Open SAT431 opened 2 weeks ago
My graphics card is an RTX 3060 with 12GB, and I have 32GB of RAM. I'm using the Tora model with BF16 precision, FP8 Transformer enabled, and CPU offload activated. The generation step is 20, and it takes around 20 minutes to generate a video. The average GPU usage is between 6-8GB, and the video output is quite good, with almost no facial distortion in the characters.
https://github.com/user-attachments/assets/dd225b45-3484-4adc-964a-00d939415e26
With the GGUF model, around 12-13GB all on GPU. Offloading works somewhat, hard to say what's the actual minimum though, if I enable it then it uses few GB less.