| 41 | } |
| 42 | |
| 43 | static void InitCudaThread() { |
| 44 | #pragma omp parallel num_threads(num_gpus) |
| 45 | { |
| 46 | auto tid = omp_get_thread_num(); |
| 47 | CUDA_CHECK(cudaSetDevice(tid), "cudaSetDevice"); |
| 48 | } |
| 49 | } |
| 50 | |
| 51 | static void InitNccl(ncclComm_t *comms) { |
| 52 | ncclUniqueId uuid; |