dylanyunlong/project_6
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

feat: asymmetric BLOCK_M/BLOCK_N search + re-add BI-V100 autotune configs

Browse Source 
bench_triton_prefill.py:
  - Split --block into --block (BLOCK_M) and --block-n (BLOCK_N)
  - Each (M, N, warps) combo triggers Triton JIT recompilation
  - Enables finding asymmetric optima like M=64,N=32 that save SMEM

triton_flash_attention.py:
  - Re-add 3 BI-V100 autotune configs (64x32, 32x64, 64x64 with warps=4)
  - These were wrongly reverted in 8c1955d -- autotune is zero-risk

run_on_bi100.sh:
  - Updated to use asymmetric block search
This commit is contained in:
Claude
2026-08-03 11:18:18 +00:00
parent fe64650681
commit a2a5dd8f00
3 changed files with 108 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
muh/run_on_bi100.sh
View File

@@ -112,7 +112,7 @@ echo "(Each BLOCK×WARPS combo triggers Triton recompilation into different PTX)
echo ""
python3 muh/bench_triton_prefill.py \
--block 16 32 64 128 \
--block 16 32 64 --block-n 16 32 64 \
--warps 1 2 4 8 \
--ctx-lens 128 512 2048 8192 \
--batch 1 --seq-len 1 \