Goto

Collaborating Authors

 meta tensor


Experiment with Billion-Parameter Models Faster using DeepSpeed and Meta Tensors

#artificialintelligence

Imagine a situation where you want to know the output shape from an operation. Typically, you would run the operation and check the size of the tensor after. With Meta Tensors, you don't have to compute the output to find the answer. Meta Tensors are just like normal tensors, except they have no data. In PyTorch meta is a device.