[QST] Variable size Gemm that can be Cuda graphed

It is common with LLMs to have variable sized inputs, which usually means that the GEMMs cannot be used with a cuda graph.
However, there is a specific case for which a solution is possible. When Computing A * B, the common dimension of the two matrices does not change the launch configuration. This means that there could be an early exit from the loop that traverses the tiles in a single strip.
If the size is stored on device (rather than passed through the kernel API), this means that the kernel can be used with cuda graph as well.
My question is:

• Does an existing Cutlass functionality already expose this varying size behavior that can be graphed?
• Is there a way to add this "early exit" with a minimal intervention into the code? Is there a relevant example?

This is super useful for low latency attention computation with small batch...

The issue does not name a file, test, or entry point. Start by reviewing existing CUTLASS GEMM and CUDA graph functionality, then determine whether device-stored varying dimensions already work. Done means documenting an applicable existing mechanism or defining and validating a minimal early-exit approach, with a relevant example if one exists.