convert : use all parts in safetensors index

[CISC]

Use all file parts in safetensors index, even if they are not named model*.safetensors. Still requires some of the files to be named that for detection to work (due to consolidated.safetensors).

Fixes #16361

[ngxson]

Still requires some of the files to be named that for detection to work (due to consolidated.safetensors).

Not quite sure if I understand this phrase, but for Mistral, we either use model-*.safetensors OR consolidated.safetensors; If all 2 types are being used, it will result in duplicated tensors

[CISC]

Still requires some of the files to be named that for detection to work (due to consolidated.safetensors).

Not quite sure if I understand this phrase, but for Mistral, we either use model-*.safetensors OR consolidated.safetensors; If all 2 types are being used, it will result in duplicated tensors

Yes, exactly, that's why this is done:

llama.cpp/convert_hf_to_gguf.py

Line 191 in 7fdac71

prefix = "model" if not self.is_mistral_format else "consolidated"

..meaning, if no files are named model*.safetensors, it won't be detected properly.

Edit: Just for clarity; I think it used to scan for *.safetensors, which resulted in duplicated tensors when consolidated.safetensors came along.