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

..

kernels

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

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

tests

fix: improve attention kernel numerical stability and test precision checks

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

__init__.py

feat: add optional CUDA kernel system (csrc/) + fused GQA decode attention

2026-07-06 12:09:58 +08:00

build.py

chore: bump version to 1.3.9

2026-07-12 21:04:20 +08:00