Nicolas Patry

> for my case， open(filename); f.read() is 2 GB/s, for memcpy(mmap(filename)) is 1.3 GB/s. which is mach more slower than read file. Something is wrong in your system, what are...

> I'm not 100% convinced yet that _any_ imports are safe. For example, [`numpy` is not, in general](https://numpy.org/doc/stable/reference/generated/numpy.distutils.exec_command.html). I think it's even worse than that: ```python import numpy as np...

threading_threshold is what you had before to get `num_threads=1 vs num_thread=all`. `(total_work - threading_threshold + 1) / threading_threshold` Is simply `ceil(total_work/threading_threshold)` (To get a heuristic on how many threads this...

The design in https://github.com/huggingface/transformers/issues/3609#issuecomment-854005760 seems very interesting ! Main comments: - I would be curious to see (and probably it would need to become a test) to prove that doing...

> After reading this thread and skimming through #14716, I must confess I still a little unsure how the scores for multi-masked prompts are computed. Based on my understanding, for...

> Is there any specific reason for this? Backward compatibility, the first pipeline wasn't built with that option in mind making it harder to support multi mask seamlessly like you...

I guess it's possible. I'm not very active on github as you can see. I don't see why it would be a lot of work. If you do it, I...

This projet was created 9 years ago, you probably need to rewrite it to be able to use it with modern Django. Sorry about this.

Feel free to use, for whatever purpose. By the way, I did not look into this for a long time, so there might be a lot of errors in there.

Thanks a lot for sharing the info ! I'll when I have time to set it up !