from_pretrained orchestration + distributed save/load#45409
Open
3outeille wants to merge 2 commits intomoe-sequence-parallelfrom
Open
from_pretrained orchestration + distributed save/load#454093outeille wants to merge 2 commits intomoe-sequence-parallelfrom
3outeille wants to merge 2 commits intomoe-sequence-parallelfrom
Conversation
- Add gather_full_state_dict() for DTensor→full tensor saving - Add convert_strided_to_shard() / restore_strided_from_shard() for DCP - Add _redistribute_dtensor() helper - Full distributed_config integration in from_pretrained/save_pretrained - Rename apply_fsdp2 → apply_fully_shard_data_parallel - save_optimizer() / load_optimizer() in distributed/utils - Trainer integration with distributed_config - Updated FSDP and TP tests for new orchestration API - DTensor shard-on-read test updates
3outeille
commented
Apr 14, 2026
| "ALL_PARALLEL_STYLES", | ||
| "translate_to_torch_parallel_style", | ||
| ] | ||
| _import_structure["tensor_parallel"] = [] |
Member
Author
There was a problem hiding this comment.
why is it empty ? do we still need it ?
| def get_correct_experts_implementation(self, requested_experts: str | None) -> str: | ||
| applicable_experts = "grouped_mm" if requested_experts is None else requested_experts | ||
| if applicable_experts not in ["eager", "grouped_mm", "batched_mm", "deepgemm"]: | ||
| if applicable_experts not in ["eager", "grouped_mm", "batched_mm"]: |
Member
Author
There was a problem hiding this comment.
why was deepgemm removed ?
| # so that custom models calling .item() during __init__ (e.g. drop-path | ||
| # schedules) don't crash on meta tensors. | ||
| init_contexts.extend([torch.device("meta"), init.meta_device_safe_creation_ops()]) | ||
| init_contexts.append(torch.device("meta")) |
| if device_mesh is not None: | ||
| shard_and_distribute_module( | ||
| self, value, param, key, None, False, device_mesh.get_local_rank(), device_mesh | ||
| from torch.distributed.tensor import DTensor |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
3outeille
force-pushed
the
moe-sequence-parallel
branch
from
e04c7d9 to
24ca327
Compare
# Conflicts: # src/transformers/distributed/utils.py
3outeille
force-pushed
the
orchestration-save-load
branch
from
815b5b2 to
7361deb
Compare
Contributor
|
View the CircleCI Test Summary for this PR: https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=45409&sha=7361de |
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
2 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
distributed_configintegration infrom_pretrained()— mesh creation, apply TP + FSDP, attachmodel.device_meshgather_full_state_dict()for streaming DTensor→full tensor saving (rank 0 only)convert_strided_to_shard()/restore_strided_from_shard()for DCP compatibility with_StridedShardsave_optimizer()/load_optimizer()indistributed/utils.pyapply_fsdp2→apply_fully_shard_data_paralleldistributed_configPart of the distributed training API chain: #44989
Chain:
main ← #44989 ← #44083 ← #44974 ← #45028 ← #45408 ← this PRReview question
Does
from_pretrainedwire things up in the right order? Is save/load round-trip correct?Test plan
from_pretrainedwith distributed_configgather_full_state_dict()roundtrip verificationsave_optimizer()/load_optimizer()roundtrip