Sync fork

Author: matthewhaynesonline

mistralrs-server-core/src/chat_completion.rs

@@ -327,9 +327,16 @@ async fn parse_request(
 
                     images.push(image);
                 }
-                RequestMessage::VisionChat { messages, images }
+                RequestMessage::VisionChat {
+                    messages,
+                    images,
+                    enable_thinking: oairequest.enable_thinking,
+                }
             } else {
-                RequestMessage::Chat(messages)
+                RequestMessage::Chat {
+                    messages,
+                    enable_thinking: oairequest.enable_thinking,
+                }
             }
         }
         Either::Right(prompt) => {
@@ -339,7 +346,10 @@ async fn parse_request(
             message_map.insert("role".to_string(), Either::Left("user".to_string()));
             message_map.insert("content".to_string(), Either::Left(prompt));
             messages.push(message_map);
-            RequestMessage::Chat(messages)
+            RequestMessage::Chat {
+                messages,
+                enable_thinking: oairequest.enable_thinking,
+            }
         }
     };
 

mistralrs-server-core/src/lib.rs

@@ -161,17 +161,17 @@ pub struct Args {
     #[arg(long)]
     pub cpu: bool,
 
-    /// Enable web searching for interactive mode.
-    #[arg(long = "interactive-search")]
-    pub interactive_search: bool,
-
     /// Enable searching compatible with the OpenAI `web_search_options` setting. This uses the BERT model specified below or the default.
     #[arg(long = "enable-search")]
     pub enable_search: bool,
 
     /// Specify a Hugging Face model ID for a BERT model to assist web searching. Defaults to Snowflake Arctic Embed L.
     #[arg(long = "search-bert-model")]
     pub search_bert_model: Option<String>,
+
+    /// Enable thinking for interactive mode and models that support it.
+    #[arg(long = "enable-thinking")]
+    pub enable_thinking: bool,
 }
 
 fn parse_token_source(s: &str) -> Result<TokenSource, String> {
@@ -345,7 +345,7 @@ pub async fn bootstrap_mistralrs_router(
         pipeline,
         scheduler_config,
         args.interactive_mode,
-        bert_model,
+        bert_model.clone(),
         args.log,
         args.truncate_sequence,
         args.no_kv_cache,

mistralrs-server-core/src/openai.rs

@@ -173,6 +173,8 @@ pub struct ChatCompletionRequest {
     pub dry_allowed_length: Option<usize>,
     #[schema(example = json!(Option::None::<String>))]
     pub dry_sequence_breakers: Option<Vec<String>>,
+    #[schema(example = json!(Option::None::<bool>))]
+    pub enable_thinking: Option<bool>,
 }
 
 #[derive(Debug, Serialize, ToSchema)]

mistralrs-server-core/src/util.rs

@@ -112,8 +112,6 @@ mod tests {
         xjApU46pnBe8fwF4pb+/8Ywv/DK9zbCKsfWXUBhf+A1dOX00S+xfgc3L3dmKWSn7iklDjthxbSaH
         c7YCVIAfi6JYn5bHjTHTGmurQJXJ8C/um928G9zK4gAAAABJRU5ErkJggg==
         ";
-        let image = parse_image_url(url).await.unwrap();
-        assert_eq!(image.dimensions(), (32, 32));
 
         let url = format!("data:image/png;base64,{}", url);
         let image = parse_image_url(&url).await.unwrap();