Lamma

Hello, in the original K-Diffusion paper the authors report FID scores for CIFAR in the low-single-digits range (eg 1.8). However, the FID scores from this repo all give in the...

### System Info Running a standard training loop where I save the optimizer state_dict using opt.state_dict(). Upon loading using opt.load_state_dict() to resume, the model immediately NaNs after the first backprop...