vllm.models.glm5next.nvidia.specialized_decode.cutedsl
¶
CuTe DSL helpers copied from Kimi K3's specialized decode (host, policy,
primitives, the split-K GEMM in splitk and projection) and from DeepSeek-V4.1's
(finalize_publish), unchanged apart from imports, so this package does not depend
on those models' trees.
Modules:
-
finalize_publish–DeepSeek-V4.1's MoE finalize (top-k weighted sum plus the shared expert)
-
host–Copy of Kimi K3's
-
policy–Copy of Kimi K3's
-
primitives–Copy of Kimi K3's
-
projection–Copy of Kimi K3's
-
splitk–Copy of Kimi K3's