MCPcopy
hub / github.com/dmlc/dgl / construct_torch_dist_launcher_cmd

Function construct_torch_dist_launcher_cmd

tools/distgraphlaunch.py:192–226  ·  view source on GitHub ↗

Constructs the torch distributed launcher command. Helper function. Args: num_trainers: num_nodes: node_rank: master_addr: master_port: Returns: cmd_str.

(
    num_trainers: int,
    num_nodes: int,
    node_rank: int,
    master_addr: str,
    master_port: int,
)

Source from the content-addressed store, hash-verified

190
191
192def construct_torch_dist_launcher_cmd(
193 num_trainers: int,
194 num_nodes: int,
195 node_rank: int,
196 master_addr: str,
197 master_port: int,
198) -> str:
199 """Constructs the torch distributed launcher command.
200 Helper function.
201
202 Args:
203 num_trainers:
204 num_nodes:
205 node_rank:
206 master_addr:
207 master_port:
208
209 Returns:
210 cmd_str.
211 """
212 torch_cmd_template = (
213 "-m torch.distributed.launch "
214 "--nproc_per_node={nproc_per_node} "
215 "--nnodes={nnodes} "
216 "--node_rank={node_rank} "
217 "--master_addr={master_addr} "
218 "--master_port={master_port}"
219 )
220 return torch_cmd_template.format(
221 nproc_per_node=num_trainers,
222 nnodes=num_nodes,
223 node_rank=node_rank,
224 master_addr=master_addr,
225 master_port=master_port,
226 )
227
228
229def wrap_udf_in_torch_dist_launcher(

Callers 1

Calls 1

formatMethod · 0.80

Tested by

no test coverage detected