Skip to content

full state checkpoint loading is broken in multi-node with FSDP #666

@RohanAwhad

Description

@RohanAwhad

I get this error:

[rank9]: Traceback (most recent call last):
[rank9]:   File "/workspace/home/lab/osilkin/training/src/instructlab/training/main_ds.py", line 821, in <module>
[rank9]:     main(args)
[rank9]:   File "/workspace/home/lab/osilkin/training/src/instructlab/training/main_ds.py", line 432, in main
[rank9]:     load_latest_full_state(args=args, accelerator=accelerator)
[rank9]:   File "/workspace/home/lab/osilkin/training/src/instructlab/training/utils.py", line 891, in load_latest_full_state
[rank9]:     accelerator.load_state(latest)
[rank9]:   File "/workspace/home/lab/osilkin/training_hub/.venv/lib/python3.12/site-packages/accelerate/accelerator.py", line 3685, in load_state
[rank9]:     load_fsdp_model(self.state.fsdp_plugin, self, model, input_dir, i)
[rank9]:   File "/workspace/home/lab/osilkin/training_hub/.venv/lib/python3.12/site-packages/accelerate/utils/fsdp_utils.py", line 197, in load_fsdp_model
[rank9]:     state_dict = torch.load(input_model_file, weights_only=True)
[rank9]:                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank9]:   File "/workspace/home/lab/osilkin/training_hub/.venv/lib/python3.12/site-packages/torch/serialization.py", line 1484, in load
[rank9]:     with _open_file_like(f, "rb") as opened_file:
[rank9]:          ^^^^^^^^^^^^^^^^^^^^^^^^
[rank9]:   File "/workspace/home/lab/osilkin/training_hub/.venv/lib/python3.12/site-packages/torch/serialization.py", line 759, in _open_file_like
[rank9]:     return _open_file(name_or_buffer, mode)
[rank9]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank9]:   File "/workspace/home/lab/osilkin/training_hub/.venv/lib/python3.12/site-packages/torch/serialization.py", line 740, in __init__
[rank9]:     super().__init__(open(name, mode))
[rank9]:                      ^^^^^^^^^^^^^^^^
[rank9]: FileNotFoundError: [Errno 2] No such file or directory: '/mnt/nvme0n1/output-checkpoints/sft/full_state/epoch_0/pytorch_model_fsdp.bin'

When using 2 nodes

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions