Int4 Plugin V1 Repack

cudaError_t trt_edgellm::kernel::launchModelOptInt4PluginV1Repack(

uint8_t const *weightNhalfK,

int8_t *packedNhalfK,

int32_t N,

int32_t K,

cudaStream_t stream

)

cudaError_t trt_edgellm::kernel::launchGptqInt4PluginV1Repack(

int32_t const *qweightK8N,

int32_t const *qzerosGN8,

int32_t const *activationPermutation,

int8_t *packedNhalfK,

int32_t N,

int32_t K,

int32_t numGroups,

int32_t groupSize,

int32_t zeroPointOffset,

cudaStream_t stream

)

cudaError_t trt_edgellm::kernel::launchAwqInt4PluginV1Repack(

int32_t const *qweightKN8,

int32_t const *qzerosGN8,

int8_t *packedNhalfK,

int32_t N,

int32_t K,

int32_t numGroups,

int32_t groupSize,

cudaStream_t stream

)