[QST] Details of Auto Schedule Execution Strategy on Blackwell

On the Blackwell architecture, which is preferred to use Auto Schedule. When running GEMM with Auto Schedule, how many SMs does the scheduler actually assign to the kernel? Does Auto Schedule internally make use of Distributed Shared Memory (DSMEM), and is there a way to inspect this?

Background:
I am currently running into an issue on Blackwell where I explicitly set cluster<_1, _1, _1>, so the kernel should be using only a single CTA cluster. However, when debugging and printing out the shared memory addresses, it appears as if two clusters are involved.

This makes me wonder:

1. Does Auto Schedule internally change the number of SMs assigned to the kernel?

2. Can Auto Schedule implicitly increase the number of clusters, even when the user specifies cluster<1,1,1>?

3. If DSMEM is being used implicitly, how can we confirm or inspect this behavior?

Any clarification on how Auto Schedule interacts with SM assignment and cluster configuration on Blackwell would be greatly appreciated.

The issue names Auto Schedule GEMM execution, SM assignment, CTA clusters, and DSMEM on Blackwell but does not identify files or tests. Start by locating the Auto Schedule GEMM entry point and cluster configuration, then determine how the configured cluster count and DSMEM usage can be inspected; done means documenting the observed behavior and inspection path.