Release overlap_comm & contiguous_gradients restrictions for ZeRO 1 #4887
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
|
@li-plus, thanks so much for this PR. Your analysis of the problem is accurate. Previously, we did not have bandwidth to evaluate the correctness of enabling those optimizations for ZeRO-1. This is a great contribution. Thanks so much. |
tjruwase
approved these changes
Jan 5, 2024
mauryaavinash95
pushed a commit
to mauryaavinash95/DeepSpeed
that referenced
this pull request
Feb 17, 2024
…icrosoft#4887) The `overlap_comm` and `contiguous_gradients` options have been ignored in ZeRO stage 1 since microsoft#1246. Back in that time, ZeRO 1 and 2 are separately implemented (see https://github.com/microsoft/DeepSpeed/tree/6ae756c03f12674f17aef90622e7664a8af9d2af/deepspeed/runtime/zero). ZeRO 1 does not have gradient hooks registered to overlap backward and gradient all-reduce, so it's fine to ignore `overlap_comm` and `contiguous_gradients`. However, in the current implementation, ZeRO 1 and 2 share almost the same implementation (`stage_1_and_2.py`). Features like `overlap_comm` and `contiguous_gradients` can also be enabled for ZeRO 1 (Please correct me if I made a mistake). With this PR, turning on `overlap_comm` and `contiguous_gradients` for ZeRO 1 on the [SFT task](https://github.com/microsoft/DeepSpeedExamples/tree/master/applications/DeepSpeed-Chat/training/step1_supervised_finetuning) produces exactly the same training curve as the latest master.  I also see a ~1.05x e2e speedup by overlapping backward and gradient all-reduce. I can confirm by the trace that backward and all-reduce do overlap, and the separate gradients are indeed copied to a flat buffer. These options are also effective for ZeRO 1.   Related issue: microsoft#2295 Co-authored-by: Olatunji Ruwase <[email protected]>
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.
The
overlap_commandcontiguous_gradientsoptions have been ignored in ZeRO stage 1 since #1246. Back in that time, ZeRO 1 and 2 are separately implemented (see https://github.com/microsoft/DeepSpeed/tree/6ae756c03f12674f17aef90622e7664a8af9d2af/deepspeed/runtime/zero). ZeRO 1 does not have gradient hooks registered to overlap backward and gradient all-reduce, so it's fine to ignoreoverlap_commandcontiguous_gradients. However, in the current implementation, ZeRO 1 and 2 share almost the same implementation (stage_1_and_2.py). Features likeoverlap_commandcontiguous_gradientscan also be enabled for ZeRO 1 (Please correct me if I made a mistake).With this PR, turning on
overlap_commandcontiguous_gradientsfor ZeRO 1 on the SFT task produces exactly the same training curve as the latest master.I also see a ~1.05x e2e speedup by overlapping backward and gradient all-reduce. I can confirm by the trace that backward and all-reduce do overlap, and the separate gradients are indeed copied to a flat buffer. These options are also effective for ZeRO 1.
Related issue: #2295