ZeRO0 does not handle BF16 gradients properly #5154
Closed
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
tohtana
changed the title
|
This PR breaks PP. Opened #5170 as another solution. |
github-merge-queue bot
pushed a commit
that referenced
this pull request
Feb 21, 2024
This PR fixes an issue with allreducing for ZeRO0 + BF16. (This replaces #5154) DeepSpeed uses `BF16_Optimizer` when ZeRO0 and BF16 are enabled. The optimizer accumulates gradients on FP32 buffer soon after a backward pass completes. However, DeepSpeed engine performs allreduce on BF16 gradients. This PR fixes the issue by performing allreduce on the FP32 buffer. It also eliminates an assertion that prohibits BF16+PP+Z1, which is actually runnable. This shows loss curves of the following conditions: - BF16/Z0,Z1,Z2,Z3/NoPP - BF16/Z0,Z1/PP(2 stages) (all used 8GPUs, gradient accumulation step: 4)  --------- Co-authored-by: Logan Adams <[email protected]>
ShellyNR
pushed a commit
to ShellyNR/DeepSpeed
that referenced
this pull request
Mar 11, 2024
This PR fixes an issue with allreducing for ZeRO0 + BF16. (This replaces microsoft#5154) DeepSpeed uses `BF16_Optimizer` when ZeRO0 and BF16 are enabled. The optimizer accumulates gradients on FP32 buffer soon after a backward pass completes. However, DeepSpeed engine performs allreduce on BF16 gradients. This PR fixes the issue by performing allreduce on the FP32 buffer. It also eliminates an assertion that prohibits BF16+PP+Z1, which is actually runnable. This shows loss curves of the following conditions: - BF16/Z0,Z1,Z2,Z3/NoPP - BF16/Z0,Z1/PP(2 stages) (all used 8GPUs, gradient accumulation step: 4)  --------- Co-authored-by: Logan Adams <[email protected]>
rraminen
pushed a commit
to ROCm/DeepSpeed
that referenced
this pull request
May 9, 2024
This PR fixes an issue with allreducing for ZeRO0 + BF16. (This replaces microsoft#5154) DeepSpeed uses `BF16_Optimizer` when ZeRO0 and BF16 are enabled. The optimizer accumulates gradients on FP32 buffer soon after a backward pass completes. However, DeepSpeed engine performs allreduce on BF16 gradients. This PR fixes the issue by performing allreduce on the FP32 buffer. It also eliminates an assertion that prohibits BF16+PP+Z1, which is actually runnable. This shows loss curves of the following conditions: - BF16/Z0,Z1,Z2,Z3/NoPP - BF16/Z0,Z1/PP(2 stages) (all used 8GPUs, gradient accumulation step: 4)  --------- Co-authored-by: Logan Adams <[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 combination of BF16 and ZeRO0 (no ZeRO optimization) has some issues with handling gradients.
BF16_Optimizerseems to be designed to accumulate gradients in FP32, but this doesn't match other parts.BF16_Optimizer. However, the engine performs allreduce on BF16 gradients after that. As the result, gradients are not properly reduced.BF16_Optimizer'sbackward(). This clears BF16 gradients. So it doesn't work when the gradient accumulation step > 1.There are two possible approaches to resolve this issue:
This PR takes the first approach to resolve the issue. ZeRO 1/2/3 follows the first approach though the second one has the advantage in terms the precision for gradient accumulation.