Configure model settings, adopt litellm bugfix

aider/main.py

@@ -570,6 +570,7 @@ def main(argv=None, input=None, output=None, force_git_root=None):
 
     res = litellm.validate_environment(args.model)
 
+    # Is the model known and are all needed keys/params available?
     missing_keys = res.get("missing_keys")
     if missing_keys:
         io.tool_error(f"To use model {args.model}, please set these environment variables:")
@@ -580,6 +581,14 @@ def main(argv=None, input=None, output=None, force_git_root=None):
         io.tool_error(f"Unknown model {args.model}.")
         return 1
 
+    # Check in advance that we have model metadata
+    try:
+        litellm.get_model_info(args.model)
+    except Exception as err:
+        io.tool_error(f"Unknown model {args.model}.")
+        io.tool_error(str(err))
+        return 1
+
     main_model = models.Model(args.model)
 
     try:

aider/models/model.py

@@ -1,5 +1,6 @@
 import json
 import math
+from dataclasses import dataclass, fields
 
 import litellm
 from PIL import Image
@@ -7,35 +8,146 @@ from PIL import Image
 from aider.dump import dump
 
 
+@dataclass
+class ModelSettings:
+    name: str
+    edit_format: str
+    weak_model_name: str = "gpt-3.5-turbo-0125"
+    use_repo_map: bool = False
+    send_undo_reply: bool = False
+
+
+# https://platform.openai.com/docs/models/gpt-4-and-gpt-4-turbo
+# https://platform.openai.com/docs/models/gpt-3-5-turbo
+# https://openai.com/pricing
+
+MODEL_SETTINGS = [
+    # gpt-3.5
+    ModelSettings(
+        "gpt-3.5-turbo-0125",
+        "whole",
+    ),
+    ModelSettings(
+        "gpt-3.5-turbo-1106",
+        "whole",
+    ),
+    ModelSettings(
+        "gpt-3.5-turbo-0613",
+        "whole",
+    ),
+    ModelSettings(
+        "gpt-3.5-turbo-16k-0613",
+        "whole",
+    ),
+    # gpt-4
+    ModelSettings(
+        "gpt-4-turbo-2024-04-09",
+        "udiff",
+        use_repo_map=True,
+        send_undo_reply=True,
+    ),
+    ModelSettings(
+        "gpt-4-0125-preview",
+        "udiff",
+        use_repo_map=True,
+        send_undo_reply=True,
+    ),
+    ModelSettings(
+        "gpt-4-1106-preview",
+        "udiff",
+        use_repo_map=True,
+        send_undo_reply=True,
+    ),
+    ModelSettings(
+        "gpt-4-vision-preview",
+        "diff",
+        use_repo_map=True,
+        send_undo_reply=True,
+    ),
+    ModelSettings(
+        "gpt-4-0613",
+        "diff",
+        use_repo_map=True,
+        send_undo_reply=True,
+    ),
+    ModelSettings(
+        "gpt-4-32k-0613",
+        "diff",
+        use_repo_map=True,
+        send_undo_reply=True,
+    ),
+    # Claude
+    ModelSettings(
+        "claude-3-opus-20240229",
+        "udiff",
+        weak_model_name="claude-3-haiku-20240307",
+        use_repo_map=True,
+        send_undo_reply=True,
+    ),
+]
+
+ALIASES = {
+    # gpt-3.5
+    "gpt-3.5-turbo": "gpt-3.5-turbo-0613",
+    "gpt-3.5-turbo-16k": "gpt-3.5-turbo-16k-0613",
+    # gpt-4
+    "gpt-4-turbo": "gpt-4-turbo-2024-04-09",
+    "gpt-4-turbo-preview": "gpt-4-0125-preview",
+    "gpt-4": "gpt-4-0613",
+    "gpt-4-32k": "gpt-4-32k-0613",
+}
+
+
 class Model:
     name = None
-    edit_format = "whole"
 
+    weak_model_name = "gpt-3.5-turbo-0125"
+    edit_format = "whole"
     use_repo_map = False
     send_undo_reply = False
+
     max_chat_history_tokens = 1024
 
     def __init__(self, model):
         self.name = model
         self.info = litellm.get_model_info(model)
-        dump(self.info)
+
+        dump(model, self.info)
 
         if self.info.get("max_input_tokens", 0) < 32 * 1024:
             self.max_chat_history_tokens = 1024
         else:
             self.max_chat_history_tokens = 2 * 1024
 
-        # TODO: set edit_format,use_repo_map,send_undo_reply for various models
+        self.configure_model_settings(model)
+
+    def configure_model_settings(self, model):
+        for ms in MODEL_SETTINGS:
+            # direct match, or match "provider/<model>"
+            if model == ms.name or model.endswith("/" + ms.name):
+                for field in fields(ModelSettings):
+                    val = getattr(ms, field.name)
+                    setattr(self, field.name, val)
+
+                return  # <--
+
+        if "gpt-4" in model or "claude-2" in model:
+            self.edit_format = "diff"
+            self.use_repo_map = True
+            self.send_undo_reply = True
+
+            return  # <--
+
+        # use the defaults
 
     def __str__(self):
         return self.name
 
     def weak_model(self):
-        model = "gpt-3.5-turbo-0125"
+        if self.name == self.weak_model_name:
-        if self.name == model:
             return self
 
-        return Model(model)
+        return Model(self.weak_model_name)
 
     def commit_message_models(self):
         return [self.weak_model()]

aider/models/openai.py

@@ -1,159 +0,0 @@
-from dataclasses import dataclass, fields
-
-import tiktoken
-
-from aider.dump import dump  # noqa: F401
-
-from .model import Model
-
-
-@dataclass
-class ModelInfo:
-    name: str
-    max_context_tokens: int
-    prompt_price: float
-    completion_price: float
-    edit_format: str
-    always_available: bool = False
-    use_repo_map: bool = False
-    send_undo_reply: bool = False
-
-
-# https://platform.openai.com/docs/models/gpt-4-and-gpt-4-turbo
-# https://platform.openai.com/docs/models/gpt-3-5-turbo
-# https://openai.com/pricing
-
-openai_models = [
-    # gpt-3.5
-    ModelInfo(
-        "gpt-3.5-turbo-0125",
-        16385,
-        0.0005,
-        0.0015,
-        "whole",
-        always_available=True,
-    ),
-    ModelInfo(
-        "gpt-3.5-turbo-1106",
-        16385,
-        0.0010,
-        0.0020,
-        "whole",
-        always_available=True,
-    ),
-    ModelInfo(
-        "gpt-3.5-turbo-0613",
-        4096,
-        0.0015,
-        0.0020,
-        "whole",
-        always_available=True,
-    ),
-    ModelInfo(
-        "gpt-3.5-turbo-16k-0613",
-        16385,
-        0.0030,
-        0.0040,
-        "whole",
-        always_available=True,
-    ),
-    # gpt-4
-    ModelInfo(
-        "gpt-4-turbo-2024-04-09",
-        128000,
-        0.01,
-        0.03,
-        "udiff",
-        use_repo_map=True,
-        send_undo_reply=True,
-    ),
-    ModelInfo(
-        "gpt-4-0125-preview",
-        128000,
-        0.01,
-        0.03,
-        "udiff",
-        use_repo_map=True,
-        send_undo_reply=True,
-    ),
-    ModelInfo(
-        "gpt-4-1106-preview",
-        128000,
-        0.01,
-        0.03,
-        "udiff",
-        use_repo_map=True,
-        send_undo_reply=True,
-    ),
-    ModelInfo(
-        "gpt-4-vision-preview",
-        128000,
-        0.01,
-        0.03,
-        "diff",
-        use_repo_map=True,
-        send_undo_reply=True,
-    ),
-    ModelInfo(
-        "gpt-4-0613",
-        8192,
-        0.03,
-        0.06,
-        "diff",
-        use_repo_map=True,
-        send_undo_reply=True,
-    ),
-    ModelInfo(
-        "gpt-4-32k-0613",
-        32768,
-        0.06,
-        0.12,
-        "diff",
-        use_repo_map=True,
-        send_undo_reply=True,
-    ),
-]
-
-openai_aliases = {
-    # gpt-3.5
-    "gpt-3.5-turbo": "gpt-3.5-turbo-0613",
-    "gpt-3.5-turbo-16k": "gpt-3.5-turbo-16k-0613",
-    # gpt-4
-    "gpt-4-turbo": "gpt-4-turbo-2024-04-09",
-    "gpt-4-turbo-preview": "gpt-4-0125-preview",
-    "gpt-4": "gpt-4-0613",
-    "gpt-4-32k": "gpt-4-32k-0613",
-}
-
-
-class OpenAIModel(Model):
-    def __init__(self, name):
-        true_name = openai_aliases.get(name, name)
-
-        try:
-            self.tokenizer = tiktoken.encoding_for_model(true_name)
-        except KeyError:
-            self.tokenizer = None
-            # raise ValueError(f"No known tokenizer for model: {name}")
-
-        model_info = self.lookup_model_info(true_name)
-        if not model_info:
-            raise ValueError(f"Unsupported model: {name}")
-
-        for field in fields(ModelInfo):
-            val = getattr(model_info, field.name)
-            setattr(self, field.name, val)
-
-        # restore the caller's specified name
-        self.name = name
-
-        # set the history token limit
-        if self.max_context_tokens < 32 * 1024:
-            self.max_chat_history_tokens = 1024
-        else:
-            self.max_chat_history_tokens = 2 * 1024
-
-    def lookup_model_info(self, name):
-        for mi in openai_models:
-            if mi.name == name:
-                return mi

aider/models/openrouter.py

@@ -1,40 +0,0 @@
-import tiktoken
-
-from .model import Model
-
-cached_model_details = None
-
-
-class OpenRouterModel(Model):
-    def __init__(self, client, name):
-        if name.startswith("gpt-4") or name.startswith("gpt-3.5-turbo"):
-            name = "openai/" + name
-
-        self.name = name
-        self.edit_format = edit_format_for_model(name)
-        self.use_repo_map = self.edit_format == "diff"
-
-        # TODO: figure out proper encodings for non openai models
-        self.tokenizer = tiktoken.get_encoding("cl100k_base")
-
-        global cached_model_details
-        if cached_model_details is None:
-            cached_model_details = client.models.list().data
-        found = next(
-            (details for details in cached_model_details if details.id == name), None
-        )
-
-        if found:
-            self.max_context_tokens = int(found.context_length)
-            self.prompt_price = round(float(found.pricing.get("prompt")) * 1000, 6)
-            self.completion_price = round(float(found.pricing.get("completion")) * 1000, 6)
-
-        else:
-            raise ValueError(f"invalid openrouter model: {name}")
-
-
-def edit_format_for_model(name):
-    if any(str in name for str in ["gpt-4", "claude-2"]):
-        return "diff"
-
-    return "whole"