Rust server: align launcher and request validation behavior (#37327)

python/sglang/srt/managers/scheduler.py

@@ -2171,6 +2171,9 @@ class Scheduler(
         # Park the idle loop on the request ring within the rank-0 rust-server
         self.idle_sleeper = RustServerIdleSleeper(rust_server)
 
+    def rust_server_tokenizer_path(self) -> str:
+        return get_serving().tokenizer_path
+
     def init_request_receiver(self) -> None:
         self.request_receiver = SchedulerRequestReceiver(
             recv_from_tokenizer=self.recv_from_tokenizer,

python/sglang/srt/rust_server/config.py

@@ -46,7 +46,7 @@ def _build_server_args(scheduler: Scheduler) -> ServerArgs:
     return ext.ServerArgs(
         model_path=get_model().model_path,
         served_model_name=get_serving().served_model_name,
-        tokenizer_path=get_serving().tokenizer_path,
+        tokenizer_path=scheduler.rust_server_tokenizer_path(),
         revision=get_model().revision,
         load_format=get_model().load_format,
         weight_version=get_serving().weight_version,
@@ -76,8 +76,6 @@ def _build_server_args(scheduler: Scheduler) -> ServerArgs:
                 **mc.get_default_sampling_params()
             ),
         ),
-        # `preferred_sampling_params` is deliberately absent: `launch`
-        # refuses to start when it is set, so the Rust server never needs it.
         preferred_sampling_params=(
             json.dumps(get_serving().preferred_sampling_params)
             if get_serving().preferred_sampling_params is not None

python/sglang/srt/rust_server/server.py

@@ -35,6 +35,7 @@ from sglang.srt.utils.flatten import (
     NestedRowColumns,
     RaggedPairColumns,
 )
+from sglang.srt.utils.network import NetworkAddress
 
 if TYPE_CHECKING:
     from sglang.srt.managers.io_struct import BatchTokenIDOutput
@@ -76,24 +77,11 @@ class RustServer:
         os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
 
         server_args = scheduler.server_args
-        # `TokenizerManager` merges these under each request's own sampling params
-        # (`{**preferred, **obj.sampling_params}`), and this server replaces that
-        # manager wholesale — so honouring the flag is not implemented here yet.
-        # Refuse rather than run: silently dropping it means generating with
-        # sampling the operator did not configure, and `/get_model_info` would go on
-        # advertising values no request ever receives.
-        if get_serving().preferred_sampling_params:
-            raise ValueError(
-                "SGLANG_RUST_SERVER does not yet apply --preferred-sampling-params "
-                "(the Python TokenizerManager merges it into every request; the rust "
-                "ingress has no equivalent). Launch without SGLANG_RUST_SERVER, or "
-                "drop --preferred-sampling-params and send those values per request."
-            )
         # Per-DP-rank HTTP port with client load balancing. `None` when DP is off,
         # so the rank is not conflated with rank 0 of a one-rank group.
         dp_rank = scheduler.ps.attn_dp_rank if scheduler.ps.dp_size > 1 else None
         listen_port = get_serving().port + (dp_rank or 0)
-        listen_addr = f"{get_serving().host}:{listen_port}"
+        listen_addr = NetworkAddress(get_serving().host, listen_port).to_host_port_str()
 
         launch_cores, server_cores = _partition_cores(
             mm_workers=(