GREENPEFT · v1.0 · MIT OPEN SOURCE

Choose how to fine-tune an LLM before you spend the energy to find out.

Zero-shot surrogate regressors estimate accuracy, peak VRAM, energy, carbon, and runtime before any benchmark sweep begins.

Python 3.10+ PyTorch 2.2+ CodeCarbon MIT PyPI · green-peft Hugging Face
Pareto Frontier · Accuracy vs VRAM 16.0 GB (13 feasible)
GPU Presets:
Methods
Drag the vertical dashed line or click a GPU preset Hover points for telemetry
01 / Decision Engine

A decision engine you can turn the dials on.

Every prediction below updates live in your browser. Move a constraint slider or change preference weights — the feasibility gate, Pareto filtering, and Green Efficiency Index (GEI) recompute instantly with zero latency.

Hardware & Objectives
16 GB

Target GPU ceiling. 10% safety margin applied (usable: 14.4 GB).

0.90

Performance floor on SST-2 text classification.

0.25 kg

Max estimated kgCO₂e emissions per run.

60 min

Maximum acceptable wall-clock training duration.

Preference Profile

Active weights: 35% Accuracy · 25% Memory · 25% Carbon · 15% Time

RECOMMENDED CANDIDATE

LISA · Qwen2.5-Medium (1.5B)

Optimal multi-objective Pareto trade-off: high accuracy within 16 GB VRAM budget with low memory activation overhead.

Predicted Accuracy
0.923
Peak VRAM
14.2GB
Energy
0.18kWh
Carbon
0.09kg
Runtime
42min
GEI Score
0.847
Feasibility
✓ FEASIBLE
Confidence
Medium
GEI Relative Score Comparison (Top 3)
Alternative Candidates
Predictions are surrogate estimates, not guarantees. Validate with a real run.

NO FEASIBLE CANDIDATE — RELAX VRAM OR ACCURACY.

No parameter-efficient fine-tuning configuration simultaneously satisfies your current hardware ceilings and performance floor.

02 / Motivation

Why pre-run decision support matters.

Parameter-efficient fine-tuning has dozens of tuning dimensions. Brute-force sweeps waste compute on configurations destined to fail.

Trial and error is expensive.

A standard grid sweep evaluating 4 model backbones, 5 PEFT methods, 3 learning rates, and 3 random seeds requires 180 full fine-tuning runs.

Running blind sweeps on cloud GPUs wastes substantial budget and accumulates hundreds of kilograms of carbon before you discover that large rank adaptations trigger Out-Of-Memory (OOM) failures.

1 sweep ≈ $340 + 12h
Based on 180-run baseline sweep on Tesla T4 cluster (45.2 kWh)

Objectives conflict.

Single-metric optimization creates severe unintended trade-offs. Chasing peak accuracy demands large model backbones and unquantized weights, driving up GPU memory requirements and energy draw.

Conversely, aggressive 4-bit quantization reduces memory footprint but can introduce performance degradation on complex domain tasks. GreenPEFT formalizes this Pareto frontier.

Downstream Accuracy Demands scale (↑ parameters, ↑ rank)
Peak VRAM Footprint Capped by GPU memory (16 GB ceiling)
Carbon & Energy Accumulates over training hours

Filter first. Validate second.

Rather than training every candidate, GreenPEFT evaluates all combinations through zero-shot surrogate models, applies physical constraints, and isolates the non-dominated Pareto frontier.

03 / Architecture

Four-stage research pipeline.

GreenPEFT follows the research paper's four core stages: benchmark, audit and coverage, modeling, and decision. Click any stage to inspect its inputs, measurements, validation gates, and outputs.

STAGE 01
Benchmark
STAGE 02
Audit & Coverage
STAGE 03
Modeling
STAGE 04
Decision Engine
GreenPEFT four-stage research pipeline Benchmark measurements flow through audit and coverage, four Ridge surrogate models, and a constraint-aware decision engine. STAGE 1 BENCHMARK 4 backbones × 5 PEFT methods SST-2 · Tesla T4 · 300 steps 3 seeds · fixed setup Accuracy · VRAM · Energy Runtime · CO₂e STAGE 2 AUDIT & COVERAGE 59.9 W retained under load 9.97 W idle floor excluded 41 valid configurations 14 / 20 matrix cells covered canonical telemetry dataset STAGE 3 MODELING 22 pre-run metadata features 4 Ridge surrogate models Accuracy · VRAM · Energy · Runtime Pass / seed / LOTO validation Energy gate: R² 0.90 · MAPE 10.6% STAGE 4 DECISION ENGINE 70 candidates → sanity filter constraints + 10% VRAM margin Pareto → GEI ranking Measured / interpolated evidence DOIs · HF · Kaggle · CLI
GreenPEFT research framework: 41 audited configurations become validated, constraint-aware recommendations. View source research figure

Normalizes heterogeneous training traces and hardware telemetry into standard units. Eliminates measurement divergence across different tracking libraries.

Empirical Ground Truth

Calibrated benchmark evidence.

Accuracy vs Peak VRAM (60 Runs)
Shapes: LoRA (▲), QLoRA (▼), LISA (■), LoRA-FA (◆), Full-FT (●)
Method Comparison (Tesla T4 · SST-2)
Method Accuracy VRAM Energy Carbon Time GEI
Full-FT (0.5B) 0.905 11.1 GB 0.0024 kWh 0.0016 kg 136s 0.490
LoRA (1.1B) 0.944 9.4 GB 0.0018 kWh 0.0012 kg 107s 0.792
QLoRA (1.1B) 0.938 4.0 GB 0.0027 kWh 0.0018 kg 162s 0.789
LoRA-FA (1.1B) 0.925 8.4 GB 0.0017 kWh 0.0011 kg 102s 0.712
LISA (1.5B) 0.938 15.2 GB 0.0025 kWh 0.0016 kg 141s 0.585
LISA fits a 1.5B model in 16 GB VRAM.
60-run NVIDIA Tesla T4 · SST-2 · 0.5B–3B params · Extrapolated entries marked with *
04 / Transparency

Trust & limitations.

Rigorous research requires explicit boundaries. GreenPEFT makes no inflated claims about optimality or universal generalization.

Always validate a selected candidate with a real run.
06 / CLI

CLI quickstart.

Integrate GreenPEFT directly into your pre-flight CI/CD pipelines and experiment scripts.

BASH · GREEN-PEFT CLI
$ pip install green-peft

$ green-peft recommend \
    --artifacts-dir ./peft_bench_export \
    --vram 16 \
    --accuracy 0.90 \
    --profile balanced
____ ____ _____ _____ _____ / ___|_ __ ___ ___ _ __ | _ \| ____| ___|_ _| | | _| '__/ _ \/ _ \ '_ \| |_) | _| | |_ | | | |_| | | | __/ __/ | | | __/| |___| _| | | \____|_| \___|\___|_| |_|_| |_____|_| |_| GreenPEFT v0.3.0 - sustainability-aware PEFT decision support Ashraful Islam Tanzil | United International University https://github.com/ai-tanzil811 surrogate status: [ok] VALIDATED data doi 10.34740/KAGGLE/DSV/20178095 | model doi 10.57967/hf/10690 → Filtering 47 candidates... → 6 feasible under constraints. → Pareto frontier: 3 candidates. RECOMMENDED method LISA backbone Qwen2.5-Medium (1.5B) accuracy 0.923 ± 0.011 peak_vram 14.2 GB energy 0.18 kWh carbon 0.09 kgCO₂e runtime 42 min GEI 0.847 Alternatives: QLoRA + TinyLlama-Small GEI 0.791 LoRA-FA + Qwen2.5-Tiny GEI 0.744

The CLI banner is written to stderr so --json output stays pipe-safe. In Windows CMD, append 2>&1 when you need to merge it into stdout.

06 / Reference

Author, links & citation.

Ashraful Islam Tanzil

AUTHOR & DEVELOPER HUB

Ashraful Islam Tanzil

Green AI & Efficient LLM Systems

✓ Author & Lead Researcher ✓ United International University ✓ Green AI & PEFT Specialist

Computer Science researcher building transparent decision frameworks for parameter-efficient fine-tuning, resource-aware experimentation, and empirical carbon accounting.

Code artifactsPyPI v0.3.0 · HF v2pip install green-peft
Benchmark grounding41 audited runsTesla T4 · SST-2
Research artifactsDOI · Kaggle · GitHubOpen-source evidence trail

PROJECT DIRECTORY

Explore the evidence stack.

Ashraful Islam Tanzil

Ashraful Islam Tanzil

Researcher · Green AI & Efficient LLM Systems

Machine learning researcher working on Green AI decision frameworks, parameter-efficient fine-tuning efficiency, and empirical carbon accounting for large language models.

Author Report

Ashraful Islam Tanzil

View source report

Computer Science undergraduate at United International University, passionate about web development and machine learning. Focused on React.js, AI, full-stack web development, and machine learning applications.

Tech stack

Languages
Python, JavaScript, HTML, CSS, C, C++
Frameworks
React, Node.js
Tools
Git, VS Code, npm
Databases
MySQL

GitHub stats

BibTeX Citation
@software{tanzil_greenpeft_2025,
  author  = {Tanzil, Ashraful Islam},
  title   = {GreenPEFT: A Multi-Objective Green AI Decision Support Framework},
  year    = {2025},
  license = {MIT},
  url     = {https://github.com/ai-tanzil811/Green-AI-Decision-Support-Frameworks}
}