eval_multigpu.py issues

[y-he2]

If directly run eval_multigpu.py without any training, after the script downloaded and attempted to loaded the shards raised:
safetensors_rust.SafetensorError: Error while deserializing header: HeaderTooLarge

Checking the model sizes under for example
all_models/adapters/gate/summary/
gives only 1KB, which suspect the model is not ready to run before a training.

But it could also be my PyTorch version, is there any version requirement for PyTorch used?

Let me know if you need further details.