Update env var RANK to NODE_RANK.

Author: qiuosier

ads/jobs/builders/runtimes/pytorch_runtime.py

@@ -205,7 +205,8 @@ def run(self, dsc_job, **kwargs):
                 if not envs:
                     envs = {}
                 # Huggingface accelerate requires machine rank
-                envs["RANK"] = str(i)
+                # Here we use NODE_RANK to store the machine rank
+                envs["NODE_RANK"] = str(i)
                 envs["WORLD_SIZE"] = str(replicas)
                 if main_run:
                     envs["MAIN_JOB_RUN_OCID"] = main_run.id

ads/jobs/templates/driver_pytorch.py

@@ -694,7 +694,7 @@ def __init__(self, code_dir: str = driver_utils.DEFAULT_CODE_DIR) -> None:
         # --multi_gpu will be set automatically if there is more than 1 GPU
         # self.multi_gpu = bool(self.node_count > 1 or self.gpu_count > 1)
         self.num_machines = self.node_count
-        self.machine_rank = os.environ["RANK"]
+        self.machine_rank = os.environ["NODE_RANK"]
         # Total number of processes across all nodes
         # Here we assume all nodes are having the same shape
         self.num_processes = (self.gpu_count if self.gpu_count else 1) * self.node_count