vllm.models.kimi_k3.nvidia

Modules:

• dense_draft –

  Dense MLA draft support for Kimi K3.1 (DSA target).

• dsa –

  Kimi K3.1 DSA (DeepSeek Sparse Attention) for the kimi_k3 stack.

• dspark_mla –

  K3 dense MLA draft model for DSpark speculative decoding.

• gqa_draft –

  GQA drafts (e.g. a Qwen3 sliding-window DSpark) on Kimi-K3: their KV cache grouping.

• index_rope –

  Triton RoPE for the Kimi K3.1 DSA indexer.

• kda –
• kda_metadata –

  Kimi-K3 specialization of GDN attention metadata.

• kv_balanced_mla –

  Kimi K3 KV-balanced MLA decode: TokenSpeed's FP8 MLA decode kernel with each

• latent_moe_runner –
• low_latency_gemm –

  Kimi-K3 decode GEMM selection for unquantized BF16 on SM90/SM100/SM103/SM107.

• mla –

  Clean Multi-head Latent Attention for Kimi-K3 (NVIDIA).

• model –

  Kimi-K3 multimodal model implementation for vLLM.

• model_state –

  Kimi K3's model runner state: the hybrid (KDA + MLA) state, plus the lifecycle of

• mtp –

  Inference-only Kimi-K3 Multi-Token-Prediction (MTP) draft model.

• ops –
• recoverssm –
• specialized_decode –