echuraev commented on pull request #8313:
URL: https://github.com/apache/tvm/pull/8313#issuecomment-870853665


   > I have no problems merging this as it stands, but I do think I have a 
bigger question:
   > 
   > Should we put some sort of logic into fusion to automatically stop fusion 
if the argument list grows too large per this setting? That should be more 
robust than arbitrarily limiting the fusion depth. It could of course be a 
second PR.
   
   It's a good point and looks reasonable to add such logic into the fusion 
algorithm. It could help us to avoid some possible problems with number of 
arguments in the future. I think it would be better to do such logic in 
separate PR, due to problem with concat layer can appear and without fusing. 
For example, in the original problem we had many inputs because for each input 
to concat we had some preprocessing. And the fusing algorithm wasn't able to 
fuse these inputs into tuple due to this preprocessing. I don't think that the 
limit on the fusion depth can solve this problem.


-- 
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