Nicolas Patry

@Wumpf Maybe ? After having matured a bit this and implemented the `Matrix` in https://github.com/huggingface/candle/ I feel like this PR should be trimmed down to only get the `MPSMatrix` support...

@altryne @ArthurZucker . While deep diving into whisper, I've notived `openai/whisper` uses timestamp ALL the time, while `transformers` doesn't (you have to ask for timestamps for us to use them)....

Hmm the `check_file_size` is pretty rough sanitation, the file might actually be OK but it's hard to tell without looking at the file. You can try deactivating the check ?...

Can you point to the actual model on the hub you're using too ? Because we don't have any issue with "official" checkpoints

> The funny thing is the code works well with 7B and 65B models but fails for 13B and 30B I converted 30B like 3 times today for quantization purposes...

> Throws a InvalidHeaderDeserialization. Loading the same files using LlamaForCausalLM works fine in a notebook. You're trying to load `pickle` files (which would work for `LLamaForCausalLM`) it seems

> - /usr/src/llama-30b-supercot/pytorch_model-00001-of-00243.bin: 537 > - /usr/src/llama-30b-supercot/model-00001-of-00243.safetensors: 48 This is probably an empty file or containing a super small tensor...

It's empty actually...

No no, I inspected the first file, it's purely empty. My paranoid flag almost expected a payload here, but no it's just empty.

> Awesome! Documenting that this PR should fix #1689 - right? Indeed it would ! Not super generally quite yet (we need to transfer part of the token counting logic...