AndrewZhaoLuo opened a new pull request #10540:
URL: https://github.com/apache/tvm/pull/10540


   Reverts apache/tvm#9829
   
   The above behavior avoids having TE output Tensors change in address 
everytime by caching the tensors instead of reallocating everytime. However in 
doing so they introduce a cyclic dependency. Each `Operation` has a list of 
`Tensors`. However, each `Tensor` has a reference to the `Operation`. As a 
result, each will always have a reference to each other and never be deleted 
under TVM's reference counting system.
   
   Possible alternatives:
   - Something like a weak_ptr, probably so a counted reference exists between 
the tensor --> operator but not in the opposite direction
   - Refactoring above PR to somehow avoid cyclic dependency
   
   Thanks to @mbrookhart for help in finding this memory error!


-- 
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]


Reply via email to