Search by

davidjln / llm-carbon-php

DavidJLN

Calculation of the energy consumed and the CO2eq emissions of an inference request to a language model (LLM), following the EcoLogits v0.4.0 methodology.

Package info

github.com/DavidJLN/llm-carbon-php

Type:project

pkg:composer/davidjln/llm-carbon-php

Security

Advisories: 0

Aikido package health analysis

Transparency log

Statistics

Installs: 9

Dependents: 1

Suggesters: 0

Stars: 0

Open Issues: 0

v1.0.0 2026-08-31 15:21 UTC

Requires

  • php: >=8.4

Requires (Dev)

Suggests

None

Provides

None

Conflicts

None

Replaces

None

MIT a430d0a48c231d8eb94aab77f498360414398e75

Published on Packagist.org on 2026-09-10 14:53 UTC

This package is auto-updated.

Last update: 2026-11-09 23:40:15 UTC

README

This document describes davidjln/llm-carbon-php as a Composer package to install in your own code. For the standalone web demo of this repository (HTML page displaying a hardcoded scenario), see README-demo-english.md.

What this package does, and for whom

This package calculates the energy consumed and the CO2eq emissions of an inference request to a language model (LLM), from three inputs: a model (number of parameters), an emission factor (geographic hosting zone of the datacenter), and a number of generated tokens. The calculation follows the EcoLogits v0.4.0 methodology.

It targets PHP developers who want to integrate a carbon estimate into their own application (dashboard, logging, reporting) rather than using the demonstration page provided by this repository — for example to compute, for each LLM call made by their application, the energy and emissions associated with that specific call.

Installation

Minimum PHP version required: 8.4 (the package uses readonly properties).

This is due to minimal version required by PHPUnit 13 and Infection.

composer require davidjln/llm-carbon-php

This package has no runtime dependency: composer require only installs the package itself.

Minimal usage

<?php

require 'vendor/autoload.php';

use LlmCarbon\EmissionFactor;
use LlmCarbon\FootprintCalculatorSimplified;
use LlmCarbon\LanguageModel;

$footprint = (new FootprintCalculatorSimplified())->calculate(
    LanguageModel::llama31_70b(),
    EmissionFactor::france(),
    500, // number of tokens generated in the response
);

echo $footprint->emissionsGco2eq, ' gCO2eq';

FootprintCalculatorSimplified::calculate() returns a Footprint (src/Footprint.php) exposing three values: energyPerTokenWh, totalEnergyWh, and emissionsGco2eq. LanguageModel and EmissionFactor each expose one factory per catalog value (see all() on each class for the full list: models llama31_70b(), gpt4(), gpt4o(), qwen3_235b_a22b(), kimiK2(), mistralLarge2(), grok1(), claudeSonnet46(), claudeOpus48(), claudeHaiku45(); zones france(), europe(), unitedStates(), world()).

A second implementation, FootprintCalculatorFull, has the same calculate() signature and additionally accounts for the GPU memory required and the number of cards needed to load the model; see README-demo-english.md for details on the difference between the two.

What the calculation covers

The scope is strictly limited to inference, based solely on the number of tokens generated in output:

  • the GPU energy consumed to generate the response tokens (EcoLogits regression on the model's active parameters);
  • with FootprintCalculatorFull only, the non-GPU server energy associated with that same generation;
  • the conversion of this energy into CO2eq emissions, via the emission factor of the electricity mix of the chosen zone.

What the calculation does not cover

  • Input tokens: the prompt sent to the model is not part of the calculation in any way; only the number of tokens generated is used.
  • Model training: the energy and emissions related to training (or fine-tuning) are not counted — only inference is.
  • Hardware manufacturing: emissions related to the manufacturing of GPUs and servers ("embodied" impact, upstream of their entry into service) are not counted — only the energy consumed while executing the request is.
  • Storage and network: neither the energy for storing the model weights nor that of the network transport of the request or response is counted.
  • Uncertainty is not a statistical confidence interval: each input value cites its Provenance (measured and published by the model provider, or a reconstructed hypothesis for lack of publication — see src/ProvenanceType.php) and the result that depends on it inherits this status, but the package does not compute any margin of error or output range: for the proprietary models in the catalog (GPT-4, GPT-4o, Claude Sonnet 4.6, Claude Opus 4.8, Claude Haiku 4.5), whose parameters are not published, a conservative hypothesis (lower bound) is retained rather than a range.

Sources and years

Among the providers of the catalog, only Mistral AI publishes a measured footprint: its life-cycle analysis of Mistral Large 2, 2025-07-22 (with Carbone 4 and ADEME) reports 1.14 gCO2eq, 45 mL of water and 0.16 mg Sb eq for a 400-token response, excluding user terminals. It is cited here as a point of comparison only, not as an input of the calculation: its scope (full life cycle, hardware manufacturing and training included) differs from this project's (inference energy only).

The full detail of each source (URL, year, exactly what it states) is accessible from the code via LanguageModel::$provenance / $totalParametersProvenance and EmissionFactor::$provenance — see README-demo-english.md for the summary and detailed limitations.