[Fix] DCP: advertise the logical KV-event block size (#35298)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

python/sglang/srt/server_args.py

@@ -9669,6 +9669,13 @@ class ServerArgs:
         ``--speculative-draft-load-format`` needs its own transfer engine."""
         return remote_instance_transfer_engine_of(self, load_format)
 
+    @property
+    def kv_event_block_size(self) -> int:
+        """Width KV events are emitted at: under DCP the radix tree pages at
+        ``page_size * dcp_size`` (``mem_cache/kv_cache_builder.py``).
+        """
+        return self.page_size * self.dcp_size
+
     def describe_kv_events_publisher(self) -> Optional[dict]:
         """Return a structured description of this server's KV-event
         publisher, or `None` if publishing is disabled / misconfigured.
@@ -9694,10 +9701,13 @@ class ServerArgs:
                 "topic": "",                      # ZMQ topic prefix on the
                                                   # SUB filter (empty =
                                                   # subscribe-all)
-                "block_size": <page_size>,        # subscribers MUST hash
+                "block_size": <kv_event_block_size>,  # subscribers MUST
-                                                  # prompts at this size
+                                                  # hash prompts at this size
-                "dp_size": <dp_size>,             # number of SUB sockets
+                "dp_size": <dp_size>,             # number of SUB sockets to
-                                                  # to open
+                                                  # open; not DCP-scaled, as
+                                                  # DCP shards within a rank
+                                                  # rather than adding
+                                                  # publishers
             }
 
         Returns None (i.e. "no publisher to describe") when any of:
@@ -9752,7 +9762,7 @@ class ServerArgs:
             "endpoint_host": host,
             "endpoint_port_base": port,
             "topic": cfg.topic,
-            "block_size": page_size,
+            "block_size": self.kv_event_block_size,
             "dp_size": self.dp_size,
         }
 

test/registered/unit/server_args/test_server_args.py

@@ -2212,5 +2212,35 @@ class TestTwoBatchOverlapBackend(CustomTestCase):
         args._check_two_batch_overlap()
 
 
+class TestDcpKvEventContract(CustomTestCase):
+    """DCP widens the radix-tree page to page_size * dcp_size, which the
+    advertised KV-event block size must reflect."""
+
+    KV_EVENTS = '{"publisher":"zmq","topic":"kv","endpoint":"tcp://*:5557"}'
+
+    def test_kv_events_descriptor_reports_logical_block_size(self):
+        """Advertising the physical page_size made every KV-aware router hash
+        prompts at a width no emitted block can match, silently pinning its
+        hit rate to zero while stores kept applying cleanly."""
+        args = ServerArgs(
+            model_path="dummy",
+            tp_size=4,
+            dcp_size=4,
+            page_size=64,
+            kv_events_config=self.KV_EVENTS,
+        )
+        self.assertEqual(args.describe_kv_events_publisher()["block_size"], 256)
+        args = ServerArgs(
+            model_path="dummy", page_size=64, kv_events_config=self.KV_EVENTS
+        )
+        self.assertEqual(args.describe_kv_events_publisher()["block_size"], 64)
+
+    def test_kv_event_block_size_widens_a_single_token_page(self):
+        # page_size=1 + DCP is a real deployment shape: the allocator is still
+        # paged, at dcp_size.
+        args = ServerArgs(model_path="dummy", tp_size=8, dcp_size=8, page_size=1)
+        self.assertEqual(args.kv_event_block_size, 8)
+
+
 if __name__ == "__main__":
     unittest.main()