Fix: Falcon tie_word_embeddings in GGUF

[MekkCyber]

What does this PR do?

In the modeling_gguf_pytorch_utils.py file, the value of tie_word_embeddings is determined by the presence (or absence) of the tensor output.weight in the GGUF file. While this approach is generally a good indicator, the Falcon architecture is an exception to the rule as you can see here : https://huggingface.co/tiiuae/falcon-7b/tree/main?show_file_info=model-00002-of-00002.safetensors (hf format) and here https://huggingface.co/tensorblock/falcon-7b-GGUF/tree/main?show_file_info=falcon-7b-Q2_K.gguf (gguf format)

In Falcon, word_embeddings are tied to the lm_head weights. Despite this, output.weight is still present in the GGUF file, and lm_head is included in the Hugging Face model format. To handle this edge case, I added an exception array for such architectures.

This issue was causing a subtle error related to parameters not being on the same device, which was only discoverable in multi-GPU settings.

Who can review ?

@SunMarc @Isotr0py

[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.

[Isotr0py]

But I remember tie_word_embeddings is not used by modeling_falcon.py?

BTW, if the issue about tie_word_embedding is the case, can we also update the falcon conversion test to cover this edge case?

[MekkCyber]

tie_word_embeddings is not used directly in the modeling file of the model, instead it's used in modeling_utils.py : https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_utils.py#L1858
It was the failure of the Falcon conversion test that made me realize the error, and now it passes successfully.

[Isotr0py]

LGTM! Thanks for fixing!

[SunMarc]

LGTM ! Let's do that for falcon but if more issues starts to appear, let's modify the tests slightly to not take into account the device of the tensors