litellm-impacts

Environmental impact metrics callback for LiteLLM. Calculates and exposes carbon footprint metrics (energy consumption, CO2 emissions, resource depletion) for LLM API calls using the EcoLogits library.

Installation

pip install litellm-impacts

Or with uv:

uv add litellm-impacts

Usage

With LiteLLM Proxy (config.yaml)

litellm_settings:
  callbacks: custom_callbacks.impacts_callback

model_list:
  - model_name: gpt-4
    litellm_params:
      model: openai/gpt-4

Create custom_callbacks.py in the same directory:

from litellm_impacts import ImpactsCallback

impacts_callback = ImpactsCallback()

With LiteLLM Python SDK

import litellm
from litellm_impacts import ImpactsCallback

litellm.callbacks = [ImpactsCallback()]

response = litellm.completion(
    model="gpt-4",
    messages=[{"role": "user", "content": "Hello!"}]
)

Configuration Options

from litellm_impacts import ImpactsCallback

callback = ImpactsCallback(
    prefix="litellm",        # Metric name prefix (default: "litellm")
    server_port=8000,        # Prometheus HTTP server port (default: 8000)
    start_server=True,       # Start built-in HTTP server (default: True)
    labels=["model", "key_alias"],  # Metric labels (default)
)

Set start_server=False if you're already running a Prometheus HTTP server or want to use a different metrics exposition method.

Metrics

The callback exposes the following Prometheus metrics (all with min/max variants):

Labels:

• model: The LLM model name
• key_alias: The LiteLLM API key alias (if using proxy)

Supported Providers

Models are automatically matched to EcoLogits providers:

• OpenAI (gpt-, o1, o3, o4, chatgpt-)
• Anthropic (claude-*)
• Google (gemini-*)
• Mistral AI (mistral-, codestral-, pixtral-*)
• Cohere (command-*)

Models can also be specified with explicit provider prefixes (e.g., openai/gpt-4, anthropic/claude-3-sonnet).

License

MIT