Domino Decoder

class DominoDecoder : public trt_edgellm::rt::DFlashDecoder

Domino: the DFlash V1 cached-block draft plus a serial GRU correction rollout.

The draft engine emits a shared candidate pool with its base logits for every proposal position, the pool’s codebook rows, and the projected draft hidden states. This decoder replays the correction head (GRU + state projection from the domino_heads.safetensors sidecar) along the greedy chain and records each position’s corrected pool scores. The DDTree builder inherited from DFlashDecoder then selects the verify tree from those scores exactly as it does from DFlash logits, so topK=1 is the chain and topK>1 fans out around it.

Public Functions

DominoDecoder(

DecodingRuntimeContext &runtime,

std::filesystem::path const &engineDir,

dflash_utils::CachedBlockDraftRuntimeConfig blockDraftConfig,

std::unique_ptr<EngineExecutor> draftExecutor,

ExternalWeightManager draftWeights,

cudaStream_t stream

)