You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The Tensor struct in the python frontend is not the same as a TensorView in nvFuser's IR that has more information. Therefore, what is the motivating use case to add dtype to the struct?
You can perform heuristic and segmentation analysis without building the cpp Fusion IR.
You wanted to do segmentation in python. How you do better than the CPP algorithm without additional information?
You already have the fusion DAG, but you need the tensor sizes and dtype information to score the segments. Most device information is already available through pytorch.
Segmentation decomposes a fusion into a directed acyclic graph (DAG) of sub-fusions. You can map a fusion directly to its component sub-fusions without building the CPP Fusion IR.
Currently, we only track the number of dimensions in the Tensor struct. Tracking the dtype of the tensor would also be useful information to track.
Reference: https://github.com/NVIDIA/Fuser/blob/main/csrc/python_frontend/fusion_definition.h#L31-L75
The text was updated successfully, but these errors were encountered: