feat: support fill in the middle templates

router/src/infer.rs

@@ -1,8 +1,8 @@
 /// Batching and inference logic
 use crate::validation::{Validation, ValidationError};
 use crate::{
-    ChatTemplateInputs, Entry, GenerateRequest, GenerateStreamResponse, HubTokenizerConfig,
-    Message, PrefillToken, Queue, Token,
+    ChatTemplateInputs, Entry, FillInMiddleInputs, GenerateRequest, GenerateStreamResponse,
+    HubTokenizerConfig, Message, PrefillToken, Queue, Token,
 };
 use futures::future::try_join_all;
 use minijinja::{Environment, ErrorKind, Template};
@@ -33,6 +33,8 @@ pub struct Infer {
     shared: Arc<Shared>,
     /// Chat template
     chat_template: Option<ChatTemplate>,
+    /// Fill in middle template
+    fill_in_middle_template: Option<FillInMiddleTemplate>,
     /// Inference limit
     limit_concurrent_requests: Arc<Semaphore>,
 }
@@ -88,6 +90,10 @@ impl Infer {
             .chat_template
             .map(|t| ChatTemplate::new(t, tokenizer_config.bos_token, tokenizer_config.eos_token));
 
+        let fill_in_middle_template = tokenizer_config
+            .fill_in_middle_template
+            .map(FillInMiddleTemplate::new);
+
         // Inference limit with a semaphore
         let semaphore = Arc::new(Semaphore::new(max_concurrent_requests));
 
@@ -96,6 +102,7 @@ impl Infer {
             queue,
             shared,
             chat_template,
+            fill_in_middle_template,
             limit_concurrent_requests: semaphore,
         }
     }
@@ -186,6 +193,25 @@ impl Infer {
             })
     }
 
+    /// Apply the fill in the middle template to the request
+    #[instrument(skip_all)]
+    pub(crate) fn apply_fill_in_middle_template(
+        &self,
+        prompt: String,
+        prefix: Option<String>,
+        suffix: Option<String>,
+    ) -> Result<String, InferError> {
+        self.fill_in_middle_template
+            .as_ref()
+            .ok_or_else(|| InferError::TemplateError(ErrorKind::TemplateNotFound.into()))?
+            .apply(prompt, prefix, suffix)
+            .map_err(|e| {
+                metrics::increment_counter!("tgi_request_failure", "err" => "template");
+                tracing::error!("{e}");
+                e
+            })
+    }
+
     /// Add a new request to the queue and return a InferResponse
     #[instrument(skip_all)]
     pub(crate) async fn generate(
@@ -342,6 +368,40 @@ impl ChatTemplate {
     }
 }
 
+#[derive(Clone)]
+struct FillInMiddleTemplate {
+    template: Template<'static, 'static>,
+}
+
+impl FillInMiddleTemplate {
+    fn new(template: String) -> Self {
+        let mut env = Box::new(Environment::new());
+        let template_str = template.into_boxed_str();
+        env.add_function("raise_exception", raise_exception);
+        // leaking env and template_str as read-only, static resources for performance.
+        let template = Box::leak(env)
+            .template_from_str(Box::leak(template_str))
+            .unwrap();
+
+        Self { template }
+    }
+
+    fn apply(
+        &self,
+        prompt: String,
+        prefix: Option<String>,
+        suffix: Option<String>,
+    ) -> Result<String, InferError> {
+        self.template
+            .render(FillInMiddleInputs {
+                prefix: prefix.as_deref(),
+                prompt: prompt.as_str(),
+                suffix: suffix.as_deref(),
+            })
+            .map_err(InferError::TemplateError)
+    }
+}
+
 /// Batching logic
 /// Will be launched in a background Tokio task
 ///

router/src/lib.rs

@@ -51,6 +51,7 @@ pub struct HubModelInfo {
 #[derive(Clone, Deserialize, Default)]
 pub struct HubTokenizerConfig {
     pub chat_template: Option<String>,
+    pub fill_in_middle_template: Option<String>,
     #[serde(deserialize_with = "token_serde::deserialize")]
     pub bos_token: Option<String>,
     #[serde(deserialize_with = "token_serde::deserialize")]
@@ -615,6 +616,13 @@ pub(crate) struct ChatTemplateInputs<'a> {
     add_generation_prompt: bool,
 }
 
+#[derive(Clone, Serialize, Deserialize)]
+pub(crate) struct FillInMiddleInputs<'a> {
+    prefix: Option<&'a str>,
+    prompt: &'a str,
+    suffix: Option<&'a str>,
+}
+
 #[derive(Clone, Deserialize, ToSchema, Serialize)]
 pub(crate) struct Message {
     #[schema(example = "user")]

router/src/server.rs

@@ -570,15 +570,28 @@ async fn completions(
     Json(req): Json<CompletionRequest>,
 ) -> Result<Response, (StatusCode, Json<ErrorResponse>)> {
     metrics::increment_counter!("tgi_request_count");
-
     let max_new_tokens = req.max_tokens.or(Some(100));
     let stream = req.stream.unwrap_or_default();
     let seed = req.seed;
-    let suffix = req.suffix.unwrap_or_default();
+
+    let inputs = match infer.apply_fill_in_middle_template(req.prompt, None, req.suffix) {
+        Ok(inputs) => inputs,
+        Err(err) => {
+            metrics::increment_counter!("tgi_request_failure", "err" => "validation");
+            tracing::error!("{err}");
+            return Err((
+                StatusCode::UNPROCESSABLE_ENTITY,
+                Json(ErrorResponse {
+                    error: err.to_string(),
+                    error_type: err.error_type().to_string(),
+                }),
+            ));
+        }
+    };
 
     // build the request passing some parameters
     let generate_request = GenerateRequest {
-        inputs: req.prompt.to_string(),
+        inputs: inputs.to_string(),
         parameters: GenerateParameters {
             best_of: None,
             temperature: req.temperature,
@@ -685,7 +698,7 @@ async fn completions(
                 finish_reason: details.finish_reason.to_string(),
                 index: 0,
                 logprobs: None,
-                text: generation.generated_text + &suffix,
+                text: generation.generated_text,
             }],
             usage: Usage {
                 prompt_tokens: details.prefill.len() as u32,