fix: fail fast for CP HiCache storage args

python/sglang/srt/server_args.py

@@ -751,6 +751,9 @@ class ServerArgs:
         # Validate SSL arguments early (before dummy-model short-circuit).
         self._handle_ssl_validation()
 
+        # Validate CP shared KV constraints early (before dummy-model short-circuit).
+        self._handle_cp_shared_kv_validation()
+
         if self.model_path.lower() in ["none", "dummy"]:
             # Skip for dummy models
             return
@@ -904,6 +907,16 @@ class ServerArgs:
                 "to be specified."
             )
 
+    def _handle_cp_shared_kv_validation(self):
+        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."
+        )
+
     def _handle_deprecated_args(self):
         # Handle deprecated tool call parsers
         deprecated_tool_call_parsers = {"qwen25": "qwen", "glm45": "glm"}
@@ -1457,15 +1470,6 @@ 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
 

test/registered/unit/server_args/test_server_args.py

@@ -372,50 +372,28 @@ class TestHiCacheArgs(CustomTestCase):
             self.assertEqual(args.decode_attention_backend, expected_decode_backend)
 
     def test_cp_shared_kv_rejects_hicache_storage_backend(self):
-        args = self._make_args(
-            enable_hierarchical_cache=True,
-            enable_nsa_prefill_context_parallel=True,
-            enable_nsa_prefill_cp_shared_kv=True,
-            nsa_prefill_cp_mode="in-seq-split",
-            disaggregation_mode=None,
-            page_size=64,
-            enable_hisparse=False,
-            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",
             ),
             patch.object(
-                args,
+                ServerArgs,
                 "get_model_config",
                 side_effect=AssertionError("get_model_config should not be called"),
             ),
         ):
-            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,
-            ),
-        )
+            ServerArgs(
+                model_path="dummy",
+                enable_hierarchical_cache=True,
+                enable_nsa_prefill_context_parallel=True,
+                enable_nsa_prefill_cp_shared_kv=True,
+                nsa_prefill_cp_mode="in-seq-split",
+                disaggregation_mode="null",
+                page_size=64,
+                enable_hisparse=False,
+                hicache_storage_backend="mooncake",
+            )
 
     def test_hicache_io_backend_and_mem_layout_compatibility(self):
         cases = [