feat(ollama): keep models in VRAM for 1h to prevent cold starts

server/src/ollama.rs

@@ -25,6 +25,8 @@ struct GenerateRequest<'a> {
     images: Option<Vec<&'a str>>,
     #[serde(skip_serializing_if = "Option::is_none")]
     options: Option<serde_json::Value>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    keep_alive: Option<&'a str>,
 }
 
 #[derive(Deserialize)]
@@ -36,6 +38,8 @@ struct GenerateResponse {
 struct EmbeddingRequest<'a> {
     model: &'a str,
     prompt: &'a str,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    keep_alive: Option<&'a str>,
 }
 
 #[derive(Deserialize)]
@@ -123,6 +127,7 @@ impl OllamaClient {
                 "num_ctx": 32768, // Massive context window win
                 "num_predict": 4096 // Give reasoning models plenty of output room
             })),
+            keep_alive: Some("1h"),
         };
 
         let res = self
@@ -166,6 +171,7 @@ impl OllamaClient {
                 "num_ctx": 8192,
                 "num_predict": 1024
             })),
+            keep_alive: Some("1h"),
         };
 
         let res = self
@@ -196,6 +202,7 @@ impl OllamaClient {
         let body = EmbeddingRequest {
             model: &self.embed_model,
             prompt: text,
+            keep_alive: Some("1h"),
         };
 
         let res = self