-
Notifications
You must be signed in to change notification settings - Fork 45
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add FMHA T5x test #442
Add FMHA T5x test #442
Conversation
Closed the old PR #327 as this one is the latest one. |
Moved to draft since there still seems to be active development on this PR. BTW, it may be better to add this to _test_t5x.yaml since we were already running pretty close to the maximum number of reusable workflows allowed in github. |
abdf793
to
a5c2f5d
Compare
e8d968a
to
89a47cb
Compare
Merginf fmha into t5x test.
Merginf fmha into t5x.
89a47cb
to
cc95535
Compare
Incorporated review comments
fixed indentation
updating the new name for _test_t5x Co-authored-by: Terry Kong <terryk@nvidia.com>
reduced number of jobs for fmha
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Tests that were added passed; metrics fail due to issue already resolved in main
Adding the JAX T5x FMHA E2E system test to check for fmha lowering support. Following are the steps implemented in the test:
FMHA lowering flag is enabled by default now, enabled the dumping of hlo to track fmha forward and backward instructions.
Added the test as part of _ci.yaml file and also added a nightly workflow file for it. We will add this test as part of performance benchmarking later and add hlo to baseline.
Also added changes for correction of seq length of decoder (should be a multiple of 64)
The test was failing with following error related to CUDNN_STATUS_BAD_PARAM. The fix for this is added in the [PR] (openxla/xla#6872) in upstream which is now merged and the test passes. Bug for this error.
run for these changes: workflow run link