.add_embeddings prevent model from training adapters

[eugene-yang]

Environment info

• adapter-transformers version: v3.0.1+ (commit 11bd9d2)
• Platform: Arch Linux
• Python version:
• PyTorch version (GPU?):
• Tensorflow version (GPU?):
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Information

Model I am using (Bert, XLNet ...): XLMR

Language I am using the model on (English, Chinese ...):

Adapter setup I am using (if any):

The problem arises when using:

• the official example scripts: (give details below)
• my own modified scripts: (give details below)

The tasks I am working on is:

• an official GLUE/SQUaD task: (give the name)
• my own task or dataset: (give details below)

To reproduce

from transformers import AdapterSetup, AutoAdapterModel, AutoTokenizer, AdapterConfig
model = AutoAdapterModel.from_pretrained('xlm-roberta-base')
tokenizer = AutoTokenizer.from_pretrained('xlm-roberta-base')
tokenizer_new = AutoTokenizer.from_pretrained('xlm-roberta-base')
model.add_embeddings('new', tokenizer_new, reference_tokenizer=tokenizer, reference_embedding='default')
model.add_adapter('test')
model.train_adapter(['test'])

This gives me the following error message.

File /expscratch/eyang/workspace/adapter/adapter-transformers/src/transformers/adapters/model_mixin.py:949, in ModelWithHeadsAdaptersMixin.train_adapter(self, adapter_setup, train_embeddings)
    947     super().train_adapter(adapter_setup, train_embeddings)
    948 else:
--> 949     self.base_model.train_adapter(adapter_setup, train_embeddings)

File /expscratch/eyang/workspace/adapter/adapter-transformers/src/transformers/adapters/model_mixin.py:287, in ModelAdaptersMixin.train_adapter(self, adapter_setup, train_embeddings)
    285 """Sets the model into mode for training the given adapters."""
    286 self.train()
--> 287 self.freeze_model(True)
    288 adapter_setup = parse_composition(adapter_setup)
    289 self.apply_to_adapter_layers(lambda i, layer: layer.enable_adapters(adapter_setup, True, False))

File /expscratch/eyang/workspace/adapter/adapter-transformers/src/transformers/adapters/model_mixin.py:726, in ModelAdaptersMixin.freeze_model(self, freeze)
    724 # first freeze/ unfreeze all model weights
    725 for param in self.base_model.parameters():
--> 726     param.requires_grad = not freeze
    727 self.model_frozen = freeze

RuntimeError: you can only change requires_grad flags of leaf variables. If you want to use a computed variable in a subgraph that doesn't require differentiation use var_no_grad = var.detach().

I used the following code to see which specific parameters is causing the issue

for name, p in model.base_model.named_parameters():
    try:
        p.requires_grad = False
    except Exception as e:
        print(name)
        print(e)

and got

embeddings.word_embeddings.weight
you can only change requires_grad flags of leaf variables. If you want to use a computed variable in a subgraph that doesn't require differentiation use var_no_grad = var.detach().

Interestingly, the following code is fine.

from transformers import AdapterSetup, AutoAdapterModel, AutoTokenizer, AdapterConfig
model = AutoAdapterModel.from_pretrained('xlm-roberta-base')
tokenizer = AutoTokenizer.from_pretrained('xlm-roberta-base')
model.add_embeddings('new', tokenizer)
model.add_adapter('test')
model.train_adapter(['test'])

So the referencing seems to be the part that breaks the computational graph. (which I'm not sure why as the parameters were cloned when adding a new embeddings...)

The following code is also fine

from transformers import AdapterSetup, AutoAdapterModel, AutoTokenizer, AdapterConfig
model = AutoAdapterModel.from_pretrained('xlm-roberta-base')
model.add_adapter('test')
model.train_adapter(['test'])
tokenizer = AutoTokenizer.from_pretrained('xlm-roberta-base')
tokenizer_new = AutoTokenizer.from_pretrained('xlm-roberta-base')
model.add_embeddings('new', tokenizer_new, reference_tokenizer=tokenizer, reference_embedding='default')

Expected behavior

Should be able to train stuff...