Dilshod Tadjibaev

@dcvz can this issue be closed as the main PR is merged?

Thank you for running numbers. It seems it's slower. Perhaps, we should use more specialized op? For example, torch uses `torch.addmm(self.bias, x, self.weight.t())` instead of matmul (https://github.com/pytorch/pytorch/blob/1474eb5f293215c995b7c224ef924c305b166be1/torch/nn/modules/linear.py#L116)

Thank you, @louisfd for your research. This helped me go another route and leave nn modules as they're. In my PR (https://github.com/tracel-ai/burn/pull/1085), I am using custom deserializer with an adapter,...

I am hesitant to accept this PR because it contains so much generated code and I am not sure who would maintain all this. I am not even sure there...

Closing for now. The comments/feedback can be address in another PR.

https://github.com/tracel-ai/burn/pull/1081 has been merged and I assume this will fix. Closing for now. Let us know if this problem still occurs.

There is a tweet saying "float16 in webGPU finally works now" https://twitter.com/nisten/status/1698796718840598850 Worth looking into this and see if we need to update anything.

https://github.com/gfx-rs/wgpu/issues/4384

There is one issue relating to Atomic locking that has not been resolved yet but this was mainly with Ndarray backend. Please see: https://github.com/tracel-ai/burn/issues/302. It might be possible with Candle...

But generally we do support no_std and our CI is building with `thumb 7m` target.