fix typo (#1184)

Co-authored-by: Zhuohan Li <[email protected]>

Author: WrRan

vllm/engine/llm_engine.py

@@ -54,8 +54,8 @@ class LLMEngine:
         scheduler_config: The configuration related to the request scheduler.
         distributed_init_method: The initialization method for distributed
             execution. See `torch.distributed.init_process_group` for details.
-        stage_devices: The list of devices for each stage. Each stage is a list
-            of (rank, node_resource, device) tuples.
+        placement_group: Ray placement group for distributed execution.
+            Required for distributed execution.
         log_stats: Whether to log statistics.
     """
 

vllm/engine/ray_utils.py

@@ -63,11 +63,10 @@ def initialize_cluster(
             the default Ray cluster address.
 
     Returns:
-        A tuple of (`distributed_init_method`, `all_stage_devices`). The
+        A tuple of (`distributed_init_method`, `placement_group`). The
         `distributed_init_method` is the address for initializing the
-        distributed backend. `all_stage_devices` includes device IDs for
-        each worker in each pipeline stage. Each device ID is a tuple of
-        (rank, node resource, device id).
+        distributed backend. `placement_group` includes the specification
+        of the resources for each distributed worker.
     """
     if parallel_config.worker_use_ray or engine_use_ray:
         if ray is None: