ViperEkura 41cd40363a perf: post-multiply attention scale in float instead of pre-scaling Q in bf16 ...

- Replace bf16 pre-scale Q loading with direct 32-bit aligned bf16x2 reads
- Apply scale in float32 after Q@K^T, before online softmax
- Reduces causal max error from 2^-6 to 2^-8 with zero perf cost

2026-07-11 00:13:32 +08:00

..

kernels

perf: post-multiply attention scale in float instead of pre-scaling Q in bf16

2026-07-11 00:13:32 +08:00

tests

refactor: rename gqa_* to attn_*, split-KV for all decode paths

2026-07-10 23:35:14 +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

refactor: rename gqa_* to attn_*, split-KV for all decode paths

2026-07-10 23:35:14 +08:00