fix: reject CP shared KV with HiCache storage

python/sglang/srt/server_args.py

@@ -1586,6 +1586,11 @@ 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

@@ -351,7 +351,7 @@ class TestSSLArgs(unittest.TestCase):
         self.assertTrue(server_args.enable_ssl_refresh)
 
 
-class TestHiCacheArgs(unittest.TestCase):
+class TestHiCacheArgs(CustomTestCase):
     def _make_args(self, **overrides) -> ServerArgs:
         args = ServerArgs(model_path="dummy")
         for key, value in overrides.items():
@@ -371,6 +371,35 @@ class TestHiCacheArgs(unittest.TestCase):
         if expected_decode_backend is not None:
             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,
+        )
+
+        with self.assertRaisesRegex(
+            AssertionError,
+            "enable_nsa_prefill_cp_shared_kv.*hicache_storage_backend",
+        ):
+            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()
+
     def test_hicache_io_backend_and_mem_layout_compatibility(self):
         cases = [
             {