SWDEV-554030 - lmx-v1-fp32-512x512 Failing to Run on GPUEP Topaz PhotoAI

[KenLagos]

Description

• Add check for implicit inputs for "If" node for destination argument index

• Disable fusion if subgraph parent node is "If" or if graph has "If" node

[KenLagos]

@apwojcik

Regarding testing the change, I have been testing with modelbench-ort and Amuse and haven't seen any regressions with the change but I'm unsure if that is good enough.

Are there any other things I should test with to confirm this change doesn't cause any regressions?

[TedThemistokleous]

I don't understand the change here, sure it fixes the issue but we're just going to turn off fusions entirely then for if? Why were we getting an out of bound index in the first place?

[KenLagos]

contined from resolved comment chain

Just to clarify, this change isn't a complete fix but an improvement to our current situation by changing our problem from a functional issue to a performance issue for models containing "If" nodes. Testing on my local machine suggests models with an "If" node just fail to run. It isn't an exhaustive test but a basic test model I made with an "If" node failed to run and would suggest "If" handling in most if not all models are probably currently broken. Resolving the wrong index error does help but still fails to run indicating another issue.

This change handles the incorrect index error by taking into account implicit inputs that are used by onnxruntime to determine the dst index when creating edges and adding it to the getSubGraph logic.

in graph.cc AddEdge() function

Added logic in GetSubGraph() migraphx_execution_provider.cc which avoids arg_index >= input_defs.size() error

What does the then_branch have in it? Should have its own list of operators or things when you click on it? Are you viewing this in netron?

I'm using netron.

Test "If" Model

The branches looks like below:

else:

then:

the failing lmx model is a bit more complicated and looks like.

else

then: