Enable evaluation during deepspeed Sequence Parallel #3917
+1
−0
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
SunMarc
reviewed
Jan 30, 2026
| seq_length_is_variable=sp_handler.sp_seq_length_is_variable, | ||
| core_attn_implementation=sp_handler.sp_attn_implementation, | ||
| micro_batch_size=batch_size_per_device, | ||
| disable_in_eval=True, |
Member
There was a problem hiding this comment.
this could become an arg in sp_handler but fine to first hardcode it. Also disable_in_eval is not available in the latest release of deepseed no ?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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Outdated suggestions cannot be applied.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.
What does this PR do?
Hello accelerate team!
First, I opened this PR to enable evaluation when performing Sequence Parallel.
The history of this PR is connected to different PRs like this:
huggingface/transformers#43517 > deepspeedai/DeepSpeed#7821 > accelerate PR
(I didn't know PRs would be chained like this lol)
So, the core maintainer in deepspeed judged that we need to ask for opinions from the accelerate side, so this PR was opened.
First, through this PR, in accelerator.py > _prepare_deepspeed's SP applying part, we add
disable_in_evalto skip Sequence Parallel during evaluation.(Why we need to skip it, check huggingface/transformers#43517 )
I hardcoded
disable_in_evalfor now, but this needs more discussion.We might need to make it configurable for users, like adding
disable_in_evaltoParallelismConfigor handling it conditionally, etc.cc: @tohtana
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
@SunMarc