bugs in accelerator.unwrap_model

[ThisisBillhe]

System Info

- `Accelerate` version: 1.7.0
- Platform: Linux-5.10.134-008.12.kangaroo.al8.x86_64-x86_64-with-glibc2.35
- `accelerate` bash location: /mnt/workspace/xxx/miniconda3/envs/blip3o/bin/accelerate
- Python version: 3.11.11
- Numpy version: 1.26.4
- PyTorch version: 2.5.1+cu124
- PyTorch accelerator: CUDA
- System RAM: 1600.00 GB
- GPU type: NVIDIA A800-SXM4-80GB
- `Accelerate` default config:
        Not found

Information

• The official example scripts
• My own modified scripts

Tasks

• One of the scripts in the examples/ folder of Accelerate or an officially supported no_trainer script in the examples folder of the transformers repo (such as run_no_trainer_glue.py)
• My own task or dataset (give details below)

Reproduction

I think there is a bug in accelerate/utils/other.py in the extract_model_from_parallel() function. When the model has some modules being compiled while the model itself has not, accessing model._orig_mod will lead to an error because there is no "model._orig_mod" attribute.

Expected behavior

When model is not compiled but model.module1 is compiled, accelerator.unwrap_model(model) successfully.

[SunMarc]

Thanks for the report @ThisisBillhe . Can you share a minimal reproducer ?

[gspeter-max]

@ThisisBillhe yes sir please provide reproducer , i spend my 30 min to find out the actual thing of this issue but i not found anything usefull

and in india time is 6:00 AM so good morning @SunMarc @ThisisBillhe