Maxime

Well it only occurs with AMD GPUs so it is related to their drivers, the stack above shows that it crashes within the driver code

This is the intended behaviour, we map virtual keys, not mapped keys.

In terms of performance it appears to be faster than WebSockets, not as fast as Telepathy but this is expected considering we are very conservative with the computing power we...

This looks handy! Having the ability to add custom rules for codegen would help us as well

We know, no need to keep spamming.

Highly interested in this as well, for domain specific pre-training before instruct/chat finetuning it would be very useful.

> @maximegmd This is awesome! Can you post some loss curves for the finetune you ran? I will complete a run during the weekend, losses looked fine but the Llama3...

device is not None when this function is called so it just passes 'mps' to torch.device() which is the expected pytorch name. But you are correct that there is room...

If I recall correctly it was around 20s/it but I suspect I was swapping a bit so I can probably improve the speed. The main issue is bitsandbytes not supporting...

I made a PR to fix the incorrect use of devices https://github.com/explosion/spacy-huggingface-pipelines/pull/23 Then setting the device is done with spacy: ```python import spacy spacy.require_gpu() ``` As for dtype, I used...