Nvfp4 Moe Scale Transform

cudaError_t trt_edgellm::kernel::launchNvfp4MoeScaleNormalizeSourceBatchPadded(

uint8_t const *const *firstWeights,

uint8_t const *const *secondWeights,

uint8_t const *const *firstScales,

uint8_t const *const *secondScales,

float const *firstAlphas,

float const *secondAlphas,

int32_t count,

int8_t *mmaLayout,

int32_t sourceRows,

int32_t sourceRowBytes,

int32_t sourceScaleColumns,

int32_t outputRows,

int32_t outputScaleColumns,

Nvfp4MoeFc1Layout layout,

cudaStream_t stream

)

Transform non-contiguous checkpoint scales into an NVFP4 MoE MMA layout. secondSources is null for FC2; otherwise layout

selects the FC1 ! arrangement required by the consuming plugin. cudaError_t launchNvfp4MoeScaleTransformSourceBatch(uint8_t const* const* firstSources,

uint8_t const* const* secondSources, int32_t count, int8_t* mmaLayout, int32_t rows, int32_t Ksf,

Nvfp4MoeFc1Layout layout, cudaStream_t stream);

! Transform scales while zero-padding to larger plugin M/K dimensions. cudaError_t launchNvfp4MoeScaleTransformSourceBatchPadded(uint8_t const* const* firstSources,

uint8_t const* const* secondSources, int32_t count, int8_t* mmaLayout, int32_t sourceRows, int32_t sourceKsf,

int32_t outputRows, int32_t outputKsf, Nvfp4MoeFc1Layout layout, cudaStream_t stream);

! Requantize provider-packed FP4 blocks and write their normalized FP8 ! scales directly in the final plugin layout. secondWeights is null for FC2.