refactor(kv-cache): one declaration per fact about the KV pools - #2147

Merged

valarLip merged 26 commits into

mainfrom

refactor/m3-index-topk-single-selector

Sep 7, 2026

Commits on Sep 5, 2026

refactor(minimax-m3): keep one indexer selector, not two
valarLip
committed
6487407
perf(deepseek-v4): raise the FP4 MQA grid floor to 4096
valarLip
committed
e9b4934

Commits on Sep 6, 2026

refactor(pool-layout): name the arena after its axis, not the pool it serves
valarLip
committed
edd1292
refactor(kv-cache): one declaration of what an MHA block holds
valarLip
committed
15952e7
refactor(kv-cache): give MLA the same one-declaration pool as MHA
valarLip
committed
9275b3a
refactor(kv-cache): declare M3's indexer cache, and hand a builder its block count
valarLip
committed
1a55b80
refactor(kv-cache): one allocation for every paged pool, one region each
valarLip
committed
02805ad
refactor(kv-cache): let the bind walk hand out pool rows, and the modules declare the pools
valarLip
committed
481f7af
refactor(selector): one attention family, decided where the backends are
valarLip
committed
a6e7ca2
refactor(kv-cache): declare what a V4 KV row and indexer block hold, once
valarLip
committed
3629301
refactor(kv-cache): give a module its pool row, instead of counting the walk
valarLip
committed
164bcc0
style(mha): declare BLOCK_TABLE_EXTENDER a ClassVar
valarLip
committed
95fd44c
fix(minimax-m3): stop the indexer tests from aborting collection
valarLip
committed
b552e94
chore(hooks): report what ruff could not fix, and clear what it reported
valarLip
committed
acb112f

Commits on Sep 7, 2026