Change head_size parameter dependent on qkv_hidden_size #12933
petermcaughan
merged
22 commits into
main
from
petermca/qkv_cuda_support
Change head_size parameter dependent on qkv_hidden_size
372fa91c
Remove bugged code, add CUDA attention UT
de3b573a
Introduce head_sizes and hidden_sizes for qkv, insert where relevant
d9664ae2
Merge with main
655dc652
Fix linter warnings
142adaad
UT passing, saving work. Needs clean
2124660e
Passing unit tests!
0abbe436
Remove print statements
318afccd
Simplify UT
a9a370f7
Remove dump objects
b42a7eeb
Resolve merge conflict
868c5f87
Undo undesired change
8c1344ab
Undo undesired change
79f0c981
tianleiwu
commented
on 2022-09-28
tianleiwu
commented
on 2022-09-28
tianleiwu
commented
on 2022-09-28
tianleiwu
commented
on 2022-09-28
tianleiwu
commented
on 2022-09-28
tianleiwu
commented
on 2022-09-28
Address comments
d202fe4a
Fix linter warnings & comments
3008846f
Fix variable names
c0dbe593
tianleiwu
commented
on 2022-09-28
tianleiwu
commented
on 2022-09-28
tianleiwu
commented
on 2022-09-28
tianleiwu
commented
on 2022-09-28
Add fp16 checks for qkv_head_size and remove unused variables
c9d7dfae
Fix longformer test failures
dafc547d
Avoid future silent errors
98337e85
Avoid ROCM execution
f49fa60f
tianleiwu
dismissed these changes
on 2022-10-05
Add support for disabling ROCM in AttentionTest
d49dfa0e
Add comments to clarify feature enablement in nonquantized CUDA only
41ec6790
tianleiwu
approved these changes
on 2022-10-10
Login to write a write a comment.
Login via GitHub