Repository navigation

Osimons/gdn mma prefill perf impro - #59

Open

ORippler wants to merge 7 commits into

am17an:aman/gdn-mma-prefillfrom

ORippler:osimons/gdn-mma-prefill-perf-impro

Open

Osimons/gdn mma prefill perf impro#59
ORippler wants to merge 7 commits into
am17an:aman/gdn-mma-prefillfrom
ORippler:osimons/gdn-mma-prefill-perf-impro

Commits on Oct 8, 2026
- Remove lmem by indexing base_output with compile-time constants
  ORippler
  committed
  83d5c9d
- Increase arithmetic density by mocked mma.sync.m16n16k16
  ORippler
  committed
  5b1b63d
Commits on Oct 9, 2026
- Hoist loop invariants
  ORippler
  committed
  4d2e4dc
- Add store4 to avoid smem bank conflicts for qk storing
  ORippler
  committed
  1c8e55a
- Load float8 to get compiler to unroll q/k fetches
  ORippler
  committed
  26de4ef
- Hide latency of vv by prefetching/doing manual scheduling
  ORippler
  committed
  9991bac
- Recompute state_out to avoid lmem
  ORippler
  committed
  ce37862