feat: support do_sample param in ChatRequest

router/src/lib.rs

@@ -709,6 +709,10 @@ pub(crate) struct ChatRequest {
     #[schema(nullable = true, example = "null")]
     #[serde(deserialize_with = "deserialize_tool_choice::deserialize")]
     pub tool_choice: Option<ToolType>,
+
+    #[serde(default)]
+    #[schema(default = "false", example = true)]
+    pub do_sample: bool,
 }
 
 fn default_tool_prompt() -> Option<String> {

router/src/server.rs

@@ -1060,7 +1060,7 @@ async fn chat_completions(
             top_k: None,
             top_p: req.top_p,
             typical_p: None,
-            do_sample: true,
+            do_sample: req.do_sample,
             max_new_tokens,
             return_full_text: None,
             stop,