mstar.utils.mla_latent_norm

mstar.utils.mla_latent_norm#

MLA without decoupled dims (mla_use_nope): both latent RMSNorms and the zero-kpe cache row in one kernel, for components/mla.py.

Functions

mla_latent_norm(fused, q_norm_weight, q_eps, ...)

mla_latent_norm_available(fused, ...)

Whether the kernel serves: triton, bf16 CUDA activations and norm weights, unit-stride rows and a power-of-two latent width.

mstar.utils.mla_latent_norm.mla_latent_norm(fused, q_norm_weight, q_eps, kv_norm_weight, kv_eps, kpe)[source]#
mstar.utils.mla_latent_norm.mla_latent_norm_available(fused, q_norm_weight, kv_norm_weight)[source]#

Whether the kernel serves: triton, bf16 CUDA activations and norm weights, unit-stride rows and a power-of-two latent width.

Parameters:
Return type:

bool