Int4 Cute Dsl Repack

int32_t trt_edgellm::kernel::int4CuteDslFragmentRows(

int32_t N,

int32_t K

)

Number of INT8 [rows, 512] rows required by Int4GroupwiseGemmPluginV2.

cudaError_t trt_edgellm::kernel::launchModelOptInt4CuteDslRepack(

uint8_t const *weightNhalfK,

int8_t *fragmentRows512,

int32_t N,

int32_t K,

cudaStream_t stream

)

Repack ModelOpt packed [N/2,K] weights directly from checkpoint storage.

cudaError_t trt_edgellm::kernel::launchGptqInt4CuteDslRepack(

int32_t const *qweightK8N,

int32_t const *qzerosGN8,

int32_t const *activationPermutation,

int8_t *fragmentRows512,

int32_t N,

int32_t K,

int32_t numGroups,

int32_t groupSize,

int32_t zeroPointOffset,

cudaStream_t stream

)

Repack GPTQ [K/8,N] weights directly from checkpoint storage.

cudaError_t trt_edgellm::kernel::launchAwqInt4CuteDslRepack(

int32_t const *qweightKN8,

int32_t const *qzerosGN8,

int8_t *fragmentRows512,

int32_t N,

int32_t K,

int32_t numGroups,

int32_t groupSize,

cudaStream_t stream

)

Repack classic AWQ [K,N/8] weights directly from checkpoint storage.