Closed
Description
nvfuser should support device
argument.
We could have factory methods in fusion that doesn't have a tensor input, so device
argument would be necessary to determine GPU device for kernel generation.
e.g. in cpp tests
Fuser/test/test_gpu_tensor_factories.cpp
Lines 51 to 56 in 44e99fe
Note that these arguments should also be plumbed all the way through to python API.
Metadata
Metadata
Assignees
Labels
No labels