Skip to content

armgit / camel

Public Imported from GitHub
0

🐫 CAMEL: The first and the best multi-agent framework. Finding the Scaling Law of Agents. https://www.camel-ai.org

feat: allow users self setting token counter from ModelFactory (#730)

Yihan Lina55aeb6dfe8e
diff --git a/camel/models/anthropic_model.py b/camel/models/anthropic_model.py
index 4d5d33554..a04267014 100644
--- a/camel/models/anthropic_model.py
+++ b/camel/models/anthropic_model.py
@@ -36,6 +36,7 @@ class AnthropicModel(BaseModelBackend):
model_config_dict: Dict[str, Any],
api_key: Optional[str] = None,
url: Optional[str] = None,
+ token_counter: Optional[BaseTokenCounter] = None,
) -> None:
r"""Constructor for Anthropic backend.
@@ -48,12 +49,16 @@ class AnthropicModel(BaseModelBackend):
Anthropic service. (default: :obj:`None`)
url (Optional[str]): The url to the Anthropic service. (default:
:obj:`None`)
+ token_counter (Optional[BaseTokenCounter]): Token counter to use
+ for the model. If not provided, `AnthropicTokenCounter` will
+ be used.
"""
- super().__init__(model_type, model_config_dict, api_key, url)
+ super().__init__(
+ model_type, model_config_dict, api_key, url, token_counter
+ )
self._api_key = api_key or os.environ.get("ANTHROPIC_API_KEY")
self._url = url or os.environ.get("ANTHROPIC_API_BASE_URL")
self.client = Anthropic(api_key=self._api_key, base_url=self._url)
- self._token_counter: Optional[BaseTokenCounter] = None
def _convert_response_from_anthropic_to_openai(self, response):
# openai ^1.0.0 format, reference openai/types/chat/chat_completion.py
diff --git a/camel/models/base_model.py b/camel/models/base_model.py
index 7df7c0ce3..aa7df0b12 100644
--- a/camel/models/base_model.py
+++ b/camel/models/base_model.py
@@ -32,6 +32,7 @@ class BaseModelBackend(ABC):
model_config_dict: Dict[str, Any],
api_key: Optional[str] = None,
url: Optional[str] = None,
+ token_counter: Optional[BaseTokenCounter] = None,
) -> None:
r"""Constructor for the model backend.
@@ -41,13 +42,16 @@ class BaseModelBackend(ABC):
api_key (Optional[str]): The API key for authenticating with the
model service.
url (Optional[str]): The url to the model service.
+ token_counter (Optional[BaseTokenCounter]): Token counter to use
+ for the model. If not provided, `OpenAITokenCounter` will
+ be used.
"""
self.model_type = model_type
-
self.model_config_dict = model_config_dict
self._api_key = api_key
self._url = url
self.check_model_config()
+ self._token_counter = token_counter
@property
@abstractmethod
diff --git a/camel/models/gemini_model.py b/camel/models/gemini_model.py
index 8a497a786..f98bfa82c 100644
--- a/camel/models/gemini_model.py
+++ b/camel/models/gemini_model.py
@@ -44,6 +44,7 @@ class GeminiModel(BaseModelBackend):
model_config_dict: Dict[str, Any],
api_key: Optional[str] = None,
url: Optional[str] = None,
+ token_counter: Optional[BaseTokenCounter] = None,
) -> None:
r"""Constructor for Gemini backend.
@@ -54,17 +55,22 @@ class GeminiModel(BaseModelBackend):
api_key (Optional[str]): The API key for authenticating with the
gemini service. (default: :obj:`None`)
url (Optional[str]): The url to the gemini service.
+ token_counter (Optional[BaseTokenCounter]): Token counter to use
+ for the model. If not provided, `GeminiTokenCounter` will be
+ used.
"""
import os
import google.generativeai as genai
from google.generativeai.types.generation_types import GenerationConfig
- super().__init__(model_type, model_config_dict, api_key, url)
+ super().__init__(
+ model_type, model_config_dict, api_key, url, token_counter
+ )
self._api_key = api_key or os.environ.get("GOOGLE_API_KEY")
genai.configure(api_key=self._api_key)
self._client = genai.GenerativeModel(self.model_type.value)
- self._token_counter: Optional[BaseTokenCounter] = None
+
keys = list(self.model_config_dict.keys())
generation_config_dict = {
k: self.model_config_dict.pop(k)
@@ -78,6 +84,12 @@ class GeminiModel(BaseModelBackend):
@property
def token_counter(self) -> BaseTokenCounter:
+ r"""Initialize the token counter for the model backend.
+
+ Returns:
+ BaseTokenCounter: The token counter following the model's
+ tokenization style.
+ """
if not self._token_counter:
self._token_counter = GeminiTokenCounter(self.model_type)
return self._token_counter
diff --git a/camel/models/litellm_model.py b/camel/models/litellm_model.py
index a8e7e0d3e..8d62e7f6d 100644
--- a/camel/models/litellm_model.py
+++ b/camel/models/litellm_model.py
@@ -16,7 +16,7 @@ from typing import Any, Dict, List, Optional
from camel.configs import LITELLM_API_PARAMS
from camel.messages import OpenAIMessage
from camel.types import ChatCompletion
-from camel.utils import LiteLLMTokenCounter
+from camel.utils import BaseTokenCounter, LiteLLMTokenCounter
class LiteLLMModel:
@@ -30,6 +30,7 @@ class LiteLLMModel:
model_config_dict: Dict[str, Any],
api_key: Optional[str] = None,
url: Optional[str] = None,
+ token_counter: Optional[BaseTokenCounter] = None,
) -> None:
r"""Constructor for LiteLLM backend.
@@ -42,11 +43,14 @@ class LiteLLMModel:
model service. (default: :obj:`None`)
url (Optional[str]): The url to the model service. (default:
:obj:`None`)
+ token_counter (Optional[BaseTokenCounter]): Token counter to use
+ for the model. If not provided, `LiteLLMTokenCounter` will
+ be used.
"""
self.model_type = model_type
self.model_config_dict = model_config_dict
self._client = None
- self._token_counter: Optional[LiteLLMTokenCounter] = None
+ self._token_counter = token_counter
self.check_model_config()
self._url = url
self._api_key = api_key
@@ -98,8 +102,10 @@ class LiteLLMModel:
tokenization style.
"""
if not self._token_counter:
- self._token_counter = LiteLLMTokenCounter(self.model_type)
- return self._token_counter
+ self._token_counter = LiteLLMTokenCounter( # type: ignore[assignment]
+ self.model_type
+ )
+ return self._token_counter # type: ignore[return-value]
def run(
self,
diff --git a/camel/models/model_factory.py b/camel/models/model_factory.py
index 78a5ec8e5..bc99c7d0f 100644
--- a/camel/models/model_factory.py
+++ b/camel/models/model_factory.py
@@ -24,6 +24,7 @@ from camel.models.stub_model import StubModel
from camel.models.vllm_model import VLLMModel
from camel.models.zhipuai_model import ZhipuAIModel
from camel.types import ModelPlatformType, ModelType
+from camel.utils import BaseTokenCounter
class ModelFactory:
@@ -38,6 +39,7 @@ class ModelFactory:
model_platform: ModelPlatformType,
model_type: Union[ModelType, str],
model_config_dict: Dict,
+ token_counter: Optional[BaseTokenCounter] = None,
api_key: Optional[str] = None,
url: Optional[str] = None,
) -> BaseModelBackend:
@@ -50,6 +52,10 @@ class ModelFactory:
created can be a `str` for open source platforms.
model_config_dict (Dict): A dictionary that will be fed into
the backend constructor.
+ token_counter (Optional[BaseTokenCounter]): Token counter to use
+ for the model. If not provided, OpenAITokenCounter(ModelType.
+ GPT_3_5_TURBO) will be used if the model platform didn't
+ provide official token counter.
api_key (Optional[str]): The API key for authenticating with the
model service.
url (Optional[str]): The url to the model service.
@@ -64,7 +70,9 @@ class ModelFactory:
if isinstance(model_type, ModelType):
if model_platform.is_open_source and model_type.is_open_source:
model_class = OpenSourceModel
- return model_class(model_type, model_config_dict, url)
+ return model_class(
+ model_type, model_config_dict, url, token_counter
+ )
if model_platform.is_openai and model_type.is_openai:
model_class = OpenAIModel
elif model_platform.is_anthropic and model_type.is_anthropic:
@@ -83,10 +91,14 @@ class ModelFactory:
elif isinstance(model_type, str):
if model_platform.is_ollama:
model_class = OllamaModel
- return model_class(model_type, model_config_dict, url)
+ return model_class(
+ model_type, model_config_dict, url, token_counter
+ )
elif model_platform.is_vllm:
model_class = VLLMModel
- return model_class(model_type, model_config_dict, url, api_key)
+ return model_class(
+ model_type, model_config_dict, url, api_key, token_counter
+ )
elif model_platform.is_litellm:
model_class = LiteLLMModel
else:
@@ -96,4 +108,6 @@ class ModelFactory:
)
else:
raise ValueError(f"Invalid model type `{model_type}` provided.")
- return model_class(model_type, model_config_dict, api_key, url)
+ return model_class(
+ model_type, model_config_dict, api_key, url, token_counter
+ )
diff --git a/camel/models/ollama_model.py b/camel/models/ollama_model.py
index 4afe6c0d3..52b55025a 100644
--- a/camel/models/ollama_model.py
+++ b/camel/models/ollama_model.py
@@ -29,6 +29,7 @@ class OllamaModel:
model_type: str,
model_config_dict: Dict[str, Any],
url: Optional[str] = None,
+ token_counter: Optional[BaseTokenCounter] = None,
) -> None:
r"""Constructor for Ollama backend with OpenAI compatibility.
@@ -40,6 +41,9 @@ class OllamaModel:
be fed into openai.ChatCompletion.create().
url (Optional[str]): The url to the model service. (default:
:obj:`None`)
+ token_counter (Optional[BaseTokenCounter]): Token counter to use
+ for the model. If not provided, `OpenAITokenCounter(ModelType.
+ GPT_3_5_TURBO)` will be used.
"""
self.model_type = model_type
self.model_config_dict = model_config_dict
@@ -50,7 +54,7 @@ class OllamaModel:
base_url=url,
api_key="ollama", # required but ignored
)
- self._token_counter: Optional[BaseTokenCounter] = None
+ self._token_counter = token_counter
self.check_model_config()
@property
@@ -61,7 +65,6 @@ class OllamaModel:
BaseTokenCounter: The token counter following the model's
tokenization style.
"""
- # NOTE: Use OpenAITokenCounter temporarily
if not self._token_counter:
self._token_counter = OpenAITokenCounter(ModelType.GPT_3_5_TURBO)
return self._token_counter
diff --git a/camel/models/open_source_model.py b/camel/models/open_source_model.py
index 35a74ef14..cfadc7094 100644
--- a/camel/models/open_source_model.py
+++ b/camel/models/open_source_model.py
@@ -19,7 +19,10 @@ from camel.configs import OPENAI_API_PARAMS
from camel.messages import OpenAIMessage
from camel.models import BaseModelBackend
from camel.types import ChatCompletion, ChatCompletionChunk, ModelType
-from camel.utils import BaseTokenCounter, OpenSourceTokenCounter
+from camel.utils import (
+ BaseTokenCounter,
+ OpenSourceTokenCounter,
+)
class OpenSourceModel(BaseModelBackend):
@@ -33,6 +36,7 @@ class OpenSourceModel(BaseModelBackend):
model_config_dict: Dict[str, Any],
api_key: Optional[str] = None,
url: Optional[str] = None,
+ token_counter: Optional[BaseTokenCounter] = None,
) -> None:
r"""Constructor for model backends of Open-source models.
@@ -43,9 +47,13 @@ class OpenSourceModel(BaseModelBackend):
api_key (Optional[str]): The API key for authenticating with the
model service. (ignored for open-source models)
url (Optional[str]): The url to the model service.
+ token_counter (Optional[BaseTokenCounter]): Token counter to use
+ for the model. If not provided, `OpenSourceTokenCounter` will
+ be used.
"""
- super().__init__(model_type, model_config_dict, api_key, url)
- self._token_counter: Optional[BaseTokenCounter] = None
+ super().__init__(
+ model_type, model_config_dict, api_key, url, token_counter
+ )
# Check whether the input model type is open-source
if not model_type.is_open_source:
diff --git a/camel/models/openai_model.py b/camel/models/openai_model.py
index 4baed2753..9edd2f190 100644
--- a/camel/models/openai_model.py
+++ b/camel/models/openai_model.py
@@ -36,6 +36,7 @@ class OpenAIModel(BaseModelBackend):
model_config_dict: Dict[str, Any],
api_key: Optional[str] = None,
url: Optional[str] = None,
+ token_counter: Optional[BaseTokenCounter] = None,
) -> None:
r"""Constructor for OpenAI backend.
@@ -48,8 +49,13 @@ class OpenAIModel(BaseModelBackend):
OpenAI service. (default: :obj:`None`)
url (Optional[str]): The url to the OpenAI service. (default:
:obj:`None`)
+ token_counter (Optional[BaseTokenCounter]): Token counter to use
+ for the model. If not provided, `OpenAITokenCounter` will
+ be used.
"""
- super().__init__(model_type, model_config_dict, api_key, url)
+ super().__init__(
+ model_type, model_config_dict, api_key, url, token_counter
+ )
self._url = url or os.environ.get("OPENAI_API_BASE_URL")
self._api_key = api_key or os.environ.get("OPENAI_API_KEY")
self._client = OpenAI(
@@ -58,7 +64,6 @@ class OpenAIModel(BaseModelBackend):
base_url=self._url,
api_key=self._api_key,
)
- self._token_counter: Optional[BaseTokenCounter] = None
@property
def token_counter(self) -> BaseTokenCounter:
diff --git a/camel/models/stub_model.py b/camel/models/stub_model.py
index 291fc02b3..dcb60078c 100644
--- a/camel/models/stub_model.py
+++ b/camel/models/stub_model.py
@@ -55,12 +55,12 @@ class StubModel(BaseModelBackend):
model_config_dict: Dict[str, Any],
api_key: Optional[str] = None,
url: Optional[str] = None,
+ token_counter: Optional[BaseTokenCounter] = None,
) -> None:
r"""All arguments are unused for the dummy model."""
- super().__init__(model_type, model_config_dict, api_key, url)
- self._token_counter: Optional[BaseTokenCounter] = None
- self._api_key = api_key
- self._url = url
+ super().__init__(
+ model_type, model_config_dict, api_key, url, token_counter
+ )
@property
def token_counter(self) -> BaseTokenCounter:
diff --git a/camel/models/vllm_model.py b/camel/models/vllm_model.py
index 61eac05b9..111ef7379 100644
--- a/camel/models/vllm_model.py
+++ b/camel/models/vllm_model.py
@@ -31,6 +31,7 @@ class VLLMModel:
model_config_dict: Dict[str, Any],
url: Optional[str] = None,
api_key: Optional[str] = None,
+ token_counter: Optional[BaseTokenCounter] = None,
) -> None:
r"""Constructor for vLLM backend with OpenAI compatibility.
@@ -44,6 +45,9 @@ class VLLMModel:
:obj:`None`)
api_key (Optional[str]): The API key for authenticating with the
model service.
+ token_counter (Optional[BaseTokenCounter]): Token counter to use
+ for the model. If not provided, `OpenAITokenCounter(ModelType.
+ GPT_3_5_TURBO)` will be used.
"""
self.model_type = model_type
self.model_config_dict = model_config_dict
@@ -54,7 +58,7 @@ class VLLMModel:
base_url=url,
api_key=api_key,
)
- self._token_counter: Optional[BaseTokenCounter] = None
+ self._token_counter = token_counter
self.check_model_config()
@property
@@ -65,7 +69,6 @@ class VLLMModel:
BaseTokenCounter: The token counter following the model's
tokenization style.
"""
- # NOTE: Use OpenAITokenCounter temporarily
if not self._token_counter:
self._token_counter = OpenAITokenCounter(ModelType.GPT_3_5_TURBO)
return self._token_counter
diff --git a/camel/models/zhipuai_model.py b/camel/models/zhipuai_model.py
index 62025ea92..5115d3c7e 100644
--- a/camel/models/zhipuai_model.py
+++ b/camel/models/zhipuai_model.py
@@ -37,6 +37,7 @@ class ZhipuAIModel(BaseModelBackend):
model_config_dict: Dict[str, Any],
api_key: Optional[str] = None,
url: Optional[str] = None,
+ token_counter: Optional[BaseTokenCounter] = None,
) -> None:
r"""Constructor for ZhipuAI backend.
@@ -49,8 +50,13 @@ class ZhipuAIModel(BaseModelBackend):
ZhipuAI service. (default: :obj:`None`)
url (Optional[str]): The url to the ZhipuAI service. (default:
:obj:`None`)
+ token_counter (Optional[BaseTokenCounter]): Token counter to use
+ for the model. If not provided, `OpenAITokenCounter(ModelType.
+ GPT_3_5_TURBO)` will be used.
"""
- super().__init__(model_type, model_config_dict)
+ super().__init__(
+ model_type, model_config_dict, api_key, url, token_counter
+ )
self._url = url or os.environ.get("ZHIPUAI_API_BASE_URL")
self._api_key = api_key or os.environ.get("ZHIPUAI_API_KEY")
if not self._url or not self._api_key:
@@ -63,7 +69,6 @@ class ZhipuAIModel(BaseModelBackend):
api_key=self._api_key,
base_url=self._url,
)
- self._token_counter: Optional[BaseTokenCounter] = None
@api_keys_required("ZHIPUAI_API_KEY")
def run(
@@ -100,7 +105,6 @@ class ZhipuAIModel(BaseModelBackend):
"""
if not self._token_counter:
- # It's a temporary setting for token counter.
self._token_counter = OpenAITokenCounter(ModelType.GPT_3_5_TURBO)
return self._token_counter
diff --git a/camel/utils/token_counting.py b/camel/utils/token_counting.py
index 0a63d821f..2a4f59614 100644
--- a/camel/utils/token_counting.py
+++ b/camel/utils/token_counting.py
@@ -232,6 +232,7 @@ class OpenAITokenCounter(BaseTokenCounter):
model (ModelType): Model type for which tokens will be counted.
"""
self.model: str = model.value_for_tiktoken
+ self.model_type = model
self.tokens_per_message: int
self.tokens_per_name: int
diff --git a/test/models/test_model_factory.py b/test/models/test_model_factory.py
index 0cd2d2549..bb666dc6c 100644
--- a/test/models/test_model_factory.py
+++ b/test/models/test_model_factory.py
@@ -13,9 +13,21 @@
# =========== Copyright 2023 @ CAMEL-AI.org. All Rights Reserved. ===========
import pytest
-from camel.configs import ChatGPTConfig
+from camel.configs import (
+ AnthropicConfig,
+ ChatGPTConfig,
+ GeminiConfig,
+ OllamaConfig,
+ OpenSourceConfig,
+)
from camel.models import ModelFactory
+from camel.models.stub_model import StubTokenCounter
from camel.types import ModelPlatformType, ModelType
+from camel.utils import (
+ AnthropicTokenCounter,
+ OpenAITokenCounter,
+ OpenSourceTokenCounter,
+)
parametrize = pytest.mark.parametrize(
'model_platform, model_type',
@@ -26,6 +38,115 @@ parametrize = pytest.mark.parametrize(
],
)
+parameterize_token_counter = pytest.mark.parametrize(
+ 'model_platform, model_type, model_config_dict, token_counter,'
+ ' expected_counter_type, expected_model_type',
+ [
+ # Test OpenAI model
+ (
+ ModelPlatformType.OPENAI,
+ ModelType.GPT_3_5_TURBO,
+ ChatGPTConfig().__dict__,
+ None,
+ OpenAITokenCounter,
+ ModelType.GPT_3_5_TURBO,
+ ),
+ (
+ ModelPlatformType.OPENAI,
+ ModelType.GPT_4,
+ ChatGPTConfig().__dict__,
+ None,
+ OpenAITokenCounter,
+ ModelType.GPT_4,
+ ),
+ # Test Stub model
+ # Stub model uses StubTokenCounter as default
+ (
+ ModelPlatformType.OPENSOURCE,
+ ModelType.STUB,
+ ChatGPTConfig().__dict__,
+ None,
+ StubTokenCounter,
+ None,
+ ),
+ (
+ ModelPlatformType.OPENSOURCE,
+ ModelType.STUB,
+ ChatGPTConfig().__dict__,
+ OpenAITokenCounter(ModelType.GPT_4),
+ OpenAITokenCounter,
+ ModelType.GPT_4,
+ ),
+ # Test Anthropic model
+ # Anthropic model uses AnthropicTokenCounter as default
+ (
+ ModelPlatformType.ANTHROPIC,
+ ModelType.CLAUDE_2_0,
+ AnthropicConfig().__dict__,
+ None,
+ AnthropicTokenCounter,
+ ModelType.CLAUDE_2_0,
+ ),
+ (
+ ModelPlatformType.ANTHROPIC,
+ ModelType.CLAUDE_2_0,
+ AnthropicConfig().__dict__,
+ OpenAITokenCounter(ModelType.GPT_3_5_TURBO),
+ OpenAITokenCounter,
+ ModelType.GPT_3_5_TURBO,
+ ),
+ # Test OpenSource model (take VICUNA as an example)
+ (
+ ModelPlatformType.OPENSOURCE,
+ ModelType.VICUNA,
+ OpenSourceConfig(
+ model_path="lmsys/vicuna-7b-v1.5",
+ server_url="http://localhost:8000/v1",
+ ).__dict__,
+ None,
+ OpenSourceTokenCounter,
+ ModelType.VICUNA,
+ ),
+ (
+ ModelPlatformType.OPENSOURCE,
+ ModelType.VICUNA,
+ OpenSourceConfig(
+ model_path="lmsys/vicuna-7b-v1.5",
+ server_url="http://localhost:8000/v1",
+ ).__dict__,
+ OpenAITokenCounter(ModelType.GPT_4),
+ OpenAITokenCounter,
+ ModelType.GPT_4,
+ ),
+ # Test OpenSource model (take VICUNA as an example)
+ (
+ ModelPlatformType.GEMINI,
+ ModelType.GEMINI_1_5_FLASH,
+ GeminiConfig().__dict__,
+ OpenAITokenCounter(ModelType.GPT_4),
+ OpenAITokenCounter,
+ ModelType.GPT_4,
+ ),
+ # Test Ollama model
+ (
+ ModelPlatformType.OLLAMA,
+ "gpt-3.5-turbo",
+ OllamaConfig().__dict__,
+ None,
+ OpenAITokenCounter,
+ ModelType.GPT_3_5_TURBO,
+ ),
+ (
+ ModelPlatformType.OLLAMA,
+ "gpt-3.5-turbo",
+ OllamaConfig().__dict__,
+ OpenAITokenCounter(ModelType.GPT_4),
+ OpenAITokenCounter,
+ ModelType.GPT_4,
+ ),
+ ],
+)
+
@parametrize
def test_model_factory(model_platform, model_type):