[BUG] [Triage] Several potential Cutlass illegal memory access violations related to SM90 EVT

[kadeng]

While working on the Pytorch Inductor CUDA/Cutlass backend, I have encountered several cases where I'm seeing Illegal Memory Accesses, which I could not trace down to an error in the user code, so these might be Cutlass bugs.

I have created a public gist here on Github where you will find the source code to 4 (generated) standalone Cutlass based
executables, each failing with some kind of error:
https://gist.github.com/kadeng/8fa35f5e42ed111f8de8d8623f16ec88

Simple build/run scripts are included in the gist above, you only need to modify them to point to the right Cutlass and CUDA Toolkit directory.

Environment: CUDA 12.1, H100 80 GB, Linux x64, Cutlass 3.3.0

You might want to try to reproduce these errors via the provided code and scripts and triage whether they are Cutlass bugs.

[hwu36]

@jackkosaian

[kadeng]

Nevermind, I found the problem. Not a Cutlass bug.