Is it still possible to use UnifiedTensor as described in PyTorch Direct paper in DGL? I see in the docs that this used to be available but was recently deprecated in DGL 1.0 [RFC] API Deprecation and Removal · Issue #4693 · dmlc/dgl · GitHub. Is there still any way to use this? I have a use case where I am implementing a custom data loader and would like to reduce host CPU contention when gathering features with multiple processes.
Hi, UnifiedTensor
can be substituted by the combination of pin_memory_inplace
and gather_pinned_tensor_rows
functions under dgl.utils
, see dgl/pin_memory.py at master · dmlc/dgl · GitHub for reference.
3 Likes
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.