Nvfp4 W4 A16 Marlin Repack

cudaError_t trt_edgellm::kernel::launchNvfp4W4A16MarlinRepackSourceBatch(

uint8_t const *const *firstWeights,

uint8_t const *const *secondWeights,

int32_t *marlinOutput,

int32_t count,

int32_t sourceN,

int32_t sourceK,

int32_t outputN,

int32_t outputK,

cudaStream_t stream

)

cudaError_t trt_edgellm::kernel::launchNvfp4W4A16ScaleRepack(

uint8_t const *const *firstSources,

uint8_t const *const *secondSources,

int8_t *marlinOutput,

int32_t count,

int32_t sourceN,

int32_t sourceGroups,

int32_t outputN,

int32_t outputGroups,

cudaStream_t stream

)