Eric Jang

Friendly ping - any updates? I use AMD gpus for hobby deep learning projects at home and it's demoralizing that I can't train my NNs very fast (not to mention...

Hi, thanks all for the recommendations. switching to fp32 seems to indeed speed things up: Dawn.py (21 sec / step): ``` epoch train time train loss train acc valid time...

I followed the fork https://github.com/mlfoundations/open_flamingo/commit/25b17319723b41f900cd52a389466b97c053695d but am unable to load the Llama weights on the 3090 - crashes when I try to load the llama weights in fp16 ``` NVIDIA-SMI...

Never mind, I was able to get things working. Had to decrease batch size to 2 to have sufficient memory. Note also that one has to cast the inputs to...