Merge pull request #17 from dataforgoodfr/feat/model-repository

Add model repository

genai_impact/compute_impacts.py

@@ -1,15 +1,31 @@
+from typing import Union
+
 from pydantic import BaseModel
 
-ENERGY_PROFILE = 1.17e-4
+GPU_ENERGY_ALPHA = 8.91e-5
+GPU_ENERGY_BETA = 1.43e-3
 
 
 class Impacts(BaseModel):
     energy: float
+    energy_range: tuple[float, float]
     energy_unit: str = "Wh"
 
 
 def compute_llm_impact(
-    model_parameter_count: float,
+    model_parameter_count: Union[float, tuple[float, float]],
     output_token_count: int,
 ) -> Impacts:
-    return Impacts(energy=ENERGY_PROFILE * model_parameter_count * output_token_count)
+    if isinstance(model_parameter_count, (tuple, list)):
+        # TODO: check tuple or list validity (min, max)
+        energy_min = output_token_count * (GPU_ENERGY_ALPHA * model_parameter_count[0] + GPU_ENERGY_BETA)
+        energy_max = output_token_count * (GPU_ENERGY_ALPHA * model_parameter_count[1] + GPU_ENERGY_BETA)
+        energy_avg = (energy_min + energy_max) / 2
+    else:
+        energy_avg = output_token_count * (GPU_ENERGY_ALPHA * model_parameter_count + GPU_ENERGY_BETA)
+        energy_min = energy_avg
+        energy_max = energy_avg
+    return Impacts(
+        energy=energy_avg,
+        energy_range=(energy_min, energy_max)
+    )

genai_impact/data/models.csv

@@ -0,0 +1,35 @@
+provider,name,total_parameters,active_parameters,warnings,sources
+openai,gpt-3.5-turbo,20;70,20;70,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-3.5-turbo-0125,20;70,20;70,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-3.5-turbo-0301,20;70,20;70,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-3.5-turbo-0613,20;70,20;70,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-3.5-turbo-1106,20;70,20;70,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-3.5-turbo-16k,20;70,20;70,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-3.5-turbo-16k-0613,20;70,20;70,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-3.5-turbo-instruct,20;70,20;70,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-3.5-turbo-instruct-0914,20;70,20;70,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-4,1760,220;880,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-4-0125-preview,1760,220;880,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-4-0613,1760,220;880,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-4-1106-preview,1760,220;880,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-4-turbo-preview,1760,220;880,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+openai,gpt-4-vision-preview,1760,220;880,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit
+mistralai,mistral-large-2402,540,135;540,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+mistralai,mistral-large-latest,540,135;540,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+mistralai,mistral-medium,70;180,45;180,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+mistralai,mistral-medium-2312,70;180,45;180,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+mistralai,mistral-medium-latest,70;180,45;180,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+mistralai,mistral-small,46.7,12.9,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+mistralai,mistral-small-2312,46.7,12.9,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+mistralai,mistral-small-2402,46.7,12.9,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+mistralai,mistral-small-latest,46.7,12.9,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+mistralai,mistral-tiny,7.3,7.3,,https://docs.mistral.ai/models/#sizes 
+mistralai,mistral-tiny-2312,7.3,7.3,,https://docs.mistral.ai/models/#sizes 
+mistralai,open-mistral-7b,7.3,7.3,,https://docs.mistral.ai/models/#sizes 
+mistralai,open-mixtral-8x7b,46.7,46.7,,https://docs.mistral.ai/models/#sizes 
+anthropic,claude-3-opus-20240229,2000,250;1000,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+anthropic,claude-3-sonnet-20240229,800,100;400,model_architecture_not_released,
+anthropic,claude-3-haiku-20240307,300,75;150,model_architecture_not_released,
+anthropic,claude-2.1,130,130,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+anthropic,claude-2.0,130,130,model_architecture_not_released,https://docs.google.com/spreadsheets/d/1O5KVQW1Hx5ZAkcg8AIRjbQLQzx2wVaLl0SqUu-ir9Fs/edit?usp=sharing
+anthropic,claude-instant-1.2,20;70,20;70,model_architecture_not_released,

genai_impact/model_repository.py

@@ -0,0 +1,84 @@
+import os
+from csv import DictReader
+from dataclasses import dataclass
+from enum import Enum
+from typing import Optional
+
+
+class Providers(Enum):
+    anthropic = "anthropic"
+    mistralai = "mistralai"
+    openai = "openai"
+
+
+class Warnings(Enum):
+    model_architecture_not_released = "model_architecture_not_released"
+
+
+@dataclass
+class Model:
+    provider: str
+    name: str
+    total_parameters: Optional[float] = None
+    active_parameters: Optional[float] = None
+    total_parameters_range: Optional[tuple[float, float]] = None
+    active_parameters_range: Optional[tuple[float, float]] = None
+    warnings: Optional[list[str]] = None
+    sources: Optional[list[str]] = None
+
+
+class ModelRepository:
+
+    def __init__(self, models: list[Model]) -> None:
+        self.__models = models
+
+    def find_model(self, provider: str, model_name: str) -> Optional[Model]:
+        for model in self.__models:
+            if model.provider == provider and model_name == model.name:
+                return model
+        return None
+
+    @classmethod
+    def from_csv(cls, filepath: Optional[str] = None) -> "ModelRepository":
+        if filepath is None:
+            filepath = os.path.join(os.path.dirname(os.path.realpath(__file__)), "data", "models.csv")
+        models = []
+        with open(filepath) as fd:
+            csv = DictReader(fd)
+            for row in csv:
+                total_parameters = None
+                total_parameters_range = None
+                if ";" in row["total_parameters"]:
+                    total_parameters_range = [float(p) for p in row["total_parameters"].split(";")]
+                elif row["total_parameters"] != "":
+                    total_parameters = float(row["total_parameters"])
+
+                active_parameters = None
+                active_parameters_range = None
+                if ";" in row["active_parameters"]:
+                    active_parameters_range = [float(p) for p in row["active_parameters"].split(";")]
+                elif row["active_parameters"] != "":
+                    active_parameters = float(row["active_parameters"])
+
+                warnings = None
+                if row["warnings"] != "":
+                    warnings = [Warnings(w).name for w in row["warnings"].split(";")]
+
+                sources = None
+                if row["sources"] != "":
+                    sources = row["sources"].split(";")
+
+                models.append(Model(
+                    provider=Providers(row["provider"]).name,
+                    name=row["name"],
+                    total_parameters=total_parameters,
+                    active_parameters=active_parameters,
+                    total_parameters_range=total_parameters_range,
+                    active_parameters_range=active_parameters_range,
+                    warnings=warnings,
+                    sources=sources
+                ))
+        return cls(models)
+
+
+models = ModelRepository.from_csv()

genai_impact/tracers/anthropic_tracer.py

@@ -3,6 +3,7 @@
 from wrapt import wrap_function_wrapper
 
 from genai_impact.compute_impacts import Impacts, compute_llm_impact
+from genai_impact.model_repository import models
 
 try:
     from anthropic import Anthropic as _Anthropic
@@ -11,31 +12,26 @@
     _Anthropic = object()
     _Message = object()
 
-_MODEL_SIZES = {
-    "claude-3-haiku-20240307": 10,
-    "claude-3-sonnet-20240229": 10,  # fake data
-    "claude-3-opus-20240229": 440,  # fake data
-}
-
 
 class Message(_Message):
     impacts: Impacts
 
 
-def _set_impacts(response: Message) -> Impacts:
-    model_size = _MODEL_SIZES.get(response.model)
-    output_tokens = response.usage.output_tokens
-    impacts = compute_llm_impact(
-        model_parameter_count=model_size, output_token_count=output_tokens
-    )
-    return impacts
-
-
 def anthropic_chat_wrapper(
     wrapped: Callable, instance: _Anthropic, args: Any, kwargs: Any  # noqa: ARG001
 ) -> Message:
     response = wrapped(*args, **kwargs)
-    impacts = _set_impacts(response)
+    model = models.find_model(provider="anthropic", model_name=response.model)
+    if model is None:
+        # TODO: Replace with proper logging
+        print(f"Could not find model `{response.model}` for anthropic provider.")
+        return response
+    output_tokens = response.usage.output_tokens
+    model_size = model.active_parameters or model.active_parameters_range
+    impacts = compute_llm_impact(
+        model_parameter_count=model_size,
+        output_token_count=output_tokens
+    )
     return Message(**response.model_dump(), impacts=impacts)
 
 

genai_impact/tracers/mistralai_tracer.py

@@ -3,6 +3,7 @@
 from wrapt import wrap_function_wrapper
 
 from genai_impact.compute_impacts import Impacts, compute_llm_impact
+from genai_impact.model_repository import models
 
 try:
     from mistralai.client import MistralClient as _MistralClient
@@ -14,14 +15,6 @@
     _ChatCompletionResponse = object()
 
 
-_MODEL_SIZES = {
-    "mistral-tiny": 7.3,
-    "mistral-small": 12.9,  # mixtral active parameters count
-    "mistral-medium": 70,
-    "mistral-large": 440,
-}
-
-
 class ChatCompletionResponse(_ChatCompletionResponse):
     impacts: Impacts
 
@@ -30,10 +23,16 @@ def mistralai_chat_wrapper(
     wrapped: Callable, instance: _MistralClient, args: Any, kwargs: Any  # noqa: ARG001
 ) -> ChatCompletionResponse:
     response = wrapped(*args, **kwargs)
-    model_size = _MODEL_SIZES.get(response.model)
+    model = models.find_model(provider="mistralai", model_name=response.model)
+    if model is None:
+        # TODO: Replace with proper logging
+        print(f"Could not find model `{response.model}` for mistralai provider.")
+        return response
     output_tokens = response.usage.completion_tokens
+    model_size = model.active_parameters or model.active_parameters_range
     impacts = compute_llm_impact(
-        model_parameter_count=model_size, output_token_count=output_tokens
+        model_parameter_count=model_size,
+        output_token_count=output_tokens
     )
     return ChatCompletionResponse(**response.model_dump(), impacts=impacts)
 

genai_impact/tracers/openai_tracer.py

@@ -5,26 +5,7 @@
 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": 440,
-    "gpt-4-0314": 440,
-    "gpt-4-0613": 440,
-    "gpt-4-32k": 440,
-    "gpt-4-32k-0314": 440,
-    "gpt-4-32k-0613": 440,
-    "gpt-3.5-turbo": 70,
-    "gpt-3.5-turbo-16k": 70,
-    "gpt-3.5-turbo-0301": 70,
-    "gpt-3.5-turbo-0613": 70,
-    "gpt-3.5-turbo-1106": 70,
-    "gpt-3.5-turbo-0125": 70,
-    "gpt-3.5-turbo-16k-0613": 70,
-}
+from genai_impact.model_repository import models
 
 
 class ChatCompletion(_ChatCompletion):
@@ -35,10 +16,16 @@ def openai_chat_wrapper(
     wrapped: Callable, instance: Completions, args: Any, kwargs: Any  # noqa: ARG001
 ) -> ChatCompletion:
     response = wrapped(*args, **kwargs)
-    model_size = _MODEL_SIZES.get(response.model)
+    model = models.find_model(provider="openai", model_name=response.model)
+    if model is None:
+        # TODO: Replace with proper logging
+        print(f"Could not find model `{response.model}` for openai provider.")
+        return response
     output_tokens = response.usage.completion_tokens
+    model_size = model.active_parameters or model.active_parameters_range
     impacts = compute_llm_impact(
-        model_parameter_count=model_size, output_token_count=output_tokens
+        model_parameter_count=model_size,
+        output_token_count=output_tokens
     )
     return ChatCompletion(**response.model_dump(), impacts=impacts)
 

pyproject.toml

@@ -89,7 +89,8 @@ ignore = [
     "PLR0913",
     "RET504",
     "RET505",
-    "COM812"
+    "COM812",
+    "PTH"
 ]
 
 fixable = [

tests/test_model_repository.py

@@ -0,0 +1,24 @@
+from genai_impact.model_repository import ModelRepository, Model
+
+
+def test_create_model_repository_default():
+    models = ModelRepository.from_csv()
+    assert isinstance(models, ModelRepository)
+    assert models.find_model(provider="openai", model_name="gpt-3.5-turbo") is not None
+
+
+def test_create_model_repository_from_scratch():
+    models = ModelRepository([
+        Model(provider="provider-test", name="model-test")
+    ])
+    assert models.find_model(provider="provider-test", model_name="model-test")
+
+
+def test_find_unknown_provider():
+    models = ModelRepository.from_csv()
+    assert models.find_model(provider="provider-test", model_name="gpt-3.5-turbo") is None
+
+
+def test_find_unknown_model_name():
+    models = ModelRepository.from_csv()
+    assert models.find_model(provider="openai", model_name="model-test") is None