Brett Kuprel

I checked to see if the embedding weights in the BART decoder were the same weights as the embedding weights in the VQGAN detokenizer. It seems they are actually different....

Interesting, I haven't tried float16 with cpu. Thanks for pointing this out

This is great, thanks! I get an error when running this `ImageFont.truetype("verdanab.ttf", 15)`, is there a built-in unicode font that could be used instead?

Would love to use this for the min-dalle space: https://huggingface.co/spaces/kuprel/min-dalle Here's how that model works with progressive outputs: https://replicate.com/kuprel/min-dalle

I'm also getting this error with the latest coremltools version 5.1 when trying to convert the generator network from [stylegan2-ada-pytorch](https://github.com/NVlabs/stylegan2-ada-pytorch) `ValueError: ('Op "x.23" (op_type: conv_transpose) Input weight must be const...

I think I've narrowed down the problem. The error "Input weight must be const" only occurs for `conv_transpose` layers and not `conv` layers. Looking at the MIL ops documentation,`conv_transpose` requires...

This worked for me. Still uses conv_transpose but with constant data ```python import torch from torch.nn import functional def conv_transpose_stride2(x: torch.Tensor, w: torch.Tensor) -> torch.Tensor: dilate = torch.nn.ConvTranspose2d(in_channels=128, out_channels=128, kernel_size=1,...

Would complications be supported?

I got it to train on the GPU using pytorch. It only takes 0.9 seconds to train on `taylorswift.txt` now. I imagine the speedup factor is more significant with larger...

Using an H100 and int16, it's now 108x speedup over the original implementation on M2 air