Gemma4 Mtp Decoder

class Gemma4MTPDecoder : public trt_edgellm::rt::DecodingStrategy

Gemma4 assistant MTP decoder.

This class owns the Gemma4-specific draft-engine binding path. The assistant has no draft KV cache: its past_key_values_* inputs are zero-copy aliases to the base target KV cache selected by kv_sharing_map.

Public Functions

Gemma4MTPDecoder(

DecodingRuntimeContext &runtime,

std::filesystem::path const &engineDir,

SpecDecodeDraftingConfig const &draftingConfig,

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 bool decodeStep(DecodingInferenceContext &context) override

virtual bool captureCudaGraphs(cudaStream_t stream) override

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