test: enforce early CP HiCache storage rejection

python/sglang/srt/server_args.py

@@ -1457,6 +1457,15 @@ class ServerArgs:
     def _handle_model_specific_adjustments(self):
         from sglang.srt.configs.model_config import is_deepseek_nsa
 
+        assert not (
+            self.enable_nsa_prefill_cp_shared_kv
+            and self.hicache_storage_backend is not None
+        ), (
+            "enable_nsa_prefill_cp_shared_kv does not support "
+            "hicache_storage_backend in the host-only CP HiCache stage. "
+            "Disable hicache_storage_backend or disable CP shared KV."
+        )
+
         if parse_connector_type(self.model_path) == ConnectorType.INSTANCE:
             return
 
@@ -1586,11 +1595,6 @@ class ServerArgs:
                         "Phase 2 shared KV initially supports NSATokenToKVPool only; "
                         "HiSparseNSATokenToKVPool is not wired yet."
                     )
-                    assert self.hicache_storage_backend is None, (
-                        "enable_nsa_prefill_cp_shared_kv does not support "
-                        "hicache_storage_backend in the host-only CP HiCache stage. "
-                        "Disable hicache_storage_backend or disable CP shared KV."
-                    )
                     if self.disaggregation_mode == "prefill":
                         assert envs.SGLANG_DISAGGREGATION_ALL_CP_RANKS_TRANSFER.get(), (
                             "Phase 2 shared KV with PD disaggregation requires "

test/registered/unit/server_args/test_server_args.py

@@ -383,22 +383,39 @@ class TestHiCacheArgs(CustomTestCase):
             hicache_storage_backend="mooncake",
             tp_size=8,
         )
+        pre_call_state = (
+            args.enable_dp_attention,
+            args.moe_dense_tp_size,
+            args.moe_a2a_backend,
+            args.ep_size,
+            args.attn_cp_size,
+            args.page_size,
+        )
 
-        with self.assertRaisesRegex(
-            AssertionError,
-            "enable_nsa_prefill_cp_shared_kv.*hicache_storage_backend",
+        with (
+            self.assertRaisesRegex(
+                AssertionError,
+                "enable_nsa_prefill_cp_shared_kv.*hicache_storage_backend",
+            ),
+            patch.object(
+                args,
+                "get_model_config",
+                side_effect=AssertionError("get_model_config should not be called"),
+            ),
         ):
-            model_config = MagicMock()
-            model_config.hf_config.architectures = ["DeepseekV3ForCausalLM"]
-            with (
-                patch.object(args, "get_model_config", return_value=model_config),
-                patch(
-                    "sglang.srt.configs.model_config.is_deepseek_nsa",
-                    return_value=True,
-                ),
-                patch("torch.cuda.get_device_capability", return_value=(9, 0)),
-            ):
-                args._handle_model_specific_adjustments()
+            args._handle_model_specific_adjustments()
+
+        self.assertEqual(
+            pre_call_state,
+            (
+                args.enable_dp_attention,
+                args.moe_dense_tp_size,
+                args.moe_a2a_backend,
+                args.ep_size,
+                args.attn_cp_size,
+                args.page_size,
+            ),
+        )
 
     def test_hicache_io_backend_and_mem_layout_compatibility(self):
         cases = [