[Feature] Support JIT set kv cache #16273
Conversation
Summary of ChangesHello @DarkSharpness, 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 aims to dramatically enhance the performance of KV cache operations in SGLang by integrating a new JIT-compiled CUDA kernel. The current PyTorch-based approach is identified as a bottleneck, and this change introduces highly optimized memory copy operations for KV cache storage. The improvements are particularly beneficial for large language models, and the PR provides extensive benchmarks validating the significant latency reductions achieved. Additionally, it includes minor corrections and structural improvements to the QK-Norm kernel. Highlights
🧠 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 AssistThe 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
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 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
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces a JIT-compiled kernel for setting the KV cache, aiming to improve performance over the existing PyTorch-based implementation. The changes include the new CUDA kernel, its Python interface, corresponding benchmarks, and unit tests. Additionally, it refactors the qknorm kernel by renaming and moving files.
My review focuses on the new JIT kernel implementation and its usage. I've identified a potential performance improvement in the Python wrapper for the new kernel and a minor style issue in one of the benchmark files. Overall, the changes look good and the performance gains shown in the benchmarks are impressive.
DarkSharpness
requested review from
Ying1123,
hanming-lu,
hnyls2002,
merrymercy,
xiezhq-hermann and
yizhang2077
as code owners
|
/tag-and-rerun-ci |
1 participant
Motivation
Currently, SGLang uses torch native API to store key/value to cache, which is highly inefficient. Even overlapped in 2 streams, the performance is still poor.
Modifications
This PR is a superset of #9775, the AOT kernel in SGLang. We introduce many aggressive optimizations to minimize the latency, especially for cases where
num_kv_head * head_dimis large (e.g. 1024 forLlama 3.1 8Bon 1 GPU)This PR also fixes some minor errors in
qknorm, and movenorm.cuhtoelementwise/qknorm.cuh.Accuracy Tests
Benchmarking and Profiling
Latency (μs) on B200. PyTorch 2 Stream is current SGLang implementation.
e2e throughput gain for
Llama 3.1 8Bon B200 insend_one: 248.5 -> 254.5Checklist
Review Process
/tag-run-ci-label,/rerun-failed-ci,/tag-and-rerun-ci) or contact authorized users to do so.