OwnerSunshine530 cdc2dd490b feat: Triton varlen因果flash attention(块对角,单kernel,消逐块调用+mask构造开销) ...

每program处理(用户段query块,head),只遍历段内<=该块的key(因果),在线softmax,
fp16读写fp32累加。CONFIG.attn=triton(默认仍chunked);_triton_block_meta每batch算一次
block→段映射8层复用;_resolve_attn在无triton/CPU时回退chunked。等价测试+bench --attn triton。
数学等价(FlashAttention同类,规则允许),不改组网。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

2026-06-17 00:14:53 +08:00

..

test_equiv.py

feat: Triton varlen因果flash attention(块对角,单kernel,消逐块调用+mask构造开销)

2026-06-17 00:14:53 +08:00