vllm.models.mistral.mistral_large4.amd.specialized_decode.mono
¶
Modules:
-
common–Mono decode: a model's decode step as one persistent kernel per layer, with the
-
geometry–Shared compile-time layouts and schedules for fused model-layer kernels.
-
geometry_config–ML4's own geometry, and the widths its kernel is built for.
-
kernel–ML4 indexed decode MonoKernel: one persistent launch per TP rank.
-
layout–Compile-time storage layout and CTA schedule for the ML4 MonoKernel.
-
op–Host wrapper for the ML4 indexed decode MonoKernel.