Does cumm support a dynamic mode to decide to use cutlass or its own gemm based on input shapes?

[umiswing]

Hello! I'm interested in cumm and find it's very fast. I read cumm's code and found it seems that cumm provides a gemm implement from scratch while also providing a cutlass mode. And cumm will use its own gemm by default. I do some experiments, and I find that for some shapes, the cutlass will get faster, and for others, cumm's gemm will be faster. Does cumm support a dynamic mode to decide to use cutlass or its own gemm based on the input shapes?

[FindDefinition]

currently not, but looks like a good feature to use new cutlass feature quickly instead of re-implement them. I will consider this feature.

[NotUr77]

How to enable and correctly build with Cutlass mode in cumm?

While exploring the codebase, I noticed that cumm contains references to a cutlass mode option (e.g. via CUTLASS_MODE). However, when I try to enable this mode and compile the code , the build fails due to various template errors or missing specializations.

My Questions

Is Cutlass mode in cumm still actively supported, or is it experimental/incomplete?
What is the correct way to enable and build cumm in Cutlass mode?