[Kernel]: Optimize DispatchFFNCombine performance

[serlar]

What this PR does / why we need it?

This PR focuses on performance optimization for the DispatchFFNCombine operator. The key optimizations include:

1. Improving communication efficiency by merging the transmission of tokens and scales;
2. Decoupling multi-core dependencies and reducing waiting bubbles in the combine process through tile-granularity communication;
3. Optimizing the full-card synchronization overhead before the umpermute operation.

These optimizations aim to reduce the overall execution latency of the DispatchFFNCombine operator and enhance the runtime performance of the model inference process on Ascend devices.

Does this PR introduce any user-facing change?

No. This PR only involves internal performance optimization of the DispatchFFNCombine operator and does not introduce any changes to user-facing APIs, interfaces, or behaviors.

How was this patch tested?

1. Enable the DispatchFFNCombine operator by setting the environment variable:

export VLLM_ASCEND_ENABLE_FUSED_MC2=1

2. Run the standard model inference test suite with the above environment variable enabled;
3. Verify the correctness of model outputs (ensuring no functional regression) and measure the performance improvement of the DispatchFFNCombine operator (reduced latency and improved throughput).

• vLLM version: v0.14.1
• vLLM main: vllm-project/vllm@dc917cc

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Summary of Changes

Hello @serlar, 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 enhances the performance of the DispatchFFNCombine operator, specifically targeting Ascend devices. The core improvements revolve around optimizing inter-core communication by transmitting tokens and their quantization scales together, and by implementing a more efficient, flag-based synchronization mechanism across processing units. These internal changes are designed to reduce execution latency and boost model inference throughput without impacting any user-facing APIs or behaviors.

Highlights

• Merged Token and Scale Transmission: The pull request introduces a significant optimization by merging the transmission of tokens and their corresponding scales. This is achieved through the new CopyGMToGMPerToken function and modifications in the quantization headers (moe_v2_fullload_dynamic_quant.h, moe_v2_gather_dynamic_quant.h) which now allocate combined buffer space (cols_scale_) for both data and scale, improving communication efficiency.
• Optimized Full-Card Synchronization: Synchronization overhead before the unpermute operation has been optimized. New functions like InitArithProgress, CrossRankSyncV2Set, and CrossRankSyncV2Wait in hccl_shmem.hpp implement a more efficient cross-rank synchronization mechanism using soft flags. The BlockMmad class now incorporates this new flag-based synchronization, reducing latency.
• Refactored Kernel Logic: The DispatchFFNCombineKernel has been refactored to merge the Dispatch and Combine operations into a single DispatchAndCombine function. This streamlines the execution flow and simplifies the kernel's overall structure. Additionally, several kernel configurations have been consolidated, and local weight/scale arrays have been removed in favor of direct GetTensorAddr calls for better memory management.
• Enhanced Workspace Management: Workspace allocation for the DispatchFFNCombine operator has been refined. The cocWorkspace calculation in dispatch_ffn_combine_tiling.cpp now includes additional space for synchronization-related data. The WorkspaceInfo struct in dispatch_ffn_combine_kernel.hpp has been updated to separate ptrC and ptrC2, and ptrPermutedToken and ptrA, along with new allocations for ptrSumBeforeRank and ptrSoftFlagBase.
• New Epilogue Dispatch Policy: A new epilogue dispatch policy, EpilogueAtlasA2PerTokenDequantV2, has been introduced and implemented in block_epilogue_pertoken_v2.hpp. This new policy is designed to handle the dequantization and result writing in conjunction with the merged token and scale data, utilizing ping-pong buffers for efficient data processing.

🧠 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.

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 significant performance optimizations for the DispatchFFNCombine operator. The changes focus on improving communication efficiency by merging the transmission of tokens and scales, and on optimizing the synchronization overhead. The implementation involves substantial refactoring across multiple kernel files, including the introduction of new epilogue and synchronization logic.

The review has identified a critical compilation error due to a typo in an #include directive. Additionally, a couple of high-severity issues have been found, including another typo in an include guard and a potential reduction in functionality where a specific case is now handled by trap(), which could lead to runtime failures for certain inputs.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.