Adding flash attention for sequence parallel #565
Conversation
dianaml0
requested review from
suchenzang,
stephenroller,
ngoyal2707,
punitkoura,
moyapchen,
klshuster,
ruanslv,
davides,
igormolybogFB and
Xirider
as code owners
dianaml0
force-pushed
the
flash_seqpar_v2
branch
from
fe89b3b
to
16c91d1
Compare
|
CircleCI failure not related to this PR |
|
Can we rebase for checks? Should we be concerned about the last bits of numerical differences? |
dianaml0
force-pushed
the
flash_seqpar_v2
branch
from
2f05657
to
cb3a4ee
Compare
|
@stephenroller just rebased the PR, should be up to date now. The rtol and atol used are the same ones we use for testing in xFormers for all flash attention bwds. I do a small training run to validate, would that be useful? |
dianaml0
force-pushed
the
flash_seqpar_v2
branch
from
cb3a4ee
to
135972e
Compare
|
Looks like everything is passing now after rebasing |
dianaml0
force-pushed
the
flash_seqpar_v2
branch
from
390b808
to
9401934
Compare
dianaml0
force-pushed
the
flash_seqpar_v2
branch
from
9401934
to
d0aa8b6
Compare
|
Hi @dianaml0! Thank you for your pull request. We require contributors to sign our Contributor License Agreement, and yours needs attention. You currently have a record in our system, but the CLA is no longer valid, and will need to be resubmitted. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with If you have received this in error or have any questions, please contact us at [email protected]. Thanks! |
Patch Description
Creating this PR off of #511, so it can be reviewed by @stephenroller
The last commit (3d709db) removes some changes from the sequence parallel code which enabled testing with world size of 1. CI is not currently running the test anyway because CI needs to be updated for the test to run.
The forward and backward tests are passing right now. However in some cases, about .2% of the elements fail
Testing steps
Unit Test
gpu_tests/test_sequence_parallel_transformer_layer.py