[Misc][MacOS] fix bfloat16 error

Signed-off-by: reidliu41 <reid201711@gmail.com>

Author: reidliu41

vllm/config.py

@@ -7,6 +7,7 @@
 import inspect
 import json
 import re
+import sys
 import textwrap
 import uuid
 import warnings
@@ -3064,7 +3065,7 @@ def _get_and_verify_dtype(
                 torch_dtype = torch.bfloat16
 
             # Deal with torch dtype fallback for device compatibility.
-            from vllm.platforms import current_platform
+            from vllm.platforms import CpuArchEnum, current_platform
             if torch_dtype not in current_platform.supported_dtypes:
                 device_name = current_platform.get_device_name()
 
@@ -3088,6 +3089,15 @@ def _get_and_verify_dtype(
                     "using float16 by default. Please specify `dtype` if you "
                     "want to use float16.")
                 torch_dtype = torch.bfloat16
+
+            if (current_platform.is_cpu() and sys.platform.startswith("darwin")
+                    and current_platform.get_cpu_architecture()
+                    == CpuArchEnum.ARM and config_dtype == torch.bfloat16):
+                logger.info(
+                    "For macOS with Apple Silicon, currently using float16 "
+                    "by default. Please specify `dtype` if you "
+                    "want to use bfloat16.")
+                torch_dtype = torch.float16
         elif dtype == "float16" and config.model_type == "plamo2":
             logger.warning(
                 "For PLaMo2, using float16 is unstable and might cause "