updating benchmarks (now we can outperforme jax cudnn sdpa attention …

…in every sequence length), caching triton kernels, autotuning
erfanzar · Nov 15, 2024 · 62b7ca2 · 62b7ca2
1 parent f5b7333
commit 62b7ca2
Show file tree

Hide file tree

Showing 574 changed files with 960 additions and 3,526 deletions.
diff --git a/benchmarks/mha/triton/README.md b/benchmarks/mha/triton/README.md
diff --git a/...h_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=128-mode=fwd.csv b/...h_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=128-mode=fwd.csv
diff --git a/...ias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=128-mode=fwd.png b/...ias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=128-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=32-mode=fwd.csv b/...ch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=32-mode=fwd.csv
diff --git a/...bias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=32-mode=fwd.png b/...bias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=32-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=64-mode=fwd.csv b/...ch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=64-mode=fwd.csv
diff --git a/...bias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=64-mode=fwd.png b/...bias=False-headdim=128-num_heads=16-blocksize_q=128-blocksize_k=64-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=128-mode=fwd.csv b/...ch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=128-mode=fwd.csv
diff --git a/...bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=128-mode=fwd.png b/...bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=128-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=32-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=32-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=32-mode=fwd.png b/...-bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=32-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=64-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=64-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=64-mode=fwd.png b/...-bias=False-headdim=128-num_heads=16-blocksize_q=32-blocksize_k=64-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=128-mode=fwd.csv b/...ch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=128-mode=fwd.csv
diff --git a/...bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=128-mode=fwd.png b/...bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=128-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=32-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=32-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=32-mode=fwd.png b/...-bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=32-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=64-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=64-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=64-mode=fwd.png b/...-bias=False-headdim=128-num_heads=16-blocksize_q=64-blocksize_k=64-mode=fwd.png
diff --git a/...h_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=128-mode=fwd.csv b/...h_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=128-mode=fwd.csv
diff --git a/...ias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=128-mode=fwd.png b/...ias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=128-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=32-mode=fwd.csv b/...ch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=32-mode=fwd.csv
diff --git a/...bias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=32-mode=fwd.png b/...bias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=32-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=64-mode=fwd.csv b/...ch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=64-mode=fwd.csv
diff --git a/...bias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=64-mode=fwd.png b/...bias=False-headdim=128-num_heads=32-blocksize_q=128-blocksize_k=64-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=128-mode=fwd.csv b/...ch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=128-mode=fwd.csv
diff --git a/...bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=128-mode=fwd.png b/...bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=128-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=32-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=32-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=32-mode=fwd.png b/...-bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=32-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=64-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=64-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=64-mode=fwd.png b/...-bias=False-headdim=128-num_heads=32-blocksize_q=32-blocksize_k=64-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=128-mode=fwd.csv b/...ch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=128-mode=fwd.csv
diff --git a/...bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=128-mode=fwd.png b/...bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=128-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=32-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=32-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=32-mode=fwd.png b/...-bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=32-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=64-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=64-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=64-mode=fwd.png b/...-bias=False-headdim=128-num_heads=32-blocksize_q=64-blocksize_k=64-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=128-mode=fwd.csv b/...ch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=128-mode=fwd.csv
diff --git a/...bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=128-mode=fwd.png b/...bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=128-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=32-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=32-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=32-mode=fwd.png b/...-bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=32-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=64-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=64-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=64-mode=fwd.png b/...-bias=False-headdim=128-num_heads=8-blocksize_q=128-blocksize_k=64-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=128-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=128-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=128-mode=fwd.png b/...-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=128-mode=fwd.png
diff --git a/...atch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=32-mode=fwd.csv b/...atch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=32-mode=fwd.csv
diff --git a/...1-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=32-mode=fwd.png b/...1-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=32-mode=fwd.png
diff --git a/...atch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=64-mode=fwd.csv b/...atch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=64-mode=fwd.csv
diff --git a/...1-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=64-mode=fwd.png b/...1-bias=False-headdim=128-num_heads=8-blocksize_q=32-blocksize_k=64-mode=fwd.png
diff --git a/...tch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=128-mode=fwd.csv b/...tch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=128-mode=fwd.csv
diff --git a/...-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=128-mode=fwd.png b/...-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=128-mode=fwd.png
diff --git a/...atch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=32-mode=fwd.csv b/...atch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=32-mode=fwd.csv
diff --git a/...1-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=32-mode=fwd.png b/...1-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=32-mode=fwd.png
diff --git a/...atch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=64-mode=fwd.csv b/...atch_size=1-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=64-mode=fwd.csv
diff --git a/...1-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=64-mode=fwd.png b/...1-bias=False-headdim=128-num_heads=8-blocksize_q=64-blocksize_k=64-mode=fwd.png
diff --git a/...h_size=1-bias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=128-mode=fwd.csv b/...h_size=1-bias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=128-mode=fwd.csv
diff --git a/...ias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=128-mode=fwd.png b/...ias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=128-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=32-mode=fwd.csv b/...ch_size=1-bias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=32-mode=fwd.csv
diff --git a/...bias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=32-mode=fwd.png b/...bias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=32-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=64-mode=fwd.csv b/...ch_size=1-bias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=64-mode=fwd.csv
diff --git a/...bias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=64-mode=fwd.png b/...bias=False-headdim=256-num_heads=16-blocksize_q=128-blocksize_k=64-mode=fwd.png
diff --git a/...ch_size=1-bias=False-headdim=256-num_heads=16-blocksize_q=32-blocksize_k=128-mode=fwd.csv b/...ch_size=1-bias=False-headdim=256-num_heads=16-blocksize_q=32-blocksize_k=128-mode=fwd.csv
diff --git a/...bias=False-headdim=256-num_heads=16-blocksize_q=32-blocksize_k=128-mode=fwd.png b/...bias=False-headdim=256-num_heads=16-blocksize_q=32-blocksize_k=128-mode=fwd.png