vllm.models.mistral.mistral_large4.amd.specialized_decode.mono.common.weights
¶
Host-side weight preparation every mono kernel needs: the storage
vocabulary (config), MX format conversion (formats), and the packing
that turns a loaded tensor into what a kernel reads (packing, layer).
Modules:
-
config–The vocabulary a mono kernel is built and fed with: how a weight is stored
-
formats–Torch-side MXFP4 and MXFP8 conversion helpers shared by MoE wrappers.
-
layer–Shared host-side weight container for model-specific MonoKernels.
-
packing–Host-side weight packing for fused model-layer launch wrappers.