[diffusion] multi-platform: support diffusion on amd and fix encoder loading on MI325 (#13760)

Co-authored-by: Sabre Shao <sabre.shao@amd.com>
Co-authored-by: Yusheng (Ethan) Su <yushengsu.thu@gmail.com>
Co-authored-by: Hubert Lu <Hubert.Lu@amd.com>
Co-authored-by: xsun <sunxiao04@gmail.com>

python/sglang/multimodal_gen/runtime/server_args.py

@@ -945,6 +945,8 @@ class ServerArgs:
             raise ValueError("pipeline_config is not set in ServerArgs")
 
         self.pipeline_config.check_pipeline_config()
+        if self.attention_backend is None:
+            self._set_default_attention_backend()
 
         # parallelism
         self.check_server_dp_args()
@@ -966,6 +968,17 @@ class ServerArgs:
                     "Please use either sequence parallelism or tensor parallelism, not both."
                 )
 
+    def _set_default_attention_backend(self) -> None:
+        """Configure ROCm defaults when users do not specify an attention backend."""
+        if current_platform.is_rocm():
+            default_backend = AttentionBackendEnum.AITER.name.lower()
+            self.attention_backend = default_backend
+            logger.info(
+                "Attention backend not specified. Using '%s' by default on ROCm "
+                "to match SGLang SRT defaults.",
+                default_backend,
+            )
+
 
 @dataclasses.dataclass
 class PortArgs: