 'sequence' sharding mode is active, preventing redundant adjustments and optimizing the processing pipeline.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Changelog

• python/sglang/multimodal_gen/configs/pipeline_configs/base.py
  • Added sp_shard_mode configuration parameter with a default value of 'frame'.
  • Modified shard_latents_for_sp to bypass frame-level sharding if sp_shard_mode is set to 'sequence'.
• python/sglang/multimodal_gen/configs/sample/sampling_params.py
  • Disabled adjust_frames when sp_shard_mode is configured as 'sequence'.
• python/sglang/multimodal_gen/runtime/models/dits/wanvideo.py
  • Imported sequence_model_parallel_all_gather and get_sp_group for distributed operations.
  • Implemented conditional logic in the forward method to manage 'sequence' sharding for hidden_states, including padding, sharding, and re-gathering.
  • Introduced _compute_rope_for_sequence_shard to calculate Rotary Positional Embeddings tailored for the new sequence sharding scheme.
• python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py
  • Propagated the sp_shard_mode from pipeline_config to the batch object for consistent use across the pipeline.

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 an efficient sequence sharding mechanism for sequence parallelism in video models. The changes allow partitioning along the frames * H * W dimension, which reduces padding and improves computational efficiency compared to the previous frame-level partitioning.

The implementation is well-structured and consistent across the modified files:

• A new sp_shard_mode configuration is added to PipelineConfig to switch between "frame" and "sequence" sharding modes.
• The sampling parameters are adjusted to disable frame adjustment when sequence sharding is enabled, as it's no longer necessary.
• The core logic in WanTransformer3DModel.forward correctly handles sequence padding, sharding, RoPE calculation for sharded sequences, and gathering the results.
• The changes are propagated correctly through the denoising pipeline stage.

The code is clean and the logic appears to be sound. I don't see any issues with the implementation. This is a great improvement for sequence parallelism efficiency.

[mickqian]

/tag-and-rerun-ci

[RubiaCx]

/tag-and-rerun-ci

[mickqian]

/tag-and-rerun-ci

[mickqian]

should we make it default enabled?