[rust][server] Add generative scoring route

rust/src/engine-core-client/src/mock_engine.rs

@@ -73,6 +73,7 @@ pub fn default_ready_response() -> EngineCoreReadyResponse {
         enable_sleep_mode: false,
         supports_draft_weight_updates: false,
         effective_attention_block_size: None,
+        use_spec_decode: false,
     }
 }
 

rust/src/engine-core-client/src/protocol/handshake.rs

@@ -107,6 +107,9 @@ pub struct EngineCoreReadyResponse {
     /// Full-attention block size in tokens after initialization, or unavailable.
     #[serde(default)]
     pub effective_attention_block_size: Option<u64>,
+    /// Whether the engine was started with speculative decoding enabled.
+    #[serde(default)]
+    pub use_spec_decode: bool,
 }
 
 /// Frontend-owned ZMQ addresses that are sent to the engine during startup

rust/src/engine-core-client/src/tests/client.rs

@@ -2968,6 +2968,7 @@ fn python_msgpack_fixtures_match_rust_encoding() {
     assert!(ready_response.enable_sleep_mode);
     assert!(ready_response.supports_draft_weight_updates);
     assert_eq!(ready_response.effective_attention_block_size, Some(64));
+    assert!(ready_response.use_spec_decode);
     let kv_events_config = ready_response.kv_events_config.expect("KV events config should decode");
     assert!(kv_events_config.enable_kv_cache_events);
     assert_eq!(kv_events_config.publisher, "zmq");

rust/src/engine-core-client/src/tests/python_compat.py

@@ -470,6 +470,7 @@ class EngineCoreReadyResponse:
     enable_sleep_mode: bool = False
     supports_draft_weight_updates: bool = False
     effective_attention_block_size: int | None = None
+    use_spec_decode: bool = False
 
 
 ready_response = EngineCoreReadyResponse(
@@ -494,6 +495,7 @@ class EngineCoreReadyResponse:
     enable_sleep_mode=True,
     supports_draft_weight_updates=True,
     effective_attention_block_size=64,
+    use_spec_decode=True,
     kv_events_config=KVEventsConfig(
         enable_kv_cache_events=True,
         publisher="zmq",

rust/src/server/src/routes.rs

@@ -100,7 +100,8 @@ fn build_router_with_options(
         .route("/v1/chat/completions", post(openai::chat_completions))
         // vLLM specific endpoints
         .route("/tokenize", post(tokenize::tokenize))
-        .route("/detokenize", post(tokenize::detokenize));
+        .route("/detokenize", post(tokenize::detokenize))
+        .route("/generative_scoring", post(inference::generative_scoring));
 
     if scale_out_endpoints_enabled {
         router = router.route("/inference/v1/generate", post(inference::generate));