Allow override when adding value to ggufwriter

[huydt84]

Discussion: #14164

This PR use a simple way to allow overriding value when add key-value to GGUF, which has been mentioned here: #14164 (comment)

This can cause drastic changes in many unexpected aspects, so please review carefully

[Copilot]

Pull Request Overview

This PR updates the behavior of the add_key_value method in gguf_writer.py to allow overriding an existing key instead of raising an exception.

• The exception for a duplicate key is replaced with a logger warning.
• The existing key in the first dictionary of kv_data is unconditionally overwritten.

Comments suppressed due to low confidence (1)

gguf-py/gguf/gguf_writer.py:274

• The change always updates the key in self.kv_data[0], which may not reflect the intended behavior if a duplicate key exists in a different dictionary within kv_data. Consider verifying that updating only at index 0 is correct, or update the key in its original location.

logger.warning(f'Duplicated key name {key!r}, overwriting it with new value {val!r} of type {vtype.name}')

[CISC]

I think this should be ok, it's only a potential problem if you did something wrong when adding a new model, but will usually be caught later on if you failed to notice the warning.

Either way, adding some more people in to review for additional opinions. :)

[ngxson]

I would admit that I myself sometimes struggle with the fact that we don't allow overwriting KV metadata.

The cleaner solution is to refactor the conversion code inside convert_hf_to_gguf.py to avoid adding the same key twice, but it proves to be quite complicated in some cases. For example, overwriting chat template is controlled by set_vocab which is quite complicated to modify.

So yes I think this can be a valid solution for now.

[compilade]

The error was originally added in #7827, mainly because duplicate keys often were mostly accidental.

It's also because before that PR, it wasn't really possible to override an existing key-value pair, and so it really was added twice to the GGUF. (that's no longer the case, and overriding the value of a key should do what you expect)

I guess it can make sense to allow it (while still warning).