[Distributed] [ROCM] Fix custom allreduce enable checks (#16010)

Signed-off-by: ilmarkov <imarkov@redhat.com>
Co-authored-by: ilmarkov <imarkov@redhat.com>

vllm/config.py

@@ -1619,13 +1619,12 @@ class ParallelConfig:
         if self.use_ray:
             from vllm.executor import ray_utils
             ray_utils.assert_ray_available()
-        device_capability = current_platform.get_device_capability()
+
-        if (current_platform.is_rocm() and device_capability is not None
+        if not current_platform.use_custom_allreduce():
-                and device_capability < (9, 4)):
             self.disable_custom_all_reduce = True
             logger.info(
                 "Disabled the custom all-reduce kernel because it is not "
-                "supported on AMD GPUs older than MI300X.")
+                "supported on current platform.")
         if self.ray_workers_use_nsight and not self.use_ray:
             raise ValueError("Unable to use nsight profiling unless workers "
                              "run with Ray.")

vllm/platforms/cuda.py

@@ -308,6 +308,10 @@ class CudaPlatformBase(Platform):
     def supports_v1(cls, model_config: ModelConfig) -> bool:
         return True
 
+    @classmethod
+    def use_custom_allreduce(cls) -> bool:
+        return True
+
 
 # NVML utils
 # Note that NVML is not affected by `CUDA_VISIBLE_DEVICES`,

vllm/platforms/interface.py

@@ -379,6 +379,13 @@ class Platform:
         """
         return False
 
+    @classmethod
+    def use_custom_allreduce(cls) -> bool:
+        """
+        Returns if custom allreduce is supported on the current platform
+        """
+        return False
+
 
 class UnspecifiedPlatform(Platform):
     _enum = PlatformEnum.UNSPECIFIED

vllm/platforms/rocm.py

@@ -302,3 +302,10 @@ class RocmPlatform(Platform):
     def supports_v1(cls, model_config: ModelConfig) -> bool:
         # V1 support on AMD gpus is experimental
         return True
+
+    @classmethod
+    def use_custom_allreduce(cls) -> bool:
+        # We only enable custom allreduce for MI300 series
+        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
+        supported_archs = ['gfx94']
+        return any(gfx in gcn_arch for gfx in supported_archs)