[Cleanup] Move mamba-max-states-per-path validation into _handle_mamba_backend (#32502)

docs_new/cookbook/autoregressive/Moonshotai/Kimi-K3.mdx

@@ -36,7 +36,7 @@ If you do not want to use a Docker image, reproduce the dependency installation
 
 Pick your hardware, then the deployment shape and operating point. Node count follows the hardware recipe (B200 2×8, GB200 4×4, H100 4×8, B300 1×8, H200 2×8, GB300 2×4, MI350X/MI355X 1×8), so it is not a separate choice.
 
-**PD Mode** — `Unified` serves prefill and decode together. `Prefill` / `Decode` split them into dedicated pools (see [PD disaggregation](#pd-disaggregation)); `Prefill` ships two strategies on the TP8 platforms, both chunked at 16k: `Default` (TP8) and `Long-Context` (`--pp-size 8 --tp-size 1`, see [Deep PP](#deep-pp-for-long-context-prefill)).
+**PD Mode** — `Unified` serves prefill and decode together. `Prefill` / `Decode` split them into dedicated pools (see [PD disaggregation](#3-4-pd-disaggregation)); `Prefill` ships two strategies on the TP8 platforms, both chunked at 16k: `Default` (TP8) and `Long-Context` (`--pp-size 8 --tp-size 1`, see [Deep PP](#deep-pp-for-long-context-prefill)).
 
 **Strategy** — the operating point within that shape:
 

python/sglang/srt/server_args.py

@@ -3409,8 +3409,6 @@ class ServerArgs:
         # _handle_model_specific_adjustments never runs.
         self._resolved_overrides = []
 
-        self._validate_mamba_max_states_per_path()
-
         if self.model_path.lower() in ["none", "dummy"]:
             return
 
@@ -3588,14 +3586,6 @@ class ServerArgs:
 
         materialize_declarations(self)
 
-    def _validate_mamba_max_states_per_path(self):
-        value = self.mamba_max_states_per_path
-        if value == 0 or value < -1:
-            raise ValueError(
-                "--mamba-max-states-per-path must be -1 (unlimited) or a positive "
-                f"integer, got {value}."
-            )
-
     def _handle_model_capability_adjustments(self):
         if parse_connector_type(self.model_path) == ConnectorType.INSTANCE:
             return
@@ -5812,6 +5802,12 @@ class ServerArgs:
         if self.mamba_cache_philox_rounds < 0:
             raise ValueError("--mamba-cache-philox-rounds must be non-negative.")
 
+        if self.mamba_max_states_per_path == 0 or self.mamba_max_states_per_path < -1:
+            raise ValueError(
+                "--mamba-max-states-per-path must be -1 (unlimited) or a positive "
+                f"integer, got {self.mamba_max_states_per_path}."
+            )
+
         if self.enable_mamba_cache_stochastic_rounding:
             if self.mamba_ssm_dtype != "float16":
                 raise ValueError(

test/registered/unit/mem_cache/test_mamba_path_state_cap.py

@@ -103,14 +103,12 @@ class TestMambaPathStateCap(unittest.TestCase):
 
     def test_server_arg_rejects_zero_and_values_below_negative_one(self):
         for value in (0, -2):
+            args = ServerArgs(model_path="dummy", mamba_max_states_per_path=value)
             with self.subTest(value=value), self.assertRaisesRegex(
                 ValueError,
                 "must be -1 \\(unlimited\\) or a positive integer",
             ):
-                ServerArgs(
+                args._handle_mamba_backend()
-                    model_path="dummy",
-                    mamba_max_states_per_path=value,
-                )
 
     def test_unified_cache_removes_only_shallow_mamba_state(self):
         component, nodes, core, cache = _build_unified_chain(cap=2)