fix(llm): only feed prompt if not empty

crates/llm-base/src/inference_session.rs

@@ -438,14 +438,16 @@ impl InferenceSession {
         let parameters = request.parameters;
 
         // Feed the initial prompt through the transformer, to update its
-        // context window with new data.
+        // context window with new data, if necessary.
+        if !request.prompt.is_empty() {
         self.feed_prompt(
             model,
             parameters,
             request.prompt,
             output_request,
             feed_prompt_callback(&mut callback),
         )?;
+        }
         stats.feed_prompt_duration = start_at.elapsed().unwrap();
         stats.prompt_tokens = self.n_past;
 

crates/llm-base/src/tokenizer/mod.rs

@@ -230,6 +230,14 @@ impl Prompt<'_> {
             }
         })
     }
+
+    /// Returns whether this prompt is empty.
+    pub fn is_empty(&self) -> bool {
+        match self {
+            Self::Text(text) => text.is_empty(),
+            Self::Tokens(tokens) => tokens.is_empty(),
+        }
+    }
 }
 impl<'a> Default for Prompt<'a> {
     fn default() -> Self {