[router][grpc] Fix proto3 default value mismatches and cleanup unused fields (#11283)

sgl-router/src/grpc_client/sglang_scheduler.rs

@@ -202,6 +202,7 @@ impl SglangSchedulerClient {
             stop: stop_sequences,
             stop_token_ids: request.stop_token_ids.clone().unwrap_or_default(),
             skip_special_tokens,
+            spaces_between_special_tokens: true, // Default from Python SamplingParams
             ignore_eos: request.ignore_eos,
             no_stop_trim: request.no_stop_trim,
             n: request.n.unwrap_or(1) as i32,
@@ -301,6 +302,8 @@ impl SglangSchedulerClient {
             top_k: -1,
             repetition_penalty: 1.0,
             n: 1,
+            skip_special_tokens: true,
+            spaces_between_special_tokens: true,
             ..Default::default()
         };
 
@@ -444,10 +447,24 @@ mod tests {
     #[test]
     fn test_sampling_params_defaults() {
         let params = proto::SamplingParams::default();
+        // Numeric fields have proto defaults (0)
         assert_eq!(params.temperature, 0.0);
-        assert_eq!(params.max_new_tokens, None);
         assert_eq!(params.top_p, 0.0);
         assert_eq!(params.top_k, 0);
+        assert_eq!(params.repetition_penalty, 0.0);
+        assert_eq!(params.n, 0);
+        // Bool fields have proto defaults (false)
+        assert!(!params.skip_special_tokens);
+        assert!(!params.spaces_between_special_tokens);
+        assert!(!params.ignore_eos);
+        assert!(!params.no_stop_trim);
+        // Optional int fields should be None
+        assert_eq!(params.max_new_tokens, None);
+        assert_eq!(params.stream_interval, None);
+        // Other non-optional fields
+        assert_eq!(params.min_p, 0.0);
+        assert_eq!(params.frequency_penalty, 0.0);
+        assert_eq!(params.presence_penalty, 0.0);
         assert!(params.stop.is_empty());
     }
 

sgl-router/src/proto/sglang_scheduler.proto

@@ -27,6 +27,11 @@ service SglangScheduler {
 // =====================
 
 // Sampling parameters matching SGLang's SamplingParams
+//
+// IMPORTANT: Do not use SamplingParams::default() directly!
+// The proto3 defaults (0 for numeric fields) do NOT match the semantic defaults
+// (temperature=1.0, top_p=1.0, top_k=-1, etc.). Always construct with explicit values
+// or use the conversion functions in sglang_scheduler.rs / grpc_server.py.
 message SamplingParams {
   float temperature = 1;
   float top_p = 2;
@@ -50,24 +55,18 @@ message SamplingParams {
     string structural_tag = 16;
   }
 
-  // LoRA adapter
-  string lora_path = 17;
-
   // Speculative decoding
-  int32 n = 18;  // Number of samples
-
-  // Token healing
-  bool token_healing = 19;
+  int32 n = 17;  // Number of samples
 
   // Additional parameters
-  int32 min_new_tokens = 20;
-  bool ignore_eos = 21;
-  bool no_stop_trim = 22;
-  int32 stream_interval = 23;
-  map<string, float> logit_bias = 24;
+  int32 min_new_tokens = 18;
+  bool ignore_eos = 19;
+  bool no_stop_trim = 20;
+  optional int32 stream_interval = 21;
+  map<string, float> logit_bias = 22;
 
   // Custom parameters for extensibility
-  google.protobuf.Struct custom_params = 25;
+  google.protobuf.Struct custom_params = 23;
 }