8000 [ROCm] CK Flash Attention Backend by xw285cornell · Pull Request #143695 · pytorch/pytorch · GitHub

[ROCm] CK Flash Attention Backend#143695

Closed

xw285cornell wants to merge 39 commits intopytorch:mainfrom

alugorey:rocm_ck_sdpa

Commits on Jan 2, 2025

[ROCm] CK Flash Attention Backend

alugorey
authored and
pytorchmergebot
committed
Fix test_meta_outplace

alugorey
authored and
pytorchmergebot
committed
Add Tri Dao to LICENSE file for functionality pulled from flash-attention

alugorey
authored and
pytorchmergebot
committed
Add ck kernel blob list

alugorey
authored and
pytorchmergebot
committed
Remove generated kernels from .gitignore

alugorey
authored and
pytorchmergebot
committed
Add generated files to git

alugorey
authored and
pytorchmergebot
committed
remove plumbing to generate files

alugorey
authored and
pytorchmergebot
committed
remove examples dir from inclusion

alugorey
authored and
pytorchmergebot
committed
Add files from example ck

alugorey
authored and
pytorchmergebot
committed
Remove references to ck examples

alugorey
authored and
pytorchmergebot
committed
Moved .h file to .hpp

alugorey
authored and
pytorchmergebot
committed
fix broken tests

xw285cornell
authored and
pytorchmergebot
committed
Re-add file generation

alugorey
authored and
pytorchmergebot
committed
Fix bad cp/paste in .gitignore

alugorey
authored and
pytorchmergebot
committed
Fix lint

alugorey
authored and
pytorchmergebot
committed
fix cmakelists

alugorey
authored and
pytorchmergebot
committed
Remove old instance files from old location

alugorey
authored and
pytorchmergebot
committed
Fix fmha_bwd/fwd.hpp

alugorey
authored and
pytorchmergebot
committed
Remove old blob lists

alugorey
authored and
pytorchmergebot
committed
add block_table for varlen_fwd

alugorey
authored and
pytorchmergebot
committed
fix sdpUtils

alugorey
authored and
pytorchmergebot
committed
re-Add generated instances

alugorey
authored and
pytorchmergebot
committed
Remove ck cmakelists

alugorey
authored and
pytorchmergebot
committed
remove calling ck cmakelists

alugorey
authored and
pytorchmergebot
committed
Add blob lists

alugorey
authored and
pytorchmergebot
committed
Prevent user from setting ck as backend on unsupported arch

alugorey
authored and
pytorchmergebot
committed
remove generated files from .gitignore

alugorey
authored and
pytorchmergebot
committed
lint

alugorey
authored and
pytorchmergebot
committed
Add guards on USE_CK_FLASH_ATTENTION

alugorey
authored and
pytorchmergebot
committed
Add target_compile_definition

alugorey
authored and
pytorchmergebot
committed
Removing stale ck_kernel_blob_list.txt

jithunnair-amd
authored and
pytorchmergebot
committed
Remove ck/fwd_blob_list.txt and ck/bwd_blob_list.txt because they contain local file paths and are not really necessary to be upstreamed

jithunnair-amd
authored and
pytorchmergebot
committed
Rename CK autogenerated files using sha1sum to address Windows failure during git clone due to long file paths eg. https://github.com/pytorch/test-infra/actions/runs/12376044163/job/34542289317#ste…

jithunnair-amd
authored and
pytorchmergebot
committed
Remove flash_api.h from list of includes for hipify

jithunnair-amd
authored and
pytorchmergebot
committed
Add informative messages when USE_CK_FLASH_ATTENTION is enabled

jithunnair-amd
authored and
pytorchmergebot
committed
Use Cmake STATUS

jithunnair-amd
authored and
pytorchmergebot
committed
Add warning to users if building for more than one gfx arch in PYTORCH_ROCM_ARCH

jithunnair-amd
authored and
pytorchmergebot
committed
makefile linter

xw285cornell
authored and
pytorchmergebot
committed
Merge branch 'ROCm:main' into rocm_ck_sdpa
xw285cornell
authored

0