masahi commented on issue #13296:
URL: https://github.com/apache/tvm/issues/13296#issuecomment-1306127542

   This is a common problem when using TVM on Windows. Can you identify which 
pass is causing the stack overflow? I think the stack overflow is coming from 
https://github.com/apache/tvm/blob/main/src/ir/transform.cc#L453, where 
different relay and TIR passes are applied. So you can add debug statements 
before / after this line to identify what pass is running when the stack 
overflow happens. The name of a pass can be printed by `LOG(INFO) << 
pass_info->name`.
   
   > Would be good to understand the kind of tools /process to debug 
such/similar issues
   
   For this one I did a very dumb thing: I knew that outputs were correct up to 
some point in the model. Starting from there, I repeated modifying 
`torchvision/models/detection/ssd.py` and gradually add more pytorch operations 
until the output became different. I noticed that the lines 
https://github.com/pytorch/vision/blob/main/torchvision/models/detection/ssd.py#L433-L435
 were converted incorrectly - indexing by mask results in a dynamic shape while 
the converted Relay model didn't have any dynamism around there. 
   


-- 
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