vllm.models.glm5next.nvidia.specialized_decode

Modules:

• all_reduce_mhc –

  Lamport TP all-reduce fused with GLM-5.3-Flash's mHC sublayer boundary.

• all_reduce_norm –

  Lamport TP all-reduce fused with a residual add and RMSNorm.

• cutedsl –

  CuTe DSL helpers copied from Kimi K3's specialized decode (host, policy,

• dense_mlp –

  The dense (first first_k_dense_replace) layers' MLP of a few decode tokens

• dsa_attention –

  DSA attention stage (dsa_attn): the MLA half of the DSA layer around the

• gemm –

  Skinny BF16 projections for decode steps: the CuTe skinny GEMM that issues its

• grouped_moe –

  Routed experts plus the shared expert of 8-64 decode tokens as one persistent

• indexer –

  DSA indexer stage: the indexer's input projections and per-token query

• indexer_query –

  The indexer's per-token epilogue in one kernel: key LayerNorm, query

• kda –

  KDA decode stage: the stock input projection and low-rank output-gate GEMMs,

• kda_split –

  Kimi K3's cluster-split KDA recurrence (KdaSplit) with GLM-5.3-Flash's low-rank

• kpool –

  The kpool indexer's decode kernels in CuTe DSL, bit for bit with the stock

• linear –

  Skinny projections for stock BF16 linear modules inside planned decode steps.

• logits_gather –

  All-gather of the vocab-parallel LM head logits straight into [M, vocab].

• mega_moe –

  Routed experts of a few decode tokens as one kernel, one 16-CTA cluster per

• policy –

  Which implementation each stage runs for a step of tokens tokens (requests x

• runtime –

  GLM-5.3-Flash specialized decode: a decode executor over the stock backbone for

• shared_expert –

  The shared expert of a few decode tokens beside the mega MoE kernel.