Gemma4 Mtp Runtime Kernels

void trt_edgellm::kernel::launchGemma4MTPBuildVerifyTokens(

int32_t const *seedTokenIds,

int32_t const *draftTokenIds,

int32_t *verifyTokenIds,

int32_t batchSize,

int32_t draftingStep,

cudaStream_t stream,

)

Build linear target-verification token IDs: [seed, draft_0, …, draft_{N-1}].

void trt_edgellm::kernel::launchGemma4MTPGatherSeedHidden(

void const *sourceHiddenStates,

void *seedHiddenStates,

int32_t const *sourceTokenIndices,

int32_t batchSize,

int64_t sourceSeqLen,

int64_t hiddenSize,

size_t elementBytes,

cudaStream_t stream,

)

Gather one target hidden-state row per batch into the assistant seed-hidden input.

void trt_edgellm::kernel::launchGemma4MTPStoreDraftToken(

int32_t const *selectedTokenIds,

int32_t *draftTokenIds,

int32_t batchSize,

int32_t draftingStep,

int32_t step,

cudaStream_t stream,

)

Store sampled assistant token IDs into the draft-token matrix at the current draft step.