Atlas Inference Products

This section describes the communication operators supported by the Atlas inference product.

  • Single-operator zero-copy: To reduce the memory copy overhead, HCCL can directly operate the memory transferred by the service to improve communication performance.
  • Communication operator retry: When the communication is intermittently interrupted due to a network fault, HCCL attempts to retry the communication operator to improve communication stability.
  • Deterministic computing: The same output is generated if a reduction communication operator is executed for multiple times with the same hardware and input.
  • The following tables list the communication operators supported in expansion mode. Modes not listed are unsupported.
  • In tables in this section, "√" indicates supported, "☓" indicates unsupported, and "N/A" indicates not involved. The Atlas inference product does not support single-operator zero-copy or operator retry.
  • Operators and network modes not listed are not supported.

HOST

Operator

Network Mode

Single-Operator Zero-Copy

Deterministic Computing

Retry

Intra-Rank Communication

Inter-Rank Communication

AllGather

Single-operator

N/A

Graph (Ascend IR)

N/A

Graph capture (ACLGraph)

N/A

AllGatherV

Single-operator

N/A

Graph (Ascend IR)

N/A

Graph capture (ACLGraph)

N/A

AllReduce

Single-operator

Graph (Ascend IR)

Graph capture (ACLGraph)

ReduceScatter

Single-operator

Graph (Ascend IR)

Graph capture (ACLGraph)

ReduceScatterV

Single-operator

Graph (Ascend IR)

Graph capture (ACLGraph)

AlltoAll

Single-operator

N/A

Graph (Ascend IR)

N/A

Graph capture (ACLGraph)

N/A

AlltoAllV

Single-operator

N/A

Graph (Ascend IR)

N/A

Graph capture (ACLGraph)

N/A

Send

Single-operator

N/A

Graph (Ascend IR)

N/A

Graph capture (ACLGraph)

N/A

Recv

Single-operator

N/A

Graph (Ascend IR)

N/A

Graph capture (ACLGraph)

N/A

BatchSendRecv

Single-operator

N/A

Graph (Ascend IR)

N/A

Graph capture (ACLGraph)

N/A

AI CPU

Operator

Network Mode

Single-Operator Zero-Copy

Deterministic Computing

Retry

Intra-Rank Communication

Inter-Rank Communication

AllReduce

Single-operator

Graph (Ascend IR)