ViperEkura 2e29ed45d3 perf: shrink decode tile to BC=16 for higher occupancy ...

- BC=32→16 halves smem (32KB→16KB for D=128), doubling blocks/SM (3→6)
- D=256 now fits STAGES=2 double-buffer in 32KB, eliminating 176-byte spill
- min_tiles_per_split=2 avoids excessive split overhead on small kv
- paged decode: require page_size multiple of BC so tiles stay page-aligned

Benchmark (L20 sm_89, D=128):
- B=1 kv=4096: 0.0134→0.0122ms (+9% BW)
- B=16 kv=2048: 0.0434→0.0352ms (+23% BW)
- B=32 kv=1024: 0.0343→0.0282ms (+22% BW)

2026-07-27 22:44:02 +08:00

..

attn_common.h

refactor: stride-based attn interface with layout and causal mask

2026-07-14 21:34:42 +08:00

attn_decode_split_kv_mma.cuh

perf: extend MMA decode to arbitrary GQA ratio, add launch bounds, vectorize combine

2026-07-27 00:35:34 +08:00

attn_decode_split_kv.cuh

perf: extend MMA decode to arbitrary GQA ratio, add launch bounds, vectorize combine

2026-07-27 00:35:34 +08:00

attn_decode.cu

refactor: extract shared dispatcher header, unify MMA/scalar dispatch format

2026-07-21 22:21:39 +08:00

attn_dispatchers.cuh

perf: shrink decode tile to BC=16 for higher occupancy

2026-07-27 22:44:02 +08:00

attn_entry_utils.cuh

perf: extend MMA decode to arbitrary GQA ratio, add launch bounds, vectorize combine

2026-07-27 00:35:34 +08:00

attn_mma_utils.cuh

refactor: adopt FA2-style KernelTraits + compile-time causal/mask dispatch

2026-07-21 21:52:46 +08:00

attn_paged_decode_split_kv_mma.cuh

perf: extend MMA decode to arbitrary GQA ratio, add launch bounds, vectorize combine

2026-07-27 00:35:34 +08:00

attn_paged_decode_split_kv.cuh

perf: extend MMA decode to arbitrary GQA ratio, add launch bounds, vectorize combine

2026-07-27 00:35:34 +08:00

attn_paged_decode.cu

refactor: extract shared dispatcher header, unify MMA/scalar dispatch format

2026-07-21 22:21:39 +08:00

attn_prefill_split_q_mma.cuh

refactor: adopt FA2-style KernelTraits + compile-time causal/mask dispatch

2026-07-21 21:52:46 +08:00

attn_prefill_split_q.cuh

fix: improve attention kernel numerical stability and test precision checks

2026-07-21 23:05:17 +08:00

attn_prefill.cu

refactor: extract shared dispatcher header, unify MMA/scalar dispatch format

2026-07-21 22:21:39 +08:00

attn_warp_utils.cuh

perf: extend MMA decode to arbitrary GQA ratio, add launch bounds, vectorize combine

2026-07-27 00:35:34 +08:00