Add openai frequency and presence penalty parameters. Closes #169

llama_cpp/llama.py

@@ -261,7 +261,7 @@ class Llama:
             ]
             self.eval_logits.extend(logits)
 
-    def _sample_top_p_top_k(
+    def _sample(
         self,
         last_n_tokens_data,  # type: llama_cpp.Array[llama_cpp.llama_token]
         last_n_tokens_size: llama_cpp.c_int,
@@ -269,6 +269,8 @@ class Llama:
         top_p: llama_cpp.c_float,
         temp: llama_cpp.c_float,
         repeat_penalty: llama_cpp.c_float,
+        frequency_penalty: llama_cpp.c_float,
+        presence_penalty: llama_cpp.c_float,
     ):
         assert self.ctx is not None
         assert len(self.eval_logits) > 0
@@ -298,6 +300,14 @@ class Llama:
             candidates=llama_cpp.ctypes.byref(candidates),  # type: ignore
             penalty=repeat_penalty,
         )
+        llama_cpp.llama_sample_frequency_and_presence_penalties(
+            ctx=self.ctx,
+            candidates=llama_cpp.ctypes.byref(candidates),  # type: ignore
+            last_tokens_data=last_n_tokens_data,
+            last_tokens_size=last_n_tokens_size,
+            alpha_frequency=frequency_penalty,
+            alpha_presence=presence_penalty,
+        )
         if float(temp.value) == 0.0:
             return llama_cpp.llama_sample_token_greedy(
                 ctx=self.ctx,
@@ -344,6 +354,8 @@ class Llama:
         top_p: float,
         temp: float,
         repeat_penalty: float,
+        frequency_penalty: float = 0.0,
+        presence_penalty: float = 0.0,
     ):
         """Sample a token from the model.
 
@@ -360,7 +372,7 @@ class Llama:
         last_n_tokens_data = [llama_cpp.llama_token(0)] * max(
             0, self.last_n_tokens_size - len(self.eval_tokens)
         ) + list(self.eval_tokens)[-self.last_n_tokens_size :]
-        return self._sample_top_p_top_k(
+        return self._sample(
             last_n_tokens_data=(llama_cpp.llama_token * self.last_n_tokens_size)(
                 *last_n_tokens_data
             ),
@@ -369,6 +381,8 @@ class Llama:
             top_p=llama_cpp.c_float(top_p),
             temp=llama_cpp.c_float(temp),
             repeat_penalty=llama_cpp.c_float(repeat_penalty),
+            frequency_penalty=llama_cpp.c_float(frequency_penalty),
+            presence_penalty=llama_cpp.c_float(presence_penalty),
         )
 
     def generate(
@@ -378,6 +392,8 @@ class Llama:
         top_p: float,
         temp: float,
         repeat_penalty: float,
+        frequency_penalty: float = 0.0,
+        presence_penalty: float = 0.0,
         reset: bool = True,
     ) -> Generator[
         llama_cpp.llama_token, Optional[Sequence[llama_cpp.llama_token]], None
@@ -431,6 +447,8 @@ class Llama:
                 top_k=top_k,
                 top_p=top_p,
                 temp=temp,
+                frequency_penalty=frequency_penalty,
+                presence_penalty=presence_penalty,
                 repeat_penalty=repeat_penalty,
             )
             tokens_or_none = yield token
@@ -505,6 +523,8 @@ class Llama:
         logprobs: Optional[int] = None,
         echo: bool = False,
         stop: Optional[List[str]] = [],
+        frequency_penalty: float = 0.0,
+        presence_penalty: float = 0.0,
         repeat_penalty: float = 1.1,
         top_k: int = 40,
         stream: bool = False,
@@ -563,6 +583,8 @@ class Llama:
             top_k=top_k,
             top_p=top_p,
             temp=temperature,
+            frequency_penalty=frequency_penalty,
+            presence_penalty=presence_penalty,
             repeat_penalty=repeat_penalty,
         ):
             if token == llama_cpp.llama_token_eos():
@@ -737,6 +759,8 @@ class Llama:
         logprobs: Optional[int] = None,
         echo: bool = False,
         stop: Optional[List[str]] = [],
+        frequency_penalty: float = 0.0,
+        presence_penalty: float = 0.0,
         repeat_penalty: float = 1.1,
         top_k: int = 40,
         stream: bool = False,
@@ -772,6 +796,8 @@ class Llama:
             logprobs=logprobs,
             echo=echo,
             stop=stop,
+            frequency_penalty=frequency_penalty,
+            presence_penalty=presence_penalty,
             repeat_penalty=repeat_penalty,
             top_k=top_k,
             stream=stream,
@@ -792,6 +818,8 @@ class Llama:
         logprobs: Optional[int] = None,
         echo: bool = False,
         stop: Optional[List[str]] = [],
+        frequency_penalty: float = 0.0,
+        presence_penalty: float = 0.0,
         repeat_penalty: float = 1.1,
         top_k: int = 40,
         stream: bool = False,
@@ -827,6 +855,8 @@ class Llama:
             logprobs=logprobs,
             echo=echo,
             stop=stop,
+            frequency_penalty=frequency_penalty,
+            presence_penalty=presence_penalty,
             repeat_penalty=repeat_penalty,
             top_k=top_k,
             stream=stream,
@@ -899,6 +929,8 @@ class Llama:
         stream: bool = False,
         stop: Optional[List[str]] = [],
         max_tokens: int = 256,
+        presence_penalty: float = 0.0,
+        frequency_penalty: float = 0.0,
         repeat_penalty: float = 1.1,
     ) -> Union[ChatCompletion, Iterator[ChatCompletionChunk]]:
         """Generate a chat completion from a list of messages.
@@ -932,6 +964,8 @@ class Llama:
             stream=stream,
             max_tokens=max_tokens,
             repeat_penalty=repeat_penalty,
+            presence_penalty=presence_penalty,
+            frequency_penalty=frequency_penalty,
         )
         if stream:
             chunks: Iterator[CompletionChunk] = completion_or_chunks  # type: ignore

llama_cpp/server/app.py

@@ -214,8 +214,6 @@ def create_completion(
             exclude={
                 "model",
                 "n",
-                "frequency_penalty",
-                "presence_penalty",
                 "best_of",
                 "logit_bias",
                 "user",
@@ -315,8 +313,6 @@ def create_chat_completion(
             exclude={
                 "model",
                 "n",
-                "presence_penalty",
-                "frequency_penalty",
                 "logit_bias",
                 "user",
             }