chore: pass presence_penalty field to completion options (#2280)

crates/http-api-bindings/src/completion/llama.rs

@@ -30,6 +30,7 @@ struct CompletionRequest {
     temperature: f32,
     stream: bool,
     penalty_last_n: i32,
+    presence_penalty: f32,
 }
 
 #[derive(Deserialize)]
@@ -47,6 +48,7 @@ impl CompletionStream for LlamaCppEngine {
             temperature: options.sampling_temperature,
             stream: true,
             penalty_last_n: 0,
+            presence_penalty: options.presence_penalty,
         };
 
         let mut request = self.client.post(&self.api_endpoint).json(&request);

crates/ollama-api-bindings/src/completion.rs

@@ -23,10 +23,12 @@ pub struct OllamaCompletion {
 #[async_trait]
 impl CompletionStream for OllamaCompletion {
     async fn generate(&self, prompt: &str, options: CompletionOptions) -> BoxStream<String> {
+        // FIXME: options.presence_penalty is not used
         let ollama_options = GenerationOptions::default()
             .num_ctx(options.max_input_length as u32)
             .num_predict(options.max_decoding_tokens)
             .seed(options.seed as i32)
+            .repeat_last_n(0)
             .temperature(options.sampling_temperature);
         let request = GenerationRequest::new(self.model.to_owned(), prompt.to_owned())
             .template("{{ .Prompt }}".to_string())

crates/tabby-inference/src/completion.rs

@@ -11,6 +11,9 @@ pub struct CompletionOptions {
     pub sampling_temperature: f32,
 
     pub seed: u64,
+
+    #[builder(default = "0.0")]
+    pub presence_penalty: f32,
 }
 
 #[async_trait]

crates/tabby/src/services/model/chat.rs

@@ -53,6 +53,7 @@ impl ChatCompletionStream for ChatCompletionImpl {
             .seed(options.seed)
             .max_decoding_tokens(options.max_decoding_tokens)
             .sampling_temperature(options.sampling_temperature)
+            .presence_penalty(options.presence_penalty)
             .build()?;
 
         let prompt = self.prompt_builder.build(messages)?;