Skip to content

vllm.models.minimax_m3.nvidia.specialized_decode.runtime.model_state

The model state of the specialized decode.

MiniMaxM3SpecializedDecodeModelState is the default model state (which MiniMax-M3 runs on without the specialized decode) plus the decode executor, built and warmed up once the weights are loaded: the model runner creates the model state after loading the target and draft models, and before memory profiling and CUDA graph capture.

The attention metadata builds still run eagerly before every step, as in the stock decode. Recording them into CUDA graphs (as DeepSeek-V4.1's metadata_graphs does) needs the MiniMax-M3 sparse and indexer builders audited for host reads first, e.g. the CUTLASS sparse decode's plan cache.