[overlap-spec] support page size > 1 #11772
Conversation
Summary of ChangesHello @hnyls2002, 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 speculative decoding mechanism, specifically for the Eagle algorithm, by introducing robust support for memory page sizes greater than one. This change allows the system to manage memory more efficiently when dealing with larger page allocations, improving performance and resource utilization in scenarios where speculative decoding is employed, and removing previous limitations for paged memory usage. Highlights
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
|
There was a problem hiding this comment.
Code Review
This pull request adds support for page_size > 1 in speculative decoding with overlap, which is a great enhancement. The changes correctly handle paged KV cache allocation and deallocation by aligning to page boundaries. I've included a couple of suggestions to improve performance by moving an import out of a loop and optimizing a tensor calculation to reduce GPU-CPU synchronization.
| last_loc = get_last_loc( | ||
| batch.req_to_token_pool.req_to_token, | ||
| batch.req_pool_indices, | ||
| self.allocate_lens, |
There was a problem hiding this comment.
if self.allocate_lens is None (first call), then I have cuda IMA. The fix:
self.allocate_lens = batch.seq_lens.clone() if self.allocate_lens is None
Test to validate.
so topk=1, pagesize=2,4,8 etc.
You can test it like:
SGLANG_ALLLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 then
python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --speculative-algorithm EAGLE3 --speculative-draft-model-path lmsys/sglang-EAGLE3-LLama3.1-Instruct-8B --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --page-size 2 --enable-beta-spec --dtype float16
I am try to impl topk > 1 and pagesize >1; based off this branch.
There was a problem hiding this comment.
The allocate_lens would not be None even when it is the first call (first decode step after prefill). The allocate_lens is initialized here
Or there are some corner cases I haven't considered. Please provide more reproduction scripts. The code on this branch can pass the unit test of test_eagle_infer_beta.py (by setting page to 64).
No description provided.