Hi, thank you for your great work!
I'm currently working on a scene completion task using MinkowskiEngine v0.5.4, and I’m also using DeepSpeed to split model parameters across two GPUs (RTX 3060 and RTX 3090).
However, DeepSpeed requires some tensors (especially custom ones) to have the "dict" attribute. Unfortunately, SparseTensor from MinkowskiEngine does not have the dict attribute, which causes the following error when trying to move parameters across devices:
Issue when use DeepSpeed to separate parameters in 2 devices:
SparseTensor doesn't have attribute "dict"
My questions are:
-
Is there any version of MinkowskiEngine where SparseTensor has a dict attribute?
-
If not, is there a workaround or recommended way to modify SparseTensor (e.g., subclass or wrapper) so that DeepSpeed can work with it?
Any suggestions or best practices would be greatly appreciated. Thank you!