ViperEkura a1a1a6bf0f perf: pack gqa q-heads per prefill block to reuse kv tiles ...

- pack HB = min(G, WARPS) q heads per block; K/V tiles stream once per block instead of once per q head
- G=1 keeps the old grid; paged path splits 64-row host Q tiles into HB blocks along grid.x (host maps unchanged)

Benchmark: NVIDIA RTX 6000D, short-q/long-kv prefill 1.4-3.4x (G=8 B=16 q=16 kv=16k 4.22 -> 1.26 ms); full prefill/MHA/paged unchanged (compute-bound); verified vs SDPA G in {1,2,3,4,8,32}, 99 tests pass

2026-08-31 13:39:47 +08:00

..

attention

perf: pack gqa q-heads per prefill block to reuse kv tiles

2026-08-31 13:39:47 +08:00

common

perf: interleave prefetch into the mma phase

2026-08-26 17:47:24 +08:00

fp8

refactor: split fp8 gemm device code into layered headers

2026-08-28 17:29:13 +08:00

rotary

refactor: reorganize CUDA kernels into per-family directories

2026-08-22 20:40:31 +08:00