Tensor Parallel Nccl Resources#

std::unique_ptr<PluginAllReducePathResources> trt_edgellm::rt::createTensorParallelNcclResources(
int32_t tpSize,
std::vector<int32_t> localRanks,
std::vector<int32_t> localDevices,
std::vector<void*> ncclComms,
bool ownsNcclComms
)#

Create and register the NCCL execution-path resources for one tensor-parallel group.