Custom all reduce fix mi250 (#247)

* disable custom all reduce when running on multiple MI250

* formatting

vllm/config.py

@@ -17,7 +17,7 @@
                                             get_hf_image_processor_config,
                                             get_hf_text_config)
 from vllm.utils import (GiB_bytes, cuda_device_count_stateless, get_cpu_memory,
-                        is_hip, is_neuron, is_openvino, is_xpu,
+                        is_hip, is_mi250, is_neuron, is_openvino, is_xpu,
                         print_warning_once)
 
 if TYPE_CHECKING:
@@ -949,6 +949,12 @@ def __init__(
         self._verify_args()
         self.rank: int = 0
 
+        if is_mi250() and self.tensor_parallel_size > 1:
+            self.disable_custom_all_reduce = True
+            logger.info(
+                "Disabled the custom all-reduce kernel because it is not "
+                "working correctly on multi AMD MI250.")
+
     @property
     def use_ray(self) -> bool:
         return self.distributed_executor_backend == "ray" or (

vllm/utils.py

@@ -487,6 +487,15 @@ def is_xpu() -> bool:
     return hasattr(torch, "xpu") and torch.xpu.is_available()
 
 
+@lru_cache(maxsize=None)
+def is_mi250() -> bool:
+    if not is_hip() or not torch.cuda.is_available():
+        return False
+    archName = torch.cuda.get_device_properties('cuda').gcnArchName
+    return (archName is not None) and \
+        ("gfx90a" in archName)
+
+
 @lru_cache(maxsize=None)
 def get_max_shared_memory_bytes(gpu: int = 0) -> int:
     """Returns the maximum shared memory per thread block in bytes."""