remove is_chat argument

evoprompt/models.py

@@ -20,6 +20,8 @@ logger = logging.getLogger(__name__)
 logging.captureWarnings(True)
 warnings.simplefilter("once")
 
+ChatMessages = list[dict[str, str]]
+
 
 class LLMModel(ABC):
     models: ClassVar[dict[str, type["LLMModel"]]] = {}
@@ -44,22 +46,17 @@ class LLMModel(ABC):
         self.kwargs = kwargs
 
         # set up caching for model calls
+        self._call_model_cached = None
         if not options.disable_cache:
             cache = Cache(Path(".cache_dir", self.model_cache_key))
-            self._call_model_cached = cache.memoize(typed=True, ignore=[0, "func"])(
-                self._call_model_cached
-            )
 
-    @abstractmethod
-    def build_model_input(
-        self,
-        prompt: str,
-        system_message: str | None,
-        messages: list[dict[str, str]],
-        history: list[dict[str, str]] | None = None,
-    ):
-        pass
+            @cache.memoize(typed=True, ignore=[0, "func"])
+            def _call_function(func, *args, **kwargs):
+                return func(*args, **kwargs)
+
+            self._call_model_cached = _call_function
 
+    @abstractmethod
     def create_completion(
         self,
         system_message: str | None,
@@ -70,26 +67,9 @@ class LLMModel(ABC):
         prompt_prefix: str = "",
         prompt_suffix: str = "",
         stop: str = None,
-        history: list[dict[str, str]] | None = None,
+        history: ChatMessages | None = None,
         **kwargs: Any,
-    ) -> tuple[str, ModelUsage]:
-        if chat is None:
-            chat = self.chat
-
-        # create prompt
-        prompt = prompt_prefix + prompt + prompt_suffix + prompt_appendix
-        messages = [self._get_user_message(prompt)]
-        model_input = self.build_model_input(prompt, system_message, messages, history)
-
-        reponse, usage = self._create_completion(
-            **model_input,
-            stop=stop,
-            use_cache=use_cache,
-            **kwargs,
-        )
-
-        messages.append(self._get_assistant_message(reponse))
-        return reponse, messages, usage
+    ) -> tuple[str, ModelUsage]: ...
 
     def _get_user_message(self, content: str):
         return {
@@ -117,16 +97,11 @@ class LLMModel(ABC):
             warnings.warn("Caching is disabled when a grammar is provided.")
             use_cache = False
 
-        if use_cache:
-            # use cached function call
+        if use_cache and self._call_model_cached is not None:
             return self._call_model_cached(model_completion_fn, **kwargs)
         else:
             return model_completion_fn(**kwargs)
 
-    def _call_model_cached(self, func, *args, **kwargs):
-        # `cache_key` is added to the cache key (e.g., to distinguish between different models), but it is not used in the function
-        return func(*args, **kwargs)
-
     @property
     def model_cache_key(self):
         unique_options_key = json.dumps(
@@ -137,7 +112,9 @@ class LLMModel(ABC):
             sort_keys=True,
         )
         cache_key = (
-            str(self.model_name) + hashlib.sha1(unique_options_key.encode()).hexdigest()
+            str(self.model_name).replace("/", "_")
+            + "/"
+            + hashlib.sha1(unique_options_key.encode()).hexdigest()
         )
         return cache_key
 
@@ -199,16 +176,35 @@ class Llama(LLMModel):
         # needs to be called after model is initialized
         super().__init__(options=options, n_ctx=n_ctx, **kwargs)
 
-    def build_model_input(
+    def create_completion(
         self,
-        prompt: str,
         system_message: str | None,
-        messages: list[dict[str, str]],
-        history: list[dict[str, str]] | None = None,
-    ):
+        prompt: str,
+        *,
+        use_cache: bool = False,
+        prompt_appendix: str = "",
+        prompt_prefix: str = "",
+        prompt_suffix: str = "",
+        stop: str = None,
+        history: ChatMessages | None = None,
+        **kwargs: Any,
+    ) -> tuple[str, ModelUsage]:
+        # create prompt
+        prompt = prompt_prefix + prompt + prompt_suffix + prompt_appendix
+        messages = [self._get_user_message(prompt)]
         if system_message is not None:
             prompt = system_message + prompt
-        return {"prompt": prompt}
+
+        reponse, usage = self._create_completion(
+            prompt=prompt,
+            stop=stop,
+            use_cache=use_cache,
+            max_tokens=self.options.max_tokens,
+            **kwargs,
+        )
+
+        messages.append(self._get_assistant_message(reponse))
+        return reponse, messages, usage
 
     def _create_completion(
         self,
@@ -264,7 +260,43 @@ class Llama(LLMModel):
         )
 
 
-class LlamaChat(Llama):
+class ChatModel:
+
+    def create_completion(
+        self,
+        system_message: str | None,
+        prompt: str,
+        *,
+        use_cache: bool = False,
+        prompt_appendix: str = "",
+        prompt_prefix: str = "",
+        prompt_suffix: str = "",
+        stop: str = None,
+        history: ChatMessages | None = None,
+        **kwargs: Any,
+    ) -> tuple[str, ModelUsage]:
+        # create prompt
+        prompt = prompt_prefix + prompt + prompt_suffix + prompt_appendix
+        messages = [self._get_user_message(prompt)]
+
+        # a history is prepended to the messages, and we assume that it also includes a system message, i.e., we never add a system message in this case
+        # TODO is it better to check for a system message in the history?
+        if history is None and system_message:
+            history = [self._get_system_message(system_message)]
+
+        reponse, usage = self._create_completion(
+            messages=messages,
+            stop=stop,
+            use_cache=use_cache,
+            max_tokens=self.options.max_tokens,
+            **kwargs,
+        )
+
+        messages.append(self._get_assistant_message(reponse))
+        return reponse, history + messages, usage
+
+
+class LlamaChat(ChatModel, Llama):
 
     def _create_completion(
         self,
@@ -281,27 +313,8 @@ class LlamaChat(Llama):
         usage = ModelUsage(**response["usage"])
         return response_text, usage
 
-    def build_model_input(
-        self,
-        prompt: str,
-        system_message: str | None,
-        messages: list[dict[str, str]],
-        history: list[dict[str, str]] | None = None,
-    ):
-
-        # a history is prepended to the messages, and we assume that it also includes a system message, i.e., we never add a system message in this case
-        # TODO is it better to check for a system message in the history?
-        if history is not None:
-            [messages.insert(index, entry) for index, entry in enumerate(history)]
-        elif system_message:
-            messages.insert(
-                0,
-                self._get_system_message(system_message),
-            )
-        return {"messages": messages}
-
 
-class OpenAI(LLMModel):
+class OpenAiChat(ChatModel, LLMModel):
     """Queries an OpenAI model using its API."""
 
     def __init__(
@@ -321,13 +334,12 @@ class OpenAI(LLMModel):
         **kwargs,
     ):
         response = self._call_model(
-            self.openai_client.completions.create,
-            model=self.model,
+            self.openai_client.chat.completions.create,
+            model=self.model_name,
             use_cache=use_cache,
             **kwargs,
         )
-        response_text = response.choices[0].text
-
+        response_text = response.choices[0].message.content
         usage = ModelUsage(**response.usage.__dict__)
         return response_text, usage
 
@@ -337,24 +349,6 @@ class OpenAI(LLMModel):
         group.add_argument("--openai-model", "-m", type=str, default="gpt-3.5-turbo")
 
 
-def OpenAiChat(OpenAI):
-
-    def _create_completion(
-        self,
-        use_cache: bool = False,
-        **kwargs,
-    ):
-        response = self._call_model(
-            self.openai_client.chat.completions.create,
-            model=self.model_name,
-            use_cache=use_cache,
-            **kwargs,
-        )
-        response_text = response.choices[0].message.content
-        usage = ModelUsage(**response.usage.__dict__)
-        return response_text, usage
-
-
 argument_group = argument_parser.add_argument_group("Model arguments")
 argument_group.add_argument(
     "--evolution-engine",
@@ -372,4 +366,3 @@ argument_group.add_argument(
     type=int,
     help="Maximum number of tokens being generated from LLM. ",
 )
-argument_group.add_argument("--chat", "-c", action="store_true")