Revert keys_to_ignore for M2M100

[younesbelkada]

What does this PR do?

This PR reverts a change that has been made on M2M100Model , to reproduce you can run:

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

src_lang = "eng_Latn"
tgt_lang = "spa_Latn"

model_id = "facebook/nllb-200-3.3B"

tokenizer = AutoTokenizer.from_pretrained(model_id, src_lang=src_lang)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

on main, users get:

│ /home/younes_huggingface_co/debug_issues/code/transformers/src/transformers/modeling_utils.py:24 │
│ 59 in _load_pretrained_model                                                                     │
│                                                                                                  │
│   2456 │   │   │   for key in missing_keys:                                                      │
│   2457 │   │   │   │   if key.startswith(prefix):                                                │
│   2458 │   │   │   │   │   key = ".".join(key.split(".")[1:])                                    │
│ ❱ 2459 │   │   │   │   param = model_state_dict[key]                                             │
│   2460 │   │   │   │   if param.device == torch.device("meta"):                                  │
│   2461 │   │   │   │   │   if not load_in_8bit:                                                  │
│   2462 │   │   │   │   │   │   set_module_tensor_to_device(model, key, "cpu", torch.empty(*para  │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
KeyError: 'encoder.embed_positions.weights'

cc @Narsil

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[Narsil]

LGTM

[sgugger]

Thanks for fixing!