graphstorm.gsf.initialize

graphstorm.gsf.initialize(ip_config=None, backend='gloo', local_rank=0, use_wholegraph=False, use_graphbolt=False)

Initialize distributed training and inference context. For GraphStorm Standalone mode, no argument is needed. For Distributed mode, users need to provide an IP address list file.

# Standalone mode
import graphstorm as gs
gs.initialize()

# distributed mode
import graphstorm as gs
gs.initialize(ip_config="/tmp/ip_list.txt")

Parameters

ip_config: str

File path of the IP address file, e.g., /tmp/ip_list.txt Default: None.

backend: str

Torch distributed backend, e.g., gloo or nccl. Default: gloo.

local_rank: int

The local rank of the current process. Default: 0.

use_wholegraph: bool

Whether to use wholegraph for feature transfer. Default: False.

use_graphbolt: bool

Whether to use GraphBolt graph representation. Requires installed DGL version to be at least 2.1.0. Default: False.

New in version 0.4.0.