Open Suhail opened 1 year ago
Not sure if related, but I'm seeing significant differences with the groupnorm implementation in AITemplate on the order of 1e-4
compared to PyTorch. Not quite sure why since the accumulation does happen in fp32.... Could also be different issue on my end misconfiguring AITemplate for what I'm doing
Using a DDIM scheduler, I get different results:
Is anyone else getting predictable results with the same seed/scheduler? It makes me think there's a bug possibly.
cc: @terrychenism