[Feature] Optimize DeepSeek's DeepEP on Ascend NPU

[iforgetmyname]

Motivation

Following our roadmap #8004, this is the first pr to support deepep expert parallelism and speed up with some fusion kernels. Now it is possible to activate --enable-deepep-moe with --deepep-mode low_latency only under pd disaggregation scenario on decode nodes. However, we only support W8A8 int8 quant method currently.

More info about our DeepEP-compatible kernels, check out our roadmap in sgl-kernel-npu

Modifications

• a new DeepEPMoE impl w/ gemm and swiglu fusion kernels
• gating_topk kernel using fp32 precision to improve accuracy
• introducing fused rms_norm kernel and add_rms_norm kernel
• introducing fused rope kernel under certain condition, waiting to be generalized

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://slack.sglang.ai to discuss your PR.

Accuracy & Performance

Accuracy with python3 -m sglang.test.few_shot_gsm8k --num-questions 1318

Performance on 2xA3 PD disaggregation w/o DP-ATTN

Code Format

[gemini-code-assist]

Summary of Changes

Hello @iforgetmyname, 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 significant optimizations for DeepSeek's DeepEP (Deep Expert Parallel) Mixture of Experts (MoE) implementation, specifically targeting Huawei Ascend NPUs. It integrates NPU-specific kernels and quantization methods for key operations like MoE gating, rotary embeddings, and W8A8 quantization, aiming to enhance performance and efficiency on Ascend hardware.

Highlights

• NPU-Specific DeepEP MoE Implementation: Introduced AscendDeepEPMoE to leverage Ascend NPU capabilities for DeepEP MoE layers. This includes specialized weight loading for W8A8 scales and an NPU-optimized forward pass utilizing torch_npu operations like npu_grouped_matmul and npu_swiglu. The get_moe_impl_class function now conditionally returns this NPU-optimized class.
• W8A8 Quantization for NPU: Added W8A8EPMoEMethod to support W8A8 integer quantization specifically for EPMoE on Ascend NPUs, enabling more efficient memory usage and computation. The EPMoE initialization logic was updated to integrate this new quantization method.
• Optimized NPU Kernels for Core Operations: Integrated torch_npu kernels for critical operations: MoE Top-K gating (npu_moe_gating_top_k) and Rotary Embedding (npu_mrope). This provides significant performance improvements for these components when running on Ascend NPUs, with fallbacks to native implementations for unsupported configurations.
• Refined W8A8 INT8 Quantization Methods: Enhanced existing W8A8 INT8 quantization methods for NPU by optimizing RMSNorm with npu_add_rms_norm and npu_rms_norm. Additionally, the bias fusion logic in linear layers was refined to ensure correct bias application and prevent double-addition in tensor parallelism scenarios.
• NPU-Aware Conditional Logic and Global Arguments: Adjusted conditional imports for certain quantization kernels to be NPU-aware, ensuring that only relevant kernels are loaded based on the hardware. The 'quantization' argument was also added to the global server arguments, allowing for dynamic configuration of quantization schemes.

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 in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

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 issue 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 or fill out our survey to provide feedback.

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 optimizations for DeepSeek's DeepEP on Ascend NPU. The changes are extensive, adding NPU-specific kernels and logic for MoE layers, quantization, and rotary embeddings.

My review has identified a few critical issues that need to be addressed:

• There's a method signature mismatch in AscendDeepEPMoE.forward that will cause runtime errors.
• The NPU-specific path in select_experts returns an incorrect type, which will lead to unpacking errors.
• There are also some bugs in the weight loading logic for NPU that need to be fixed.

Additionally, I've pointed out some areas for code improvement, such as moving local imports to the top level and removing dead code, to enhance maintainability.

Overall, this is a significant feature addition. Please address the critical issues to ensure correctness.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[Alcanderian]

I cannot see any npu deepep calls in this PR. Are they hidden behind deepep-compatible interfaces?

And VLLM test failed

[iforgetmyname]

I cannot see any npu deepep calls in this PR. Are they hidden behind deepep-compatible interfaces?

yes, refer to the interface

And VLLM test failed

looks like fixed in #8859 and #8881

[ping1jing2]

I cannot see any npu deepep calls in this PR. Are they hidden behind deepep-compatible interfaces?

And VLLM test failed

seems like #8837 is the cause. 8837/checks

[ping1jing2]

I cannot see any npu deepep calls in this PR. Are they hidden behind deepep-compatible interfaces?

And VLLM test failed

see #8886