feat: overlap shared experts with send/recv

[AlpinDale]

wip

[gemini-code-assist]

Summary of Changes

Hello @AlpinDale, 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 significant optimization for Mixture-of-Experts (MoE) models that include shared experts. The primary goal is to enhance computational efficiency by allowing the processing of shared experts to occur concurrently with the data communication required for the routed experts. This is achieved through a new architectural pattern involving asynchronous kernel operations and a dedicated SharedFusedMoE layer, which orchestrates the overlapping of compute and communication, ultimately leading to faster inference for such models.

Highlights

• Shared Experts Overlap: Introduced a new SharedFusedMoE class and modified the core FusedMoE layer to enable overlapping the computation of shared experts with the all-to-all communication (send/recv) operations required for routed experts. This aims to improve performance by hiding communication latency.
• Asynchronous MoE Kernel Operations: Implemented prepare_async and supports_async methods in various MoE kernels (deepep_ht_prepare_finalize.py, deepep_ll_prepare_finalize.py, pplx_prepare_finalize.py). These methods allow the dispatch phase of MoE communication to proceed asynchronously, returning a 'receiver' callback that can be invoked later to collect results, facilitating overlap with other computations like shared experts.
• New Custom Operator for Shared MoE: Added a new custom PyTorch operator, moe_forward_shared, along with its fake implementation, to specifically handle the forward pass when both shared and fused experts are present, returning a tuple of outputs for both.
• Model Integration: Updated model implementations for DeepseekV2 and Llama4 to utilize the new SharedFusedMoE class, correctly integrating the shared expert computation and handling the new tuple return type from the MoE layer.
• Type Hinting and Refactoring: Extensively updated type hints across various MoE-related files to reflect the new possibility of methods returning a Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]], accommodating the separate outputs from shared and fused experts. Also, refactored FusedMoE's forward pass logic to manage shared expert processing and output reduction.

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 significant optimization to overlap shared expert computation with the all-to-all communication in MoE layers. This is achieved by refactoring the prepare methods into an asynchronous prepare_async that returns a receiver callable. The core logic is well-encapsulated in FusedMoEModularKernel and the new SharedFusedMoE class. The changes are extensive but follow a clear pattern across multiple files to support this new asynchronous execution flow. My review focuses on improving the robustness of module type checking.

[gemini-code-assist]

Using __class__.__name__ for checking the module type is brittle and can break with subclassing or refactoring. A more robust approach would be to check for the presence of specific attributes, as suggested in the TODO comment. This avoids issues with subclassing or renaming.

            if hasattr(module, "quant_method") and hasattr(
                module.quant_method, "init_prepare_finalize"
            )

[gemini-code-assist]

Using __class__.__name__ for checking the module type is brittle and can break with subclassing or refactoring. A more robust approach would be to check for the presence of specific attributes, which avoids issues with subclassing or renaming.

            if hasattr(module, "quant_method") and hasattr(
                module.quant_method, "init_prepare_finalize"
            )