Closed lukstafi closed 1 year ago
For the GPU backends, with a global on-device memory, both are needed. For CPU backends, maybe only ~final would be used, for cases of "non-recurrent" tensors (without read-before-write).
~final
The current approach is that cross-device memory management will be manual.
For the GPU backends, with a global on-device memory, both are needed. For CPU backends, maybe only
~final
would be used, for cases of "non-recurrent" tensors (without read-before-write).