perf: stop generation on default turn boundaries

crates/inference-server/src/compat.rs

@@ -514,7 +514,17 @@ pub fn build_sampler(req: &ChatCompletionRequest) -> inference_core::SamplerConf
     sampler.temperature = req.temperature.clamp(0.0, 2.0);
     sampler.top_p = req.top_p.clamp(0.0, 1.0);
     sampler.max_tokens = req.max_tokens.max(1);
-    sampler.stop_sequences = req.stop.clone();
+    sampler.stop_sequences = if req.stop.is_empty() {
+        vec![
+            "\nUser:".to_owned(),
+            "\nSystem:".to_owned(),
+            "\nDeveloper:".to_owned(),
+            "\nAssistant:".to_owned(),
+            "\nTool:".to_owned(),
+        ]
+    } else {
+        req.stop.clone()
+    };
     sampler.repeat_penalty = if req.tools.is_empty() { 1.05 } else { 1.0 };
     sampler.seed = req.seed.unwrap_or(0);
     sampler
@@ -565,8 +575,8 @@ pub fn ensure_supported_mode(req: &ChatCompletionRequest) -> Result<()> {
 #[cfg(test)]
 mod tests {
     use super::{
-        authorize_request, finalize_assistant_message, prepare_prompt, validate_request,
-        FinalizedAssistant, ModelCatalog, ServerConfig, StreamingMode,
+        authorize_request, build_sampler, finalize_assistant_message, prepare_prompt,
+        validate_request, FinalizedAssistant, ModelCatalog, ServerConfig, StreamingMode,
     };
     use api_contracts::chat::ChatCompletionRequest;
     use axum::http::{HeaderMap, HeaderValue, StatusCode};
@@ -718,4 +728,18 @@ mod tests {
             Some("Use references and lifetimes.")
         );
     }
+
+    #[test]
+    fn adds_default_turn_stops_when_request_has_none() {
+        let request = parse_request(
+            r#"{
+                "model": "alias-model",
+                "messages": [{"role":"user","content":"hi"}]
+            }"#,
+        );
+
+        let sampler = build_sampler(&request);
+        assert!(sampler.stop_sequences.contains(&"\nUser:".to_owned()));
+        assert!(sampler.stop_sequences.contains(&"\nAssistant:".to_owned()));
+    }
 }