[rust-server] fix p/d bootstrap across dp listeners (#36234)

python/sglang/srt/disaggregation/base/conn.py

@@ -74,6 +74,8 @@ class KVArgs:
     page_size: int
     # for system dp
     system_dp_rank: int
+    # Local Rust /route registry port; None on scheduler ranks without a listener.
+    rust_http_port: Optional[int]
     # for pp prefill
     pp_rank: int
     prefill_start_layer: int

python/sglang/srt/disaggregation/common/conn.py

@@ -828,6 +828,28 @@ class CommonKVManager(BaseKVManager):
             "prefill_http_port": get_serving().port,
         }
 
+        if envs.SGLANG_RUST_SERVER.get() and self.attn_dp_size > 1:
+            topology_rows = get_world_group().all_gather_object(payload)
+            # Every scheduler contributes a topology row. Only the scheduler
+            # ranks that own a Rust listener populate their local registry.
+            if self.kv_args.rust_http_port is None:
+                return
+            registry_host = {"0.0.0.0": "127.0.0.1", "::": "::1"}.get(
+                self.bootstrap_host, self.bootstrap_host
+            )
+            registry = NetworkAddress(registry_host, self.kv_args.rust_http_port)
+            url = f"{registry.to_url()}/route"
+            for topology_row in topology_rows:
+                registry_row = {
+                    **topology_row,
+                    "prefill_http_port": self.kv_args.rust_http_port,
+                }
+                self._register_topology_row(url, registry_row)
+            return
+
+        self._register_topology_row(url, payload)
+
+    def _register_topology_row(self, url: str, payload: Dict) -> None:
         max_retries, initial_delay, max_delay = 5, 1.0, 30.0
         for attempt in range(max_retries):
             try:

python/sglang/srt/disaggregation/prefill.py

@@ -208,6 +208,11 @@ class PrefillBootstrapQueue:
         kv_args.engine_rank = self.tp_rank
         kv_args.pp_rank = self.pp_rank
         kv_args.system_dp_rank = self.scheduler.ps.dp_rank
+        kv_args.rust_http_port = (
+            self.scheduler.rust_server.http_port
+            if self.scheduler.rust_server is not None
+            else None
+        )
         kv_args.kv_cache_dtype_str = (
             self.scheduler.tp_worker.model_runner.kv_cache_dtype_str
         )

python/sglang/srt/rust_server/server.py

@@ -55,10 +55,12 @@ class RustServer:
     def __init__(
         self,
         server: Server,
+        http_port: int,
         mm_spec: Optional[RustMmSpec] = None,
         max_per_poll: int = 256,
     ):
         self.server = server
+        self.http_port = http_port
         self.mm_spec = mm_spec
         self._max_per_poll = max_per_poll
 
@@ -151,7 +153,7 @@ class RustServer:
             dp_note,
         )
 
-        return cls(server, mm_spec=mm_spec)
+        return cls(server, http_port=listen_port, mm_spec=mm_spec)
 
     def wait_request(self, timeout_ms: int) -> None:
         """Block until a request is pushed into the in-process ring or the timeout