Set dropout in SDPA to 0.0 when not in training mode #1803
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
facebook-github-bot
added
the
CLA Signed
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/1803
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit eb9a317 with merge base 5de5001 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
RdoubleA
approved these changes
Oct 10, 2024
mori360
pushed a commit
to mori360/torchtune
that referenced
this pull request
Oct 14, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
As pointed out by @zjost in #1791, we probably should just manually force dropout to be 0 outside of training. This is actually what we did originally, but along the way we switched to just always using the value from
attn_dropoutdirectly, which will not be correct at inference time if someone passes a nonzero value. So this PR changes back to coercing dropout to 0.0 outside of training mode. I checked with the author of the PR who first dropped the if/else dropout logic and it was not done by design.I also make two other changes: (1) reverting the doc update I made last night, and (2) raising an error in case someone tries to use FlexAttention with nonzero dropout (it's not currently supported).