Conversation
Move FSDP2 wrapping and plan verification to distributed/fsdp.py, keep integrations/fsdp.py as a backward-compatible re-export, and update core call sites to import from transformers.distributed.fsdp.
…uggingface/transformers into split/a-pr-3-dual-path-loading
naming Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
…huggingface/transformers into split/a-pr-4-fsdp-orchestration
Member
Author
|
flagging #47253 for here |
Contributor
|
[For maintainers] Suggested jobs to run (before merge) run-slow: cohere2_moe, deepseek_v4, glm_moe_dsa, gpt_oss |
Contributor
CI recapDashboard: View test results in Grafana |
This was referenced Jul 16, 2026
Closed
Member
Author
ArthurZucker
approved these changes
Jul 17, 2026
ArthurZucker
left a comment
ArthurZucker
left a comment
Collaborator
There was a problem hiding this comment.
overall LGTM, would like to remove onliners, many are introduced where. we might not need them!
Comment on lines
+3452
to
+3458
| if self._tp_size is not None: | ||
| state_dict, needs_dist_barrier = self._gather_tp_state_dict_for_save( | ||
| state_dict, is_checkpoint_writer=save_on_this_rank | ||
| ) | ||
| elif is_fsdp_managed_module(model_to_save): | ||
| state_dict, needs_dist_barrier = self._gather_fsdp_state_dict_for_save(model_to_save) | ||
|
|
Collaborator
There was a problem hiding this comment.
should be a single distributed function with the if else modeling does not care
| "metadata": {"total_parameters": self.num_parameters(), **state_dict_split.metadata}, | ||
| "weight_map": weight_map, | ||
| } | ||
| # TODO: it would be very nice to do the writing concurrently, but safetensors never releases the GIL, |
| "torch", | ||
| docker_image=[{"image": "huggingface/transformers-torch-light"}], | ||
| marker="not generate", | ||
| marker="not (generate or is_training_test or is_tensor_parallel_test or is_fsdp_test)", |
|
|
||
| distributed_config = DistributedConfig(enable_expert_parallel=True) | ||
| distributed_config = DistributedConfig( | ||
| tp_size=int(os.environ["WORLD_SIZE"]), |
Collaborator
There was a problem hiding this comment.
do you actually need this?
| raise ValueError( | ||
| "FSDP+TP is not supported yet. " | ||
| "Use DistributedConfig(fsdp_size=N) or DistributedConfig(tp_size=N), not both. " | ||
| "2D support will come soon." |
Collaborator
There was a problem hiding this comment.
link the PR, maybe next release?!
Comment on lines
+81
to
+88
| if not is_torch_available(): | ||
| raise RuntimeError("PyTorch is required to use DistributedConfig.") | ||
|
|
||
| if not torch.distributed.is_available() or not torch.distributed.is_initialized(): | ||
| raise RuntimeError( | ||
| "torch.distributed must be initialized before using DistributedConfig with tp_size > 1 or " | ||
| "fsdp_size > 1. Call dist.init_process_group(...) first, or launch with torchrun." | ||
| ) |
Collaborator
There was a problem hiding this comment.
this file can't be imported already no?
| elif distributed_config.fsdp_size > 1: | ||
| device_map, device_mesh = initialize_fully_sharded_data_parallelism(distributed_config) | ||
|
|
||
| distributed_config.validate() |
Collaborator
There was a problem hiding this comment.
weird that we validate last
Comment on lines
+255
to
+274
| def _gather_tp_state_dict_for_save( | ||
| self, | ||
| local_state_dict: dict, | ||
| *, | ||
| is_checkpoint_writer: bool = True, | ||
| ) -> tuple[dict, bool]: | ||
| """All-gather TP-sharded weights for checkpoint writing.""" | ||
| full_state_dict = gather_state_dict_for_save(local_state_dict, self._tp_plan, self._device_mesh, self._tp_size) | ||
| if not is_checkpoint_writer: | ||
| full_state_dict = {} | ||
| return full_state_dict, True | ||
|
|
||
| def _gather_fsdp_state_dict_for_save(self, model_to_save) -> tuple[dict, bool]: | ||
| """Gather FSDP-sharded weights to full CPU tensors on rank 0.""" | ||
| if not _is_torch_distributed_initialized(): | ||
| raise ValueError( | ||
| "Saving an FSDP-wrapped model requires torch.distributed to be initialized. " | ||
| "Call save_pretrained from every rank after init_process_group." | ||
| ) | ||
| return gather_full_state_dict(model_to_save), True |
Collaborator
There was a problem hiding this comment.
loads of one liners, we usually avoid them when possible
Comment on lines
+3642
to
+3647
| # `gather_full_state_dict` concentrates the full state on rank 0 only; | ||
| # other ranks then loop over an empty shard list and would race ahead | ||
| # of rank 0's safetensors writes. Barrier so any subsequent | ||
| # `from_pretrained` on this path sees the consolidated files. | ||
| if needs_dist_barrier: | ||
| _distributed_barrier() |
Collaborator
There was a problem hiding this comment.
not sure this should be here
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
4 participants
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
base_fsdp_plan. Will do another PR to edit every other models laterfrom_pretrainedshard-on-Read+ saving like TP (DCP optional)DistributedConfigeverywhere (no more tp_plan=auto)Stack