Problems with locally downloading an adapter model with save_pretrained

[ashleys0]

I'm trying to simply download model weights locally. model.save_pretrained works with regular huggingface transformers models, but doesn't seem to be the case with adapters models. When I tried the below code, the model loaded straight from huggingface and the model loaded from my locally downloaded version don't produce the same embeddings. What's the proper way to download locally?

model_name = "allenai/specter2"
model = AutoAdapterModel.from_pretrained(model_name)
model.load_adapter(model_name, source="hf", load_as="specter2_proximity", set_active=True)
config = AutoConfig.from_pretrained(model_name)
config.save_pretrained(local_dir_path)
model.save_pretrained(local_dir_path)

config = AutoConfig.from_pretrained(local_dir_path)
model = AutoModel.from_pretrained(local_dir_path, config=config)

Thank you so much!!

[hSterz]

Hey @ashleys0 , I tried to reproduce the problem. The following code works for me:

model = AutoAdapterModel.from_pretrained("allenai/specter2_base")
model.load_adapter("allenai/specter2", source="hf", load_as="specter2_proximity", set_active=True)
model.save_pretrained(local_dir_path)
model.set_active_adapters("specter2_proximity")
model.eval()

model_local = AutoAdapterModel.from_pretrained(local_dir_path)
model_local.set_active_adapters("specter2_proximity")
model_local.eval()

tokenizer = AutoTokenizer.from_pretrained("allenai/specter2_base")

text = "Some text"
input_data = tokenizer(text, return_tensors="pt")

out1 = model(**input_data)
out2 = model_local(**input_data)

assert torch.equal(out1.last_hidden_state, out2.last_hidden_state)

Could you go into more detail on what the problem is?