HcclGroupEnd

Applicability

For Atlas A2 training product/Atlas A2 inference product, only the Atlas 800T A2 training server, Atlas 900 A2 PoD cluster basic unit, and Atlas 200T A2 Box16 heterogeneous subrack are supported.

Function

Ends a group call.

Prototype

1

HcclResult HcclGroupEnd()

Parameters

None.

Returns

HcclResult: HCCL_SUCCESS on success, or else failure.

Constraints

• Using the group call APIs for communicator management is exclusively supported in single-node environments.
• In a group call, mixing communicator management, collective communication, and point-to-point communication APIs is not allowed.
• Calling the HcclBatchSendRecv API is not supported when grouping multiple point-to-point communications.
• The HcclGroupStart and HcclGroupEnd APIs must be used in pairs, with HcclGroupStart preceding HcclGroupEnd.

Call Example

• Example 1: Single-process multi-thread NPU management

  HcclComm hccl_comms[devCount];
  HcclGroupStart();
  for(int i = 0; i < ndev; i++){
      //
          aclrtSetDevice(i);
          HcclCommInitRootInfo(devCount, &rootInfo, global_rank, &(hccl_comms[i]));
      }
  HcclGroupEnd();

• Example 2: Grouping multiple collective communication operations

  HcclGroupStart();
      HCCLCHECK(HcclReduceScatter(sendBuf, recvBuf, 1, HCCL_DATA_TYPE_FP32, HCCL_REDUCE_SUM, hcclComm, stream));
      HCCLCHECK(HcclAllGather(recvBuf, sendBuf, 1, HCCL_DATA_TYPE_FP32, hcclComm, stream));
  HcclGroupEnd();

• Example 3: Grouping multiple point-to-point communication operations

  HcclGroupStart();
  for(int i = 0; i < devCount; i++){
      HCCLCHECK(HcclSend(sendBuf[i], count, HCCL_DATA_TYPE_FP32, i, hcclComm, stream));
      }
  for(int i = 0; i < devCount; i++){
      HCCLCHECK(HcclRecv(recvBuf[i], count, HCCL_DATA_TYPE_FP32, i, hcclComm, stream));
      }
  HcclGroupEnd();