initial code for working with openrouter

aider/coders/base_coder.py

@@ -636,7 +636,7 @@ class Coder:
                 if len(chunk.choices) == 0:
                     continue
 
-                if chunk.choices[0].finish_reason == "length":
+                if hasattr(chunk.choices[0], "finish_reason") and chunk.choices[0].finish_reason == "length":
                     raise ExhaustedContextWindow()
 
                 try:
@@ -901,5 +901,6 @@ class Coder:
 
 def check_model_availability(main_model):
     available_models = openai.Model.list()
+    print(available_models)
     model_ids = [model.id for model in available_models["data"]]
     return main_model.name in model_ids

aider/commands.py

@@ -5,7 +5,6 @@ import sys
 from pathlib import Path
 
 import git
-import tiktoken
 from prompt_toolkit.completion import Completion
 
 from aider import prompts, voice
@@ -24,7 +23,7 @@ class Commands:
             voice_language = None
 
         self.voice_language = voice_language
-        self.tokenizer = tiktoken.encoding_for_model(coder.main_model.name)
+        self.tokenizer = coder.main_model.tokenizer
 
     def is_command(self, inp):
         if inp[0] == "/":

aider/history.py

@@ -85,7 +85,7 @@ class ChatSummary:
             dict(role="user", content=content),
         ]
 
-        summary = simple_send_with_retries(model=models.GPT35.name, messages=messages)
+        summary = simple_send_with_retries(model=self.model.weak_model, messages=messages)
         summary = prompts.summary_prefix + summary
 
         return [dict(role="user", content=summary)]

aider/main.py

@@ -449,8 +449,6 @@ def main(argv=None, input=None, output=None, force_git_root=None):
             )
         return 1
 
-    main_model = models.Model(args.model)
-
     openai.api_key = args.openai_api_key
     for attr in ("base", "type", "version", "deployment_id", "engine"):
         arg_key = f"openai_api_{attr}"
@@ -460,6 +458,8 @@ def main(argv=None, input=None, output=None, force_git_root=None):
             setattr(openai, mod_key, val)
             io.tool_output(f"Setting openai.{mod_key}={val}")
 
+    main_model = models.Model(args.model, openai)
+
     try:
         coder = Coder.create(
             main_model,

aider/models/__init__.py

@@ -0,0 +1,7 @@
+from .openai import OpenAIModel
+from .openrouter import OpenRouterModel
+from .model import Model
+
+GPT4 = Model("gpt-4")
+GPT35 = Model("gpt-3.5-turbo")
+GPT35_16k = Model("gpt-3.5-turbo-16k")

aider/models/model.py

@@ -0,0 +1,60 @@
+import importlib
+
+using_openrouter = False
+
+class Model:
+    name = None
+    edit_format = None
+    max_context_tokens = 0
+    tokenizer = None
+
+    always_available = False
+    use_repo_map = False
+    send_undo_reply = False
+
+    prompt_price = None
+    completion_price = None
+
+    def __init__(self, name, openai=None):
+        global using_openrouter
+        if (openai and "openrouter.ai" in openai.api_base):
+            using_openrouter = True
+
+        from .openai import OpenAIModel
+        from .openrouter import OpenRouterModel
+        model = None
+        if using_openrouter:
+            if name == 'gpt-4':
+                name = 'openai/gpt-4'
+            elif name == 'gpt-3.5-turbo':
+                name = 'openai/gpt-3.5-turbo'
+            elif name == 'gpt-3.5.turbo-16k':
+                name = 'openai/gpt-3.5-turbo-16k'
+
+            model = OpenRouterModel(name, openai)
+        else:
+            model = OpenAIModel(name)
+
+        self.name = model.name
+        self.edit_format = model.edit_format
+        self.max_context_tokens = model.max_context_tokens
+        self.tokenizer = model.tokenizer
+        self.prompt_price = model.prompt_price
+        self.completion_price = model.completion_price
+        self.always_available = model.always_available
+        self.use_repo_map = model.use_repo_map
+
+    def __str__(self):
+        return self.name
+
+    @staticmethod
+    def strong_model():
+        return Model('gpt-4')
+
+    @staticmethod
+    def weak_model():
+        return Model('gpt-3.5-turbo')
+
+    @staticmethod
+    def commit_message_models():
+        return [Model('gpt-3.5-turbo'), Model('gpt-3.5-turbo-16k')]

aider/models/openai.py

@@ -1,4 +1,6 @@
+import tiktoken
 import re
+from .model import Model
 
 known_tokens = {
     "gpt-3.5-turbo": 4,
@@ -6,14 +8,7 @@ known_tokens = {
 }
 
 
-class Model:
-    always_available = False
-    use_repo_map = False
-    send_undo_reply = False
-
-    prompt_price = None
-    completion_price = None
-
+class OpenAIModel(Model):
     def __init__(self, name):
         self.name = name
 
@@ -31,6 +26,7 @@ class Model:
             raise ValueError(f"Unknown context window size for model: {name}")
 
         self.max_context_tokens = tokens * 1024
+        self.tokenizer = tiktoken.encoding_for_model(name)
 
         if self.is_gpt4():
             self.edit_format = "diff"
@@ -66,11 +62,3 @@ class Model:
 
     def is_gpt35(self):
         return self.name.startswith("gpt-3.5-turbo")
-
-    def __str__(self):
-        return self.name
-
-
-GPT4 = Model("gpt-4")
-GPT35 = Model("gpt-3.5-turbo")
-GPT35_16k = Model("gpt-3.5-turbo-16k")

aider/models/openrouter.py

@@ -0,0 +1,13 @@
+import tiktoken
+from .model import Model
+
+
+class OpenRouterModel(Model):
+    def __init__(self, name, openai):
+        self.name = name
+        self.edit_format = "diff"
+        self.use_repo_map = True
+        self.max_context_tokens = 1024 * 8
+
+        # TODO: figure out proper encodings for non openai models
+        self.tokenizer = tiktoken.get_encoding("cl100k_base")

aider/repo.py

@@ -109,8 +109,8 @@ class GitRepo:
             dict(role="user", content=content),
         ]
 
-        for model in [models.GPT35.name, models.GPT35_16k.name]:
-            commit_message = simple_send_with_retries(model, messages)
+        for model in models.Model.commit_message_models():
+            commit_message = simple_send_with_retries(model.name, messages)
             if commit_message:
                 break
 

aider/repomap.py

@@ -9,7 +9,6 @@ from collections import Counter, defaultdict
 from pathlib import Path
 
 import networkx as nx
-import tiktoken
 from diskcache import Cache
 from pygments.lexers import guess_lexer_for_filename
 from pygments.token import Token
@@ -104,7 +103,7 @@ class RepoMap:
         else:
             self.use_ctags = False
 
-        self.tokenizer = tiktoken.encoding_for_model(main_model.name)
+        self.tokenizer = main_model.tokenizer
         self.repo_content_prefix = repo_content_prefix
 
     def get_repo_map(self, chat_files, other_files):

aider/sendchat.py

@@ -50,6 +50,12 @@ def send_with_retries(model, messages, functions, stream):
     if hasattr(openai, "api_engine"):
         kwargs["engine"] = openai.api_engine
 
+    if "openrouter.ai" in openai.api_base:
+        kwargs["headers"] = {
+            "HTTP-Referer": "http://aider.chat",
+            "X-Title": "Aider"
+        }
+
     key = json.dumps(kwargs, sort_keys=True).encode()
 
     # Generate SHA1 hash of kwargs and append it to chat_completion_call_hashes