Module bound 

2026-09-25: K3BoundLayer, the Kimi K3 decoder layer the engine binds, and K3HostShared, the state its layers share.

Decode runs on the host between a copy of the hidden row out and a copy back (host_decode.rs). The KDA mixer uses CUDA kda_decode with per-sequence device state unless K3_CUDA_KDA=0, the MLA mixer uses CUDA mla_decode unless K3_CUDA_MLA=0, and packed LatentMoE experts use moe_w4a16_grouped_gemm_ptrtable_e8m0.

Owner: model-arch, Kimi K3. Invariants:

• decode_graph_unsupported, decode_multi_seq_unsupported, decode_verify_multi_unsupported and decode_rollback_unsupported are all true.
• Layer state is always a K3CpuFallbackState; the layer never uses the SSM pool.

Structs

K3BoundLayer

2026-09-25: One decoder layer. weights must be BF16 or FP32: the host bind refuses any other dtype.

K3HostShared

2026-09-25: One per loaded model; every decoder layer holds the same Arc.

WeightMeta

2026-09-25: Checkpoint name, dtype and element count of one resident weight. The first decode uses them to copy the weight to host FP32.