wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'ae' into ae-flash-ampere

Browse Source
This commit is contained in:
Hansung Kim
2025-01-30 01:16:21 -08:00
parent 5ab2cc4334 6bdc6af607
commit 30001b7677
1 changed files with 6 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
kernels/flash_attention/compile_flash.sh
View File

@@ -26,7 +26,7 @@ ln -sf input.b.rand.fp32.seqlen1024headdim64.row.bin input.b.bin
ln -sf input.c.rand.fp32.seqlen1024headdim64.row.bin input.c.bin ln -sf input.c.rand.fp32.seqlen1024headdim64.row.bin input.c.bin
for arch in "${archs[@]}"; do for arch in "${archs[@]}"; do
git checkout ae-$arch git checkout ae-flash-$arch
# re-compile libvortexrt.a # re-compile libvortexrt.a
# FIXME after restructure # FIXME after restructure
@@ -34,7 +34,6 @@ for arch in "${archs[@]}"; do
make make
popd popd
for dim in "${dims[@]}"; do
echo "compiling flash_attn kernel for $arch with seqlen 1024, headdim 64" echo "compiling flash_attn kernel for $arch with seqlen 1024, headdim 64"
# touch source file to force re-building, as the Makefile does not track # touch source file to force re-building, as the Makefile does not track
@@ -42,5 +41,4 @@ for arch in "${archs[@]}"; do
touch kernel.cpp touch kernel.cpp
make CONFIG=flash.$arch.seqlen1024.headdim64 make CONFIG=flash.$arch.seqlen1024.headdim64
done
done done