tvm icon indicating copy to clipboard operation
tvm copied to clipboard
verified publisher icon pytorch

Question on FuseSupportedOps pass

Open vdantu opened this issue 6 years ago • 2 comments

I was just playing around with this optimization pass with the main PyTorch project. I was curious to understand why control flow operators are not handled here. What was the issue? I had lodged a question on the PyTorch discuss forum about the issue I was facing in my experimentation. Would anyone here be able to help with it?

TL;DR : In my experimentation, I was finding that subgraph_utils in PyTorch repo doesn't handle merging of nodes, when the inputs of the node belongs to outputs of another block's nodes. Is this understanding correct? Any pointers would be highly appreciated :)

vdantu avatar Oct 21 '19 05:10 vdantu

@serhaty I think you saw the same issue and was working on a fix?

lly-zero-one avatar Oct 22 '19 00:10 lly-zero-one

@lly-zero-one : Thanks for responding. Is there any communication channel (slack/messenger) thats open to public? I would love to join it and try to see if my understanding is right.

vdantu avatar Oct 22 '19 16:10 vdantu