CreateHcclCrossMulitComm

Description

Creates a multi-server multi-device sub-communicator in ranktable mode when the backend is HCCL.

Definition

HcclComm CreateHcclCrossMulitComm(const char *rankTableFile, uint32_t subCommRankId, std::vector<uint32_t> &rankIds, uint64_t subCommId, uint32_t hcclBufferSize, char *commName);

Parameters

Member	Input/Output	Type	Default Value	Description
rankTableFile	Input	char *	-	Path of the ranktable file. The value is of the string type and contains a maximum of 4096 bytes, including the end-of-text character. For details about the ranktable configuration, see .
subCommRankId	Input	uint32_t	-	Rank ID of the current rank in the sub-communicator.
rankIds	Input	vector<uint32_t>	-	Array consisting of the IDs of the ranks in the sub-communicator in the global communicator. Note: The array should be ordered. The subscript of each rank in the array is mapped to its rank ID in the sub-communicator.
subCommId	Input	uint64_t	-	Current sub-communicator ID.
hcclBufferSize	Input	uint32_t	-	Buffer size of the shared data in the sub-communicator. Unit: MB
commName	Output	char *	-	Name of the communicator obtained after a sub-communicator is created. The value can contain a maximum of 128 characters.

Restrictions

The current rank cannot be created in multiple sub-communicators in parallel. For example, if the current rank is 1 and the two sub-communicators are [0,1] and [1,2], parallel creation is not supported but serial creation is supported. If the current rank is 0 or 2 and the two sub-communicators are [0,1] and [1,2], parallel creation is supported.
subCommRankId < rankIds.size().
Multi-server communication across chip types is not supported.
For the same global communicator, you cannot use the same subCommId to create a sub-communicator.
Supported only by Atlas A2 training products/Atlas A2 inference products and Atlas A3 inference products/Atlas A3 training products.

Parent topic: atb/comm.h