Fix to tuple conversion with config

[qubvel]

What does this PR do?

setting return_dict=False with config fails for models with sub-models wrapped with can_return_tuple or check_model_inputs

import torch
from transformers import LlamaConfig, LlamaForCausalLM

config = LlamaConfig(vocab_size=256, hidden_size=128, num_hidden_layers=2, num_attention_heads=4, intermediate_size=256)
model = LlamaForCausalLM(config)

# default: ModelOutput
input_ids = torch.tensor([[0, 1, 2, 3]])
with torch.no_grad():
    output = model(input_ids)

print(output)


# passing return_dict=False as a kwarg 
input_ids = torch.tensor([[0, 1, 2, 3]])
with torch.no_grad():
    output = model(input_ids, return_dict=False)

print(output)


# ERROR: setting return_dict=False in the config
model.config.return_dict = False
with torch.no_grad():
    output = model(input_ids)

print(output)

# Traceback (most recent call last):
#   File "/home/ubuntu/projects/transformers/test_llama_small.py", line 17, in <module>
#     output = model(input_ids)
#   File "/home/ubuntu/projects/transformers/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
#     return self._call_impl(*args, **kwargs)
#   File "/home/ubuntu/projects/transformers/.venv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
#     return forward_call(*args, **kwargs)
#   File "/home/ubuntu/projects/transformers/src/transformers/utils/generic.py", line 962, in wrapper
#     output = func(self, *args, **kwargs)
#   File "/home/ubuntu/projects/transformers/src/transformers/models/llama/modeling_llama.py", line 506, in forward
#     hidden_states = outputs.last_hidden_state
# AttributeError: 'tuple' object has no attribute 'last_hidden_state'

[qubvel]

This way, it's going to work in case **kwargs: [TransformersKwargs] are properly propagated from the top module up to each wrapped module.

Just for the context, previously, we were recursively setting the module attribute _is_top_module to avoid passing **kwargs everywhere.

[qubvel]

no longer needed

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.