Revert "fix"

This reverts commit 6907571.

Signed-off-by: Cody Yu <hao.yu.cody@gmail.com>

Author: comaniac

examples/offline_inference/rlhf.py

@@ -44,9 +44,8 @@ def __init__(self, *args, **kwargs):
 train_model.to("cuda:0")
 """
 Start the inference process, here we use vLLM to hold a model on GPU 1 and 
-GPU 2 by creating a Ray placement group. The placement group will be passed
-to the worker processes spawned by vLLM. For the details on how to use Ray,
-please refer to the Ray documentation https://docs.ray.io/en/latest/ .
+GPU 2. For the details on how to use ray, please refer to the ray 
+documentation https://docs.ray.io/en/latest/ .
 """
 os.environ["CUDA_VISIBLE_DEVICES"] = "1,2"
 ray.init()

vllm/envs.py

@@ -86,7 +86,6 @@
     VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON: bool = False
     VLLM_RAY_PER_WORKER_GPUS: float = 1.0
     VLLM_RAY_BUNDLE_INDICES: str = ""
-    VLLM_RAY_PLACEMENT_GROUP: Optional[str] = None
     VLLM_CUDART_SO_PATH: Optional[str] = None
     VLLM_USE_HPU_CONTIGUOUS_CACHE_FETCH: bool = True
     VLLM_DP_RANK: int = 0
@@ -578,12 +577,6 @@ def maybe_convert_int(value: Optional[str]) -> Optional[int]:
     "VLLM_RAY_BUNDLE_INDICES":
     lambda: os.getenv("VLLM_RAY_BUNDLE_INDICES", ""),
 
-    # Ray placement group (serialized string), if it is set and
-    # ray.util.get_current_placement_group() is None, it will be used as the
-    # placement group in vLLM Ray executor.
-    "VLLM_RAY_PLACEMENT_GROUP":
-    lambda: os.getenv("VLLM_RAY_PLACEMENT_GROUP", None),
-
     # In some system, find_loaded_library() may not work. So we allow users to
     # specify the path through environment variable VLLM_CUDART_SO_PATH.
     "VLLM_CUDART_SO_PATH":

vllm/executor/ray_utils.py

@@ -1,14 +1,12 @@
 # SPDX-License-Identifier: Apache-2.0
 
-import json
 import os
 import time
 from collections import defaultdict
 from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
 
 import msgspec
 
-import vllm.envs as envs
 import vllm.platforms
 from vllm.config import ParallelConfig
 from vllm.executor.msgspec_utils import decode_hook, encode_hook
@@ -164,41 +162,6 @@ def assert_ray_available():
                          "`pip install ray`.") from ray_import_err
 
 
-def serialize_placement_group_to_str(placement_group: "PlacementGroup") -> str:
-    """Serialize a placement group to a string.
-    FIXME: This should be implemented in Ray.
-
-    Args:
-        placement_group: The placement group to serialize.
-
-    Returns:
-        A string representation of the placement group.
-    """
-    placement_group_data = {
-        "id": placement_group.id.hex(),
-        "bundle_cache": placement_group.bundle_cache,
-    }
-    return json.dumps(placement_group_data)
-
-
-def deserialize_placement_group_from_str(
-        placement_group_str: str) -> "PlacementGroup":
-    """Deserialize a placement group from a string.
-    FIXME: This should be implemented in Ray.
-
-    Args:
-        placement_group_str: The string representation of the placement group.
-
-    Returns:
-        A placement group.
-    """
-    placement_group_data = json.loads(placement_group_str)
-    return PlacementGroup(
-        id=ray._raylet.PlacementGroupID.from_hex(placement_group_data["id"]),
-        bundle_cache=placement_group_data["bundle_cache"],
-    )
-
-
 def _verify_bundles(placement_group: "PlacementGroup",
                     parallel_config: ParallelConfig, device_str: str):
     """Verify a given placement group has bundles located in the right place.
@@ -345,19 +308,12 @@ def initialize_ray_cluster(
 
     # Create or get the placement group for worker processes
     if parallel_config.placement_group:
-        logger.info(
-            "Using the existing Ray placement group from parallel config")
         current_placement_group = parallel_config.placement_group
-    elif envs.VLLM_RAY_PLACEMENT_GROUP:
-        logger.info("Using the existing Ray placement group from "
-                    "VLLM_RAY_PLACEMENT_GROUP")
-        current_placement_group = deserialize_placement_group_from_str(
-            envs.VLLM_RAY_PLACEMENT_GROUP)
     else:
-        logger.info("Trying to get the existing Ray placement group")
         current_placement_group = ray.util.get_current_placement_group()
 
     if current_placement_group:
+        logger.info("Using the existing placement group")
 
         # We are in a placement group
         bundles = current_placement_group.bundle_specs