[object Object][object Object][object Object]undefined[object Object]

ProcessGroupHCCL继承自[object Object]，实现[object Object]后端的相关接口，用于通信算子调用。

[object Object]

torch_npu\csrc\distributed\ProcessGroupHCCL.hpp

[object Object][object Object]

• 为了更好的并发和性能，该类提供的所有HCCL通信都是异步函数，用户需要确保通过WorkHCCL::wait()或WorkHCCL::synchronize()来保证任务完成。
• 该类继承自原生[object Object]，实现的基本通信算子相关接口资料可参考，不在本文档额外补充。基本通信算子列表如下：[object Object] broadcast[object Object] allreduce[object Object] allreduce_coalesced[object Object] reduce[object Object] allgather[object Object] allgather_togather[object Object] allgather_into_tensor_coalesced[object Object] reduce_scatter[object Object] reduce_scatter_tensor_coalesced[object Object] barrier[object Object] gather[object Object] scatter[object Object] send[object Object] recv[object Object] recvAnysource[object Object] alltoall_base[object Object] alltoall[object Object]