Opt:ascend kv separation; dsv3 support graph; fused dequant+swiglu+quant

[chenxu140]

Motivation

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.

[gemini-code-assist]

Summary of Changes

Hello @chenxu140, 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 SGLang's performance on Ascend NPUs. The core changes involve a fundamental redesign of the Key-Value (KV) cache to store Key and Value components separately, enhancing memory efficiency and data transfer. This architectural change directly benefits the integration and performance of the DeepSeekV2 model on Ascend, particularly when operating in graph mode. Additionally, the PR includes targeted improvements to the attention mechanism, Mixture-of-Experts (MoE) gating, and quantization processes specifically for the Ascend platform.

Highlights

• Ascend KV Cache Separation: The KV cache management for Ascend NPUs has been fundamentally refactored to store Key and Value components in separate memory buffers. This separation aims to optimize memory access patterns and data transfer efficiency, particularly beneficial for disaggregated systems.
• DeepSeekV2 Model Optimization for Ascend: DeepSeekV2 model support on Ascend has been significantly enhanced. This includes integrating the model with the Ascend attention backend and adapting its KV cache handling to leverage the newly separated Key and Value buffers, especially for graph mode execution.
• Enhanced Ascend Attention Backend with Graph Mode Support: The Ascend attention backend now includes a dedicated forward_decode_graph path. This new function is designed to optimize attention calculations when running in graph mode, improving performance and better handling of RoPE (Rotary Positional Embeddings) within the KV cache.
• MoE Gating Improvements on Ascend: Improvements have been made to the Mixture-of-Experts (MoE) gating mechanism on Ascend NPUs. The changes enable re-normalization of top-k weights, which can lead to more accurate and stable expert selection.
• Quantization Process Refinements for Ascend: Adjustments have been made to the W8A8_INT8 quantization process for Ascend. These include refining the input offset data type and modifying parameters used in the NPU quantization function, contributing to more precise and efficient quantization.

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. 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 optimizations for Ascend hardware, including KV cache separation and graph mode support for DeepSeekV3 models. The changes are extensive and touch several parts of the system, from memory management to attention backends. My review identifies a critical issue in the torch_native_backend where arguments to scaled_dot_product_attention appear to be incorrect, a medium-severity issue regarding code duplication in the new Ascend attention logic, and a minor redundancy in an assertion. Overall, the changes are well-aligned with the PR's objectives, but the identified issues should be addressed.