Add GQA support for ROCm #21032
cloudhan
merged
18 commits into
main
from
guangyunhan/rocm-gqa
feat: init rocm gqa
c877adcb
feat: extend strided copy to support runtime tok idx
f845099d
more case
0ea33352
feat: local
99b2feb6
feat: rotary
816249c8
feat: allow rotary to read and write in a strided way, so that we don…
6024dc96
fix: rotary for packed qkv
48092eee
remove debug print
de2f30ae
github-advanced-security
commented
on 2024-06-13
workaround: add flash_attn test to ci
6c4e6125
add gpu arch checking warning log
e9f6d13f
fix: build without ck tile
2b0c46ed
test: update ci pytorch and triton version to fix tests which have fa…
6091a69b
format
8ca06340
remove unused param is_input_bnsh_format from strided version LaunchR…
e22dfb9e
make onnxruntime_USE_COMPOSABLE_KERNEL_CK_TILE depends on onnxruntime…
789fee73
skip test_flash_attn_rocm on CUDA platform
b973217f
github-advanced-security
commented
on 2024-07-01
fix lint and ci
c3c7089d
fix typo
f4355d46
tianleiwu
approved these changes
on 2024-07-02
mszhanyi
approved these changes
on 2024-07-03
Reviewers
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub