diff --git a/vllm/core/scheduler.py b/vllm/core/scheduler.py
index de1988eead97b..a4a4285cdf3a2 100644
--- a/vllm/core/scheduler.py
+++ b/vllm/core/scheduler.py
@@ -1088,8 +1088,9 @@ def _can_append_slots(self, seq_group: SequenceGroup) -> bool:
         )
 
     def _allow_async_output_proc(self, seq_group: SequenceGroup) -> bool:
-        no_beam_search = (seq_group.sampling_params.best_of == 1
-                          and not seq_group.sampling_params.use_beam_search)
+        no_beam_search = seq_group.sampling_params is None or (
+            seq_group.sampling_params.best_of == 1
+            and not seq_group.sampling_params.use_beam_search)
 
         return no_beam_search
 
diff --git a/vllm/platforms/rocm.py b/vllm/platforms/rocm.py
index 3f6f5adee5a56..28525e8ff8811 100644
--- a/vllm/platforms/rocm.py
+++ b/vllm/platforms/rocm.py
@@ -1,10 +1,21 @@
+import os
 from functools import lru_cache
 from typing import Tuple
 
 import torch
 
+from vllm.logger import init_logger
+
 from .interface import Platform, PlatformEnum
 
+logger = init_logger(__name__)
+
+if os.environ.get("VLLM_WORKER_MULTIPROC_METHOD", None) in ["fork", None]:
+    logger.warning("`fork` method is not supported by ROCm. "
+                   "VLLM_WORKER_MULTIPROC_METHOD is overridden to"
+                   " `spawn` instead.")
+    os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
 
 class RocmPlatform(Platform):
     _enum = PlatformEnum.ROCM