Merge EmbeddedLLM/vllm-rocm into vLLM main (#1836)

Co-authored-by: Philipp Moritz <pcmoritz@gmail.com>
Co-authored-by: Amir Balwel <amoooori04@gmail.com>
Co-authored-by: root <kuanfu.liu@akirakan.com>
Co-authored-by: tjtanaa <tunjian.tan@embeddedllm.com>
Co-authored-by: kuanfu <kuanfu.liu@embeddedllm.com>
Co-authored-by: miloice <17350011+kliuae@users.noreply.github.com>

vllm/engine/ray_utils.py

@@ -3,6 +3,7 @@ from typing import Optional, Tuple, TYPE_CHECKING
 
 from vllm.config import ParallelConfig
 from vllm.logger import init_logger
+from vllm.utils import is_hip
 
 logger = init_logger(__name__)
 
@@ -73,7 +74,12 @@ def initialize_cluster(
                 "Ray is not installed. Please install Ray to use distributed "
                 "serving.")
         # Connect to a ray cluster.
-        ray.init(address=ray_address, ignore_reinit_error=True)
+        if is_hip():
+            ray.init(address=ray_address,
+                     ignore_reinit_error=True,
+                     num_gpus=parallel_config.world_size)
+        else:
+            ray.init(address=ray_address, ignore_reinit_error=True)
 
     if not parallel_config.worker_use_ray:
         # Initialize cluster locally.