SetSkipLocalRankCopy

Function

Sets whether the compute result of the communication algorithm on the local rank is to be output to the address of the destination data buffer.

Prototype

1
uint32_t SetSkipLocalRankCopy(uint8_t skipLocalRankCopy)

Parameters

Table 1 Parameters

Parameter

Input/Output

Description

skipLocalRankCopy

Input

Whether the computing result of the current device communication algorithm is to be output to the recvBuf (address of the destination data buffer).

For the Atlas A2 training products/Atlas A2 inference products, this parameter is configured only for the AllGather and AlltoAll algorithms. The value is of the uint8_t type. The options are as follows:

  • 0: the computing result of the current device communication algorithm is output (default behavior when this API is not called).
  • 1: the computing result of the current device communication algorithm is not output. In this case, the communication result data of the current device is not copied, and the operator performance is improved. For example, when eight devices are used, the current device obtains only part of data of other cards. In this case, this parameter can be set to 1.

For the Atlas A3 training products/Atlas A3 inference products, this parameter is reserved and does not take effect after being configured.

Returns

  • 0: Setting succeeded.
  • Other values: Setting failed.

Restrictions

None

Example

For details about the calling example of this API, see Example.