Error when specifying adapter_id with invalid base_model_name_or_path about lorax

[thincal]

input_dict = {
"inputs": test_data[idx],
"parameters": {
"adapter_id": "merror/llama_13b_lora_beauty",
"max_new_tokens": 256,
"top_p": 0.7
}
}

if the adapter_id is from local filesystem, "adapter_source": "local" is required also.

[sleepwalker2017]

input_dict = {
"inputs": test_data[idx],
"parameters": {
"adapter_id": "merror/llama_13b_lora_beauty",
"max_new_tokens": 256,
"top_p": 0.7
}
}

if the adapter_id is from local filesystem, "adapter_source": "local" is required also.

hi, the adapter is from hub.
I wonder why it looks for this repo: llama-13b-hf.

[tgaddair]

Hey @sleepwalker2017, I think I see the issue here. Looking at the adapter_confg.json for this adapter, it states the base_mdel_name_or_path as decapoda-research/llama-13b-hf. Because your base model is /data/vicuna-13b/vicuna-13b-v1.5 (different from decapoda-research/llama-13b-hf), LoRAX will attempt to do an additional check here to see if the architectures of the adapter and base model are compatible.

In this case, the architecture compatibility check fails because decapoda-research/llama-13b-hf no longer exists (or was made private).

I think the architecture check should not cause a hard failure if it can't be performed. I'll put together a PR to make this check non-fatal in this case.