Merge pull request #12 from dataforgoodfr/feat/mistral-7

Add support for MistralAI client

README.md

@@ -63,7 +63,7 @@ pip install poetry
 ### Install project
 
 ```shell
-poetry install --with dev,docs
+poetry install --all-extras --with dev,docs
 ```
 
 

docs/index.md

@@ -12,7 +12,7 @@ pip install genai_impact
 ## Basic example for OpenAI
 
 ```python
-from genai_impact.client_wrapper import OpenAI
+from genai_impact import OpenAI
 
 client = OpenAI()
 

genai_impact/__init__.py

@@ -1,3 +1,3 @@
-from genai_impact.client_wrapper import OpenAI
+from genai_impact.client import MistralClient, OpenAI
 
-__all__ = ["OpenAI"]
+__all__ = ["OpenAI", "MistralClient"]

genai_impact/client/__init__.py

@@ -0,0 +1,4 @@
+from .mistralai_wrapper import MistralClient
+from .openai_wrapper import OpenAI
+
+__all__ = ["OpenAI", "MistralClient"]

genai_impact/client/mistralai_wrapper.py

@@ -0,0 +1,45 @@
+from typing import Any, Callable
+
+from wrapt import wrap_function_wrapper
+
+from genai_impact.compute_impacts import Impacts, compute_llm_impact
+
+try:
+    from mistralai.client import MistralClient as _MistralClient
+    from mistralai.models.chat_completion import (
+        ChatCompletionResponse as _ChatCompletionResponse,
+    )
+except ImportError:
+    _MistralClient = object()
+    _ChatCompletionResponse = object()
+
+
+_MODEL_SIZES = {
+    "mistral-tiny": 7.3,
+    "mistral-small": 12.9,  # mixtral active parameters count
+    "mistral-medium": 70,
+    "mistral-large": 220,
+}
+
+
+class ChatCompletionResponse(_ChatCompletionResponse):
+    impacts: Impacts
+
+
+def chat_wrapper(
+    wrapped: Callable, instance: _MistralClient, args: Any, kwargs: Any  # noqa: ARG001
+) -> ChatCompletionResponse:
+    response = wrapped(*args, **kwargs)
+    model_size = _MODEL_SIZES.get(response.model)
+    output_tokens = response.usage.completion_tokens
+    impacts = compute_llm_impact(
+        model_parameter_count=model_size, output_token_count=output_tokens
+    )
+    return ChatCompletionResponse(**response.model_dump(), impacts=impacts)
+
+
+class MistralClient(_MistralClient):
+    def __init__(self, **kwargs: Any) -> None:
+        super().__init__(**kwargs)
+
+    wrap_function_wrapper("mistralai.client", "MistralClient.chat", chat_wrapper)

genai_impact/client/openai_wrapper.py

@@ -0,0 +1,53 @@
+from typing import Any, Callable
+
+from openai import OpenAI as _OpenAI
+from openai.resources.chat import Completions
+from openai.types.chat import ChatCompletion as _ChatCompletion
+from wrapt import wrap_function_wrapper
+
+from genai_impact.compute_impacts import Impacts, compute_llm_impact
+
+_MODEL_SIZES = {
+    "gpt-4-0125-preview": None,
+    "gpt-4-turbo-preview": None,
+    "gpt-4-1106-preview": None,
+    "gpt-4-vision-preview": None,
+    "gpt-4": 220,
+    "gpt-4-0314": 220,
+    "gpt-4-0613": 220,
+    "gpt-4-32k": 220,
+    "gpt-4-32k-0314": 220,
+    "gpt-4-32k-0613": 220,
+    "gpt-3.5-turbo": 20,
+    "gpt-3.5-turbo-16k": 20,
+    "gpt-3.5-turbo-0301": 20,
+    "gpt-3.5-turbo-0613": 20,
+    "gpt-3.5-turbo-1106": 20,
+    "gpt-3.5-turbo-0125": 20,
+    "gpt-3.5-turbo-16k-0613": 20,
+}
+
+
+class ChatCompletion(_ChatCompletion):
+    impacts: Impacts
+
+
+def chat_wrapper(
+    wrapped: Callable, instance: Completions, args: Any, kwargs: Any  # noqa: ARG001
+) -> ChatCompletion:
+    response = wrapped(*args, **kwargs)
+    model_size = _MODEL_SIZES.get(response.model)
+    output_tokens = response.usage.completion_tokens
+    impacts = compute_llm_impact(
+        model_parameter_count=model_size, output_token_count=output_tokens
+    )
+    return ChatCompletion(**response.model_dump(), impacts=impacts)
+
+
+class OpenAI(_OpenAI):
+    def __init__(self, **kwargs: Any) -> None:
+        super().__init__(**kwargs)
+
+    wrap_function_wrapper(
+        "openai.resources.chat.completions", "Completions.create", chat_wrapper
+    )

genai_impact/compute_impacts.py

@@ -1,30 +1,9 @@
-from dataclasses import dataclass
+from pydantic import BaseModel
 
 ENERGY_PROFILE = 1.17e-4
 
-MODEL_SIZES = {
-    "gpt-4-0125-preview": None,
-    "gpt-4-turbo-preview": None,
-    "gpt-4-1106-preview": None,
-    "gpt-4-vision-preview": None,
-    "gpt-4": 200,
-    "gpt-4-0314": 200,
-    "gpt-4-0613": 200,
-    "gpt-4-32k": 200,
-    "gpt-4-32k-0314": 200,
-    "gpt-4-32k-0613": 200,
-    "gpt-3.5-turbo": 20,
-    "gpt-3.5-turbo-16k": 20,
-    "gpt-3.5-turbo-0301": 20,
-    "gpt-3.5-turbo-0613": 20,
-    "gpt-3.5-turbo-1106": 20,
-    "gpt-3.5-turbo-0125": 20,
-    "gpt-3.5-turbo-16k-0613": 20,
-}
 
-
-@dataclass
-class Impacts:
+class Impacts(BaseModel):
     energy: float
     energy_unit: str = "Wh"