PR #15279 ROCm MHA - SemanticDiff

ROCm MHA #15279

cloudhan merged 11 commits into main from guangyunhan/rocm-mha

cloudhan requested a review from

tianleiwu 3 years ago

cloudhan requested a review from

kailums 3 years ago

cloudhan requested a review from

PeixuanZuo 3 years ago

cloudhan requested a review from

abudup 3 years ago

cloudhan requested a review from

zhangyaobit 3 years ago

github-advanced-security commented on 2023-03-30

cloudhan force pushed from 29d1aa7b to 63e2b509 3 years ago

github-advanced-security commented on 2023-03-31

cloudhan force pushed from be7a812d to ec54f81b 3 years ago

abudup requested changes on 2023-03-31

Classify qkv format

4a78362a

Refactor for adding multiple qkv format support

4da92bba

Add MHA and additional qkv_format support for underlying flash attention

173fb73f

Allow Q_K_V_BSNH

b187e868

Make linter happy

48c7f75e

Cleanup leftover when hipifying from CUDA

7252848f

Make it clear what we don't support and note them as TODOs

ae4f1dd5

Minor

05f02972

Fix after rebase

f86af6e0

cloudhan force pushed from 9a0335c2 to f86af6e0 3 years ago

tianleiwu commented on 2023-04-03

cloudhan requested a review from

abudup 3 years ago

Partially enable MHA tests for ROCm

3843112a

abudup dismissed these changes on 2023-04-06

cloudhan requested a review from

tianleiwu 3 years ago

cloudhan requested a review from

gh-yewang 3 years ago

Leave qkv_format unspecified

152ea5a4

cloudhan dismissed their stale review via 152ea5a4 3 years ago

tianleiwu approved these changes on 2023-04-11

cloudhan merged 9acbfc6a into main 3 years ago

cloudhan deleted the guangyunhan/rocm-mha branch 3 years ago

Reviewers

tianleiwu

abudup

gh-yewang

github-advanced-security

kailums

PeixuanZuo

zhangyaobit

Assignees

No one assigned

Labels

None yet

Milestone

No milestone

onnxruntime ROCm MHA #15279 Merged

ROCm MHA #15279

onnxruntime
ROCm MHA
#15279

Merged