Flash-Attn Single-Arch Compilation | Generated by AI
Status update:
- 1 nvcc process — flash-attn compiles
.cufiles sequentially regardless of MAX_JOBS/ninja - 2 .o files done, currently on
flash_bwd_hdim128_bf16_sm80.cu - 9 tmpxft temp files — still early (backward pass, hdim 128)
- Arch: sm_89 only — this is the real speedup (~4x faster per file vs 4-arch build)
MAX_JOBS=32 can’t help here — flash-attn’s build system compiles one .cu file at a time. The bottleneck is per-file CUDA compilation. With single-arch, each file should take 1-3 min instead of 5-12 min. Estimate ~30-40 min total remaining.
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋