vllm.models.kimi_k3.nvidia.specialized_decode
¶
Modules:
-
cutedsl– -
drafter–Kimi K3 / K3.1 DSpark drafter decode path (dense MLA draft layers).
-
recoverssm–The specialized decode's deferred RecoverSSM (backend
fp32_deferred). -
runtime–Kimi K3 specialized decode: a CuTe decode executor over the stock backbone, enabled
-
symm–Symmetric-memory rendezvous that spends NVLink multicast objects only where a