prain: fix vLLM on ROCm (no CUDA capability families), smoke test and benchmark scripts

patches/vllm-0.32-rocm-no-cuda-families.patch

@@ -0,0 +1,17 @@
+diff --git a/vllm/platforms/rocm.py b/vllm/platforms/rocm.py
+index 807e571..bbc642a 100644
+--- a/vllm/platforms/rocm.py
++++ b/vllm/platforms/rocm.py
+@@ -983,6 +983,12 @@ class RocmPlatform(Platform):
+         major, minor = torch.cuda.get_device_capability(device_id)
+         return DeviceCapability(major=major, minor=minor)
+ 
++    @classmethod
++    def is_device_capability_family(cls, capability: int, device_id: int = 0) -> bool:
++        # Families are a CUDA notion, and gfx majors collide with them (gfx1032 reads as
++        # 10.3, i.e. Blackwell sm_103): no CUDA family check may match a ROCm GPU.
++        return False
++
+     @classmethod
+     @with_amdsmi_context
+     def is_fully_connected(cls, physical_device_ids: list[int]) -> bool: