NVIDIA / NVIDIA/recsys-examples

[QST] Timeline for "native page tables in decode attention" for SID-GR — variable-length online throughput gap

Open
#445 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

question
Dominant language
Python
Stars
322
Forks
83
Avg merge
3d 3h
Merged PRs (30d)
8

Description

README §10 (ContextKV memory strategy) lists "eventually support native page tables in decode attention".

On variable-length online input (128–256 tok, beam=50), SID-GR trails SGLang by ~2x - 3x on L20, with decode collapsing to all batch=1. This gap can't really close until paged decode attention lands — the interim steps (multi context buckets, page-backed ContextKV) only mitigate.

Is there a rough timeline / priority for native paged decode attention? Are the interim steps planned sooner?

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with README §10, especially the ContextKV memory strategy and its note about native page tables in decode attention. Review the reported variable-length online benchmark conditions and determine whether the project has documented plans for paged decode attention, multi-context buckets, or page-backed ContextKV; done means providing a confirmed priority or timeline.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
performance
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.