[AMD] Use tilelang as default NSA attention backend dispatch on AMD Instinct (#18319)

python/sglang/srt/layers/attention/nsa_backend.py

@@ -1477,7 +1477,9 @@ class NativeSparseAttnBackend(
                 page_size=1,
             )
         else:
-            raise ValueError(f"Unsupported {nsa_impl = }")
+            raise ValueError(
+                f"Unsupported {nsa_impl = } for forward_extend. Consider using an other attention backend."
+            )
 
     def forward_decode(
         self,

python/sglang/srt/server_args.py

@@ -1150,7 +1150,10 @@ class ServerArgs:
         user_set_prefill = self.nsa_prefill_backend is not None
         user_set_decode = self.nsa_decode_backend is not None
 
-        if kv_cache_dtype == "fp8_e4m3":
+        if not user_set_prefill and not user_set_decode and is_hip():
+            self.nsa_prefill_backend = "tilelang"
+            self.nsa_decode_backend = "tilelang"
+        elif kv_cache_dtype == "fp8_e4m3":
             # flashmla_auto dispatches to flashmla_sparse/flashmla_kv based on hardware and heuristics
             if not user_set_prefill:
                 self.nsa_prefill_backend = "flashmla_auto"