Question: _set_cluster_n() in RMSNorm fwd

[karthik-man]

In RMSNorm fwd _set_cluster_n(), why are the N thresholds (for picking larger cluster_n), smaller for 16-bit types than for 32-bit types?

[tridao]

It's empirical. I'm not sure why.

[karthik-man]

Thanks. Did you tune this on Blackwell too? Do you have any numbers for reduction kernels on Blackwell?

[tridao]

It was tuned on Hopper. Blackwell might need slightly different numbers, I'm not sure