Fix Muon optimizer checkpoint resume with bf16 mode#7748
Open
yurekami wants to merge 1 commit intodeepspeedai:masterfrom
Open
Fix Muon optimizer checkpoint resume with bf16 mode#7748yurekami wants to merge 1 commit intodeepspeedai:masterfrom
yurekami wants to merge 1 commit intodeepspeedai:masterfrom
Conversation
When resuming training from a checkpoint with bf16 enabled, the Muon optimizer's momentum_buffer was loaded as fp32 (from the checkpoint) while gradients were bf16, causing a dtype mismatch error in the lerp_() operation. This fix adds a load_state_dict override to all Muon optimizer classes (Muon, SingleDeviceMuon, MuonWithAuxAdam, SingleDeviceMuonWithAuxAdam) that casts the momentum_buffer (and exp_avg/exp_avg_sq for hybrid classes) to match the parameter dtype after loading the checkpoint. Fixes: deepspeedai#7746 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> Signed-off-by: yurekami <yurekami@users.noreply.github.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
load_state_dictoverride to castmomentum_bufferto match parameter dtype after loadingRoot Cause
When resuming training from a checkpoint with
bf16enabled:momentum_bufferis saved asfp32in the checkpointload_state_dict(),momentum_bufferremainsfp32bf16momentum.lerp_(grad, 1 - beta)crashes due to dtype mismatchFix
Added
load_state_dictoverride to all Muon optimizer classes that casts optimizer state buffers to match the parameter dtype after loading:Muonmomentum_bufferSingleDeviceMuonmomentum_bufferMuonWithAuxAdammomentum_buffer,exp_avg,exp_avg_sqSingleDeviceMuonWithAuxAdammomentum_buffer,exp_avg,exp_avg_sqTest Plan
TestMuonBF16CheckpointResumetest class that:Fixes: #7746
🤖 Generated with Claude Code