Add support for logit bias (#697)

llms/mlx_lm/server.py

@@ -128,6 +128,7 @@ class APIHandler(BaseHTTPRequestHandler):
         self.top_p = self.body.get("top_p", 1.0)
         self.repetition_penalty = self.body.get("repetition_penalty", 1.0)
         self.repetition_context_size = self.body.get("repetition_context_size", 20)
+        self.logit_bias = self.body.get("logit_bias", None)
 
         # Get stop id sequences, if provided
         stop_words = self.body.get("stop", [])
@@ -247,6 +248,7 @@ class APIHandler(BaseHTTPRequestHandler):
                 top_p=self.top_p,
                 repetition_penalty=self.repetition_penalty,
                 repetition_context_size=self.repetition_context_size,
+                logit_bias=self.logit_bias,
             ),
             range(self.max_tokens),
         ):

llms/mlx_lm/utils.py

@@ -117,6 +117,7 @@ def generate_step(
     repetition_penalty: Optional[float] = None,
     repetition_context_size: Optional[int] = 20,
     top_p: float = 1.0,
+    logit_bias: Optional[Dict[int, float]] = None,
 ) -> Generator[Tuple[mx.array, mx.array], None, None]:
     """
     A generator producing text based on the given prompt from the model.
@@ -135,6 +136,7 @@ def generate_step(
     """
 
     def sample(logits: mx.array) -> Tuple[mx.array, float]:
+        logits = logits + logit_bias if logit_bias else logits
         softmax_logits = mx.softmax(logits)
 
         if temp == 0:
@@ -203,6 +205,7 @@ def generate(
     repetition_penalty: Optional[float] = None,
     repetition_context_size: Optional[int] = None,
     top_p: float = 1.0,
+    logit_bias: Optional[Dict[int, float]] = None,
 ) -> str:
     """
     Generate text from the model.
@@ -241,6 +244,7 @@ def generate(
             repetition_penalty,
             repetition_context_size,
             top_p,
+            logit_bias,
         ),
         range(max_tokens),
     ):