tqchen commented on issue #13678:
URL: https://github.com/apache/tvm/issues/13678#issuecomment-1369820183
Thanks @vvchernov for the explaination. First of all, I am not questioning
the need of zero copy, just how we implement it carefully.
The main concern I have is the API itself and its usage in say VM, which is
not always the correct intention.
Having developer to directly(and intentionally) create an DLManagedTensor,
convert it to NDArray, then pass into SetupInput might be a better apporach
here. Get the ORT integration need, can you point me to the related threads?
One possiblity is to implement a ORTTensorToDLPack in that integration point,
which specifically setup a DLManagedTensor. If there is no ref counting
mechanism there, the worst case is thatw e can follow the same behavior as
`FromExternalDLTensor`, but the surface area is minimized so there will be less
possible misused.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
