Fix --enable_each_rank_log when used with PDSH multi-node runner (#6863)

This PR addresses fixes #6859 by threading this argument into the deepspeed launcher command build by PDSHRunner. --------- Co-authored-by: Logan Adams <[email protected]>
microsoft · Dec 17, 2024 · a964e43 · a964e43
1 parent da771ed
commit a964e43
Showing 1 changed file with 2 additions and 0 deletions.
diff --git a/deepspeed/launcher/multinode_runner.py b/deepspeed/launcher/multinode_runner.py
@@ -104,6 +104,8 @@ def get_cmd(self, environment, active_resources):
             deepspeed_launch.append("--no_local_rank")
         if self.args.save_pid:
             deepspeed_launch += ["--save_pid", f"{os.getpid()}"]
+        if self.args.enable_each_rank_log:
+            deepspeed_launch.append(f"--enable_each_rank_log={self.args.enable_each_rank_log}")
         if self.args.elastic_training:
             deepspeed_launch.append("--enable_elastic_training")
             deepspeed_launch.append(f"--max_elastic_nodes={self.args.max_elastic_nodes}")