vllm.models.kimi_k3.nvidia.kv_balanced_mla.backend

KIMI_K3_KV_BALANCED_MLA: an FP8 MLA decode backend on the KV-balanced kernel.

Selected with --attention-backend KIMI_K3_KV_BALANCED_MLA (and the speculative config's attention_backend for the DSpark draft); it is never chosen automatically. Its cache layout, decode metadata and LSE convention are those the kernel was written for (TokenSpeed MLA's): an FP8 cache of 576 values per token in pages of 32 or 64, uniform decode query lengths, and a log2 LSE for the DCP combine. Every decode step runs on the kernel (decode.py), which raises for a step it can't run. Prefill uses the configured MLA prefill backend.