Pedro Cuenca

Hi, thanks for reporting this! It's strange, we do serve CORS headers and the site works fine for me in Safari and Brave. I just tested in Firefox and got...

See also this effort: https://huggingface.slack.com/archives/C025LJDP962/p1627724095070700

I don't think this needs to go into `diffusers@main` yet, as it will require users to install jax to use the `pndm` scheduler. It's cool to have the tests, but...

Hi @kabachuha! Work towards that goal is being coordinated in issue #672, where you have participated :) It looks like @rashmimarganiatgithub is taking a go at it!

> generating 1 image took 27 mins on dalle-playground (using 117% CPU), whereas this pytorch model runs in 2.7 mins (using 145% CPU)! GPU looks less-than-half utilized. haven't checked whether...

@Birch-san These are my changes so far: https://github.com/kuprel/min-dalle/compare/main...pcuenca:min-dalle:mps-device I tried to use workarounds for unsupported ops, except for `multinomial`. You need to use `PYTORCH_ENABLE_MPS_FALLBACK=1` for the backend to automatically fall...

> it's worth knowing that the MPS backend _does_ have some silent errors where it will produce incorrect output. here's the _really_ wacky one I found: [pytorch/pytorch#79383](https://github.com/pytorch/pytorch/issues/79383) That's very interesting....

The neural engine is much faster than the GPU, so it makes sense to apply those optimizations. Not all operations are supported, however, and it's hard to know whether the...

Yes, except when using JAX, where a multiple of 8 is often required for parallelization and batch sizes are therefore fixed, usually. But this could be handled by the user...

/cc @hollance too. I can do some testing here.