ColossalAI icon indicating copy to clipboard operation
ColossalAI copied to clipboard

Published 20 hours ago • verified publisher icon hpcaitech

[autoparallel] support more flexible data types

Open YuliangLiu0306 opened this issue 2 years ago • 0 comments

Previous problem

The output of majority nodes are a torch.Tensor. However, some kernels, such as split, have a tuple type output, which will lead to the previous implementation broken.

What does this PR do

In this PR, outputs of nodes are not supposed to be a tensor, more flexible data types are supported now.

YuliangLiu0306 avatar Nov 16 '22 13:11 YuliangLiu0306