Open subho406 opened 1 year ago
Not sure actually, never had a chance to work with TPUs.
I imagine it would be trivial for serial training usecase --serial_mode=True
, just replacing "cuda" with TPU everywhere.
Although SF2 relies on shared memory tensors a lot. If TPU shared tensors behave the same way as CUDA tensors, then I assume it would be easy to implement.
This is an amazing library! I was wondering if the library has any support for TPUs? If not how difficult will it get it working in TPUs? I would be willing to contribute to this if its not super time consuming.