[RAG] official facebook example code for RAG is not working anymore.

[giobin]

Environment info

• transformers version: 4.6.0.dev0
• Platform: Linux-3.10.0-1127.10.1.el7.x86_64-x86_64-with-debian-buster-sid
• Python version: 3.6.13
• PyTorch version (GPU?): 1.8.1+cu102 (True)
• Tensorflow version (GPU?): 2.4.1 (False)
• Using GPU in script?: no
• Using distributed or parallel set-up in script?: no

Who can help

rag: @patrickvonplaten, @lhoestq

Models:
RAG model

Information

Model I am using RAG

The problem arises when using the official example scripts from https://huggingface.co/facebook/rag-sequence-nq, which I copied here:

from transformers import RagTokenizer, RagRetriever, RagSequenceForGeneration 

tokenizer = RagTokenizer.from_pretrained("facebook/rag-sequence-nq") 
retriever = RagRetriever.from_pretrained("facebook/rag-sequence-nq", index_name="exact", use_dummy_dataset=True) 
model = RagSequenceForGeneration.from_pretrained("facebook/rag-sequence-nq", retriever=retriever) 

input_dict = tokenizer.prepare_seq2seq_batch("how many countries are in europe", return_tensors="pt") 

generated = model.generate(input_ids=input_dict["input_ids"]) 
print(tokenizer.batch_decode(generated, skip_special_tokens=True)[0]) 

The tasks I am working on is:
I am trying to ran the sample code from above. Note that this same error occurs also when finetuning rag with the official code in transformers/examples/research_projects/rag/finetune_rag.sh

To reproduce

Steps to reproduce the behavior:

1. run the sample code above

The error is:

Traceback (most recent call last):
  File "prova_rag.py", line 5, in <module>
    model = RagSequenceForGeneration.from_pretrained("facebook/rag-sequence-nq", retriever=retriever, _fast_init=False)
  File "/nlu/users/giovanni_bonetta/transformers/src/transformers/modeling_utils.py", line 1208, in from_pretrained
    model, state_dict, pretrained_model_name_or_path
  File "/nlu/users/giovanni_bonetta/transformers/src/transformers/modeling_utils.py", line 1278, in _load_state_dict_into_model
    model._init_weights(module)
  File "/nlu/users/giovanni_bonetta/miniconda2/envs/hf_venv/lib/python3.6/site-packages/torch/nn/modules/module.py", line 948, in __getattr__
    type(self).__name__, name))
AttributeError: 'RagSequenceForGeneration' object has no attribute '_init_weights'

Looking at last commits i suppose that the error was introduced in "Pytorch - Lazy initialization of models #11471" a couple of weeks ago, where the line model._init_weights(module) was introduced.

Expected behavior

It should initialize the model without errors.