feat: support logprobs in streaming and non streaming chat

router/src/lib.rs

@@ -213,7 +213,7 @@ pub(crate) struct ChatCompletionComplete {
     pub index: u32,
     pub message: Message,
     pub logprobs: Option<Vec<f32>>,
-    pub finish_reason: Option<String>,
+    pub finish_reason: String,
 }
 
 #[derive(Clone, Deserialize, Serialize)]
@@ -227,24 +227,26 @@ impl ChatCompletion {
     pub(crate) fn new(
         model: String,
         system_fingerprint: String,
-        ouput: String,
+        output: String,
         created: u64,
         details: Details,
+        return_logprobs: bool,
     ) -> Self {
         Self {
-            id: "".to_string(),
-            object: "text_completion".to_string(),
+            id: String::new(),
+            object: "text_completion".into(),
             created,
             model,
             system_fingerprint,
             choices: vec![ChatCompletionComplete {
                 index: 0,
                 message: Message {
-                    role: "assistant".to_string(),
-                    content: ouput,
+                    role: "assistant".into(),
+                    content: output,
                 },
-                logprobs: None,
-                finish_reason: details.finish_reason.to_string().into(),
+                logprobs: return_logprobs
+                    .then(|| details.tokens.iter().map(|t| t.logprob).collect()),
+                finish_reason: details.finish_reason.to_string(),
             }],
             usage: Usage {
                 prompt_tokens: details.prompt_token_count,
@@ -269,7 +271,7 @@ pub(crate) struct ChatCompletionChunk {
 pub(crate) struct ChatCompletionChoice {
     pub index: u32,
     pub delta: ChatCompletionDelta,
-    pub logprobs: Option<Vec<f32>>,
+    pub logprobs: Option<f32>,
     pub finish_reason: Option<String>,
 }
 
@@ -286,7 +288,7 @@ impl ChatCompletionChunk {
         delta: String,
         created: u64,
         index: u32,
-        logprobs: Option<Vec<f32>>,
+        logprobs: Option<f32>,
         finish_reason: Option<String>,
     ) -> Self {
         Self {
@@ -340,12 +342,10 @@ pub(crate) struct ChatRequest {
     #[serde(default)]
     pub logit_bias: Option<Vec<f32>>,
 
-    /// UNUSED
     /// Whether to return log probabilities of the output tokens or not. If true, returns the log probabilities of each
-    /// output token returned in the content of message. This option is currently not available on the gpt-4-vision-preview
-    /// model.
+    /// output token returned in the content of message. 
     #[serde(default)]
-    pub logprobs: Option<u32>,
+    pub logprobs: Option<bool>,
 
     /// UNUSED
     /// An integer between 0 and 5 specifying the number of most likely tokens to return at each token position, each with

router/src/server.rs

@@ -564,6 +564,7 @@ async fn chat_completions(
         .frequency_penalty
         // rescale frequency_penalty from (-2.0, 2.0) to (0.0, 4.0)
         .map(|x| x + 2.0);
+    let logprobs = req.logprobs.unwrap_or(false);
 
     // apply chat template to flatten the request into a single input
     let inputs = match infer.apply_chat_template(req) {
@@ -626,13 +627,16 @@ async fn chat_completions(
                     stream_token.token.text,
                     current_time,
                     stream_token.index,
-                    None,
+                    logprobs.then(|| stream_token.token.logprob),
                     stream_token.details.map(|d| d.finish_reason.to_string()),
                 ))
-                .unwrap_or_else(|e| {
-                    println!("Failed to serialize ChatCompletionChunk: {:?}", e);
-                    Event::default()
-                })
+                .map_or_else(
+                    |e| {
+                        println!("Failed to serialize ChatCompletionChunk: {:?}", e);
+                        Event::default()
+                    },
+                    |data| data,
+                )
         };
 
         let (headers, response_stream) =
@@ -655,6 +659,7 @@ async fn chat_completions(
             system_fingerprint,
             current_time,
             generation.details.unwrap(),
+            logprobs,
         );
 
         // wrap generation inside a Vec to match api-inference