SWA Prefix Cache #7367
SWA Prefix Cache #7367
Conversation
There was a problem hiding this comment.
Summary of Changes
Hello @hanming-lu, 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 integrates a new, specialized radix tree data structure for managing the KV cache within the SGLang framework. The SWARadixCache is tailored to optimize memory efficiency, particularly for models employing Sliding Window Attention, by enabling intelligent prefix sharing and eviction strategies.
Highlights
- New Component Introduction: This PR introduces a new module,
swa_radix_cache.py, which defines theSWARadixCacheclass and its supporting structures. This class is a specialized radix tree implementation. - KV Cache Management: The primary purpose of the
SWARadixCacheis to efficiently manage the Key-Value (KV) cache for large language models. It stores and retrieves token prefixes to optimize memory usage. - Sliding Window Attention (SWA) Integration: The cache is specifically designed to support models utilizing Sliding Window Attention (SWA). It incorporates SWA-aware logic for prefix matching, eviction, and memory allocation, including concepts like 'swa_tombstone' nodes.
- Memory Optimization and Eviction: It provides mechanisms to track and manage both 'full' and 'SWA' evictable token counts, along with methods to increment/decrement lock references on cache nodes, ensuring that in-use KV cache entries are protected from eviction.
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 is currently in preview and 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 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
-
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. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces SWARadixCache, a new radix tree implementation for managing KV cache, tailored for Sliding Window Attention (SWA). Key suggestions include using time.monotonic() for more reliable LRU timestamps and aligning the match_prefix return type with the BasePrefixCache interface.
0eddd7d to
a8a6de6
Compare
acc52a2 to
7823c12
Compare
7823c12 to
a21d35d
Compare
b40f8a2 to
c2682b2
Compare
Co-authored-by: Ying Sheng <sqy1415@gmail.com>
Co-authored-by: Ying Sheng <sqy1415@gmail.com>
Co-authored-by: Ying Sheng <sqy1415@gmail.com>
Co-authored-by: Ying Sheng <sqy1415@gmail.com>
Motivation
Modifications
Testing
Loogle:
python python/sglang/eval/loogle_eval.py --api-url http://127.0.0.1:30002/v1llama4 fa3 disable swa radix cache: average BERTScore (F1): 84.50%:
python3 -m sglang.launch_server --model-path meta-llama/Llama-4-Scout-17B-16E-Instruct --port 30002 --tp 8 --mem-fraction-static 0.8 --context-length 100000 --attention-backend fa3 --disable-radix-cache --hybrid-kvcache-ratio 0.95llama4 fa3 enable swa radix cache: average BERTScore (F1): 84.32%
python3 -m sglang.launch_server --model-path meta-llama/Llama-4-Scout-17B-16E-Instruct --port 30002 --tp 8 --mem-fraction-static 0.8 --context-length 100000 --attention-backend fa3 --hybrid-kvcache-ratio 0.95Checklist