NickLucche
commented
2 years ago - Introduces the Model Parallel option: model is split into 4 main components/stages, that can be distributed over multiple GPUs. Worker takes care of moving tensors from one device to the other. Enables to run model on multiple "small" GPUs
- improvements to nsfw filter removal (saving on memory and inference speed)
- fixes to seed fixing for data parallel and env variables reading
- readme samples update