Mtp Decoder

class MTPDecoder : public trt_edgellm::rt::DecodingStrategy

Public Functions

MTPDecoder(

DecodingRuntimeContext &runtime,

SpecDecodeDraftingConfig const &draftingConfig,

std::unique_ptr<EngineExecutor> draftExecutor,

ExternalWeightManager draftWeights,

cudaStream_t stream

)

inline virtual DecodingStrategyKind kind() const noexcept override

inline virtual char const *name() const noexcept override

inline virtual bool isSpeculative() const noexcept override

virtual DecodingKvHeadroom requiredKvHeadroom() const override

virtual bool decodeStep(DecodingInferenceContext &context) override

virtual bool captureCudaGraphs(cudaStream_t stream) override

virtual bool initializeForGeneration(

DecodingInferenceContext &context

) override

Hybrid+MTP endpoint reuse runs the draft prefill pre-publication (mirrors EagleDecoder). Default MTP keeps its decode-round-0 draft prefill: this override is a no-op unless context.hybridMtpEndpointReuse is set.

virtual int64_t getRequiredContextMemorySize(

) const noexcept override

virtual void setContextMemory(Tensor &memory) override

virtual bool hasSystemPromptKVCache(

SystemPromptCacheKey const &key

) const override

virtual void restoreSystemPromptKVCache(

SystemPromptCacheKey const &key,

int32_t batchIdx,

cudaStream_t stream

) override

virtual bool runSystemPromptPrefill(

DecodingInferenceContext &context

) override

virtual void saveSystemPromptKVCache(

SystemPromptCacheKey const &key,

std::string const &prompt,

std::vector<tokenizer::Rank> const &tokenizedPrompt,

int32_t promptIdsLength,

cudaStream_t stream

) override

virtual void resetForNewSequences(

Tensor &reuseLengths,

cudaStream_t stream

) override

virtual void onBatchEvict(

std::vector<int32_t> const &batchMapping,

int32_t oldActiveBatch,

int32_t newActiveBatch,

Tensor &deviceBatchMapping,

cudaStream_t stream

) override