For model-parallel&multi-gpu training and inference

Hi,

1. I am looking for materials for how GPU memory in DeepSpeed is used for model-parallel&multi-gpu training setting (=means all weights are not fit into single GPU memory even DeepSpeed is applied).

For my current understanding, the following post only visualizes data-parallel&multi-gpu setting. https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/. Any reference materials are appreciated to understand the internals. (both DeepSpeed doc or outside doc is ok).

2. Is DeepSpeed provides inference or serving API for model-parallel&multi-gpu environment? (Because in model-parallel setting, weights are partitioned, I think this is non-trivial to serve. I currently looking for using typical inference function in the training process, however.)

Thank you.

Start with the linked ZeRO/DeepSpeed blog post and the existing DeepSpeed documentation on model parallelism, multi-GPU training, and inference. The issue does not name a file, test, or entry point; done would require a decided documentation scope or a defined serving API requirement, plus references that explain memory use and supported inference behavior.