Gemma-4 SWA: per-group physical KV pools and data-plane routing
Nobody has claimed this yet.
- Dominant language
- C++
- Stars
- 423
- Forks
- 53
- Avg merge
- 20h 26m
- Merged PRs (30d)
- 310
Description
Gemma-4 publishes one oversized FullAttention KV group on current main. Generic sliding infrastructure already exists (SlidingWindowSpec, SlidingWindowManager, hybrid coordinator, per-layer specs, group-name masks), but all coordinator managers share one block_pool. Distinct per-group physical pool sizes, conjunctive per-pool admission, and Gemma data-plane routing of the sliding group's own block table/metadata are absent.
Row: ROCM-GEMMA4-SWA-PHYSICAL
This issue tracks the residual only: per-group physical pools plus Gemma publication/data-plane routing. It is not a generic SWA rewrite. Default stays OFF / explicit promotion until CPU and device identity gates. No GPU work on this filing.
Spec-first. Implementation waits on an approved committed spec.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reviewing the existing SlidingWindowSpec, SlidingWindowManager, hybrid coordinator, per-layer specs, and group-name masks mentioned in the issue. Wait for an approved committed spec before implementation; done means per-group physical pools and Gemma sliding-group data-plane routing are defined and implemented with default-off CPU and device-identity gates, without GPU work.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- backend, machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 25/100