nvidia/llama-3.1-nemoguard-8b-content-safety · Unable to load model with PEFT due to nested file structure

Hey folks, thanks for releasing this model! Very excited to test it out.

Currently, I am not able to load the adapter model using PEFT. For example, running the following:

model = AutoPeftModelForCausalLM.from_pretrained(
   'nvidia/llama-3.1-nemoguard-8b-content-safety',
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
)

will return the error ValueError: Can't find 'adapter_config.json' at 'nvidia/llama-3.1-nemoguard-8b-content-safety'.

The huggingface "Use this model" feature also detects that there is an issue:

I suspect that this is because the adapter files are stored within a nested llama-3.1-nemoguard-8b-content-safety-lora-adapter directory, as opposed to the base nvidia/llama-3.1-nemoguard-8b-content-safety directory.