fix: add pull to solve the problem that fsdp2 traninng in shared will timeout #68
Open
gygdh-001 wants to merge 2 commits into
Open
fix: add pull to solve the problem that fsdp2 traninng in shared will timeout #68gygdh-001 wants to merge 2 commits into
gygdh-001 wants to merge 2 commits into
Conversation
gygdh-001
changed the title
Previously _resume_checkpoint() only restored optimizer and scheduler states but never loaded model weights. This caused resumed training to use the original pretrained weights instead of the checkpointed trained weights, making resume effectively a from-scratch restart. Fix: add _load_model_weights_from_checkpoint() to load per-module diffusion_pytorch_model.bin files for all trainable modules (video_dit, video_dit_2, audio_dit, dual_tower_bridge) before restoring optimizer and scheduler states. Verified by: 50-parameter anchor comparison (max_diff=0), loss continuity (step-6 loss=0.3047 → step-7 loss=0.3047), and optimizer state key audit (3735 keys match). Signed-off-by: gygdh-001 <leiliandong@foxmail.com>
Author
|
@tianyilt Please review the current PR. If you have any questions, feel free to ask. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
During multi-node multi-GPU FSDP training, checkpoint saving and resume frequently hit timeout errors.
The root cause is that
accelerate_trainer.pyusesFullStateDictConfig/FullOptimStateDictConfigfor FSDP,which gathers the full state dict from all ranks onto rank 0 during every
save_state/load_statecall.Transferring the full parameter set across nodes over the network causes timeouts,
and rank 0 is also at risk of OOM.
This PR applies the following changes:
Replaces
FullStateDictConfig/FullOptimStateDictConfigwithShardedStateDictConfig/ShardedOptimStateDictConfigfor FSDP,so each rank only holds its own shard without cross-node full-state gather.
Under FSDP,
save_checkpointnow saves per-rank files:optimizer_{rank}.binandscheduler_{rank}.bin,bypassing
accelerator.save_state()which would trigger full-state gather.Under FSDP,
_resume_checkpointnow loads per-rank optimizer and schedulerstate dicts independently, bypassing
accelerator.load_state()and its full-state gather path.Validation:
Verified on a multi-node (2 nodes × 8 Ascend 910B) FSDP training setup: