[torch.compile] support all attention backends (#10558)

Signed-off-by: youkaichao <youkaichao@gmail.com>

vllm/platforms/rocm.py

@@ -36,6 +36,7 @@ if os.environ.get("VLLM_WORKER_MULTIPROC_METHOD", None) in ["fork", None]:
 class RocmPlatform(Platform):
     _enum = PlatformEnum.ROCM
     device_type: str = "cuda"
+    dispatch_key: str = "CUDA"
 
     @classmethod
     def get_default_attn_backend(cls, selected_backend: _Backend) -> _Backend: