Arthur comments

Results 795 comments of


                                            Arthur

gpt2 results with past_key_values not the same as when computed from scratch

Hey! Yes as mentioned before, the positional IDS in GPT2 are not created on the fly contrary to other of our models. A fix is in the makinf, see #21853,...

Add DEBERTA-base model for usage in EncoderDecoderModel.

Yes! Feel free to ping me for an early review if you have any doubts

Blip2ForConditionalGeneration.from_pretrained is limited by 100% CPU usability (on one single core)

It's unrelated to this issue, but setting `device_map = "auto"` should suffice

Felix

I think you can take this on!

Understanding why TorchInductor cannot speed-up huggingface transformer inference

Yes! Static KV cache is not supported but coming soon!

ffmpeg_microphone does not use current input device on Mac/Darwin

We should be able to close this as #31846 was merged

With deepspeed zero3 enabled, loading from_pretrained() and resize_token_embeddings() do not work correctly

@haixpham can you rather provide a reproducer?

With deepspeed zero3 enabled, loading from_pretrained() and resize_token_embeddings() do not work correctly

cc @muellerzr if you can have a look!

Possible bug in case of prepending chars in a pretokenizer

I am not getting this on 0.19:

Possible bug in case of prepending chars in a pretokenizer

This is expected, the `pre_tokenizer` does not have access to the information about the special tokens, so it will always prepend regardless of whether the first token is a special...