Fix the sequence-parallelism for the dense model architecture #4530
Conversation
This file contains 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
RezaYazdaniAminabadi
requested review from
jeffra,
tjruwase,
samyam and
mrwyattii
as code owners
RezaYazdaniAminabadi
changed the title
samadejacobs
requested review from
samadejacobs
and removed request for
samyam
tjruwase
reviewed
Oct 21, 2023
tjruwase
reviewed
Oct 26, 2023
| @@ -1395,15 +1397,15 @@ def allreduce_bucket(self, bucket, rank=None, log=None): | |||
|
|
|||
| tensor_to_allreduce = tensor | |||
|
|
|||
| if pg_correctness_test: | |||
| if pg_correctness_test or self.sequence_parallel_size > 1: | |||
There was a problem hiding this comment.
self.sequence_parallel_size >1 is now redundant given the ds_config flag, right?
baodii
pushed a commit
to baodii/DeepSpeed
that referenced
this pull request
Nov 7, 2023
…oft#4530) Co-authored-by: Masahiro Tanaka <mtanaka@microsoft.com> Co-authored-by: Olatunji Ruwase <olruwase@microsoft.com> Co-authored-by: Sam Ade Jacobs <samjacobs@microsoft.com> Co-authored-by: Michael Wyatt <michaelwyatt@microsoft.com>
mauryaavinash95
pushed a commit
to mauryaavinash95/DeepSpeed
that referenced
this pull request
Feb 17, 2024
…oft#4530) Co-authored-by: Masahiro Tanaka <mtanaka@microsoft.com> Co-authored-by: Olatunji Ruwase <olruwase@microsoft.com> Co-authored-by: Sam Ade Jacobs <samjacobs@microsoft.com> Co-authored-by: Michael Wyatt <michaelwyatt@microsoft.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
This PR fixes some convergence issues for when SP > 1.
We have seen that the gradients were lower when using SP=2 for a dense model, and by further investigation, we find that the gradients were scaled with the total world size, however, they should have been summed across the SP ranks and averaged on the DP-world. Here is the initial curve comparing grad norm of SP=1 (grey) vs SP=2 (green):
After adding the fix for scaling the gradients using the right scale, we get parity for the grad_norm, however, it keeps gradually increasing over time, and results in inferior LM validation (orange: SP1, grey: SP-2).
Fortunately, we are able to fix this by increasing the precision of the gradients before summing them up. The following curves show the LM validation loss of different cases of debugging the SP convergence issue (orange: SP1, grey: SP-2(bf16 gradient), blue: SP2(fp32 gradient)):
cc: @samadejacobs @tohtana