ZoneFv

about inference

1

run inference.py with 7b_tiva_v0, I found that the model can't stable generate image/video/audio. The LLM always output without signal tokens。like this: 

multi-GPU utilization is not high when training, and cpu is the bottleneck.

3

we use 8 A100 80G to training this code，found that most of the time the GPU utilization rate is 0, and the cpu load is high. Is there something wrong...