greentensor 0.1.0

Energy-efficient middleware for PyTorch ML workloads — tracks carbon, optimizes GPU, reports savings.

GreenTensor

Energy-efficient middleware for PyTorch ML workloads.

GreenTensor wraps your training loop with one line of code and gives you:

• Real carbon emissions and energy tracking (via CodeCarbon)
• Automatic GPU optimizations (cuDNN benchmark, mixed precision)
• Idle GPU detection to reduce wasted energy
• A detailed savings report comparing against your baseline

---

Install

pip install greentensor

---

Quickstart

from greentensor import GreenTensor

with GreenTensor() as gt:
    with gt.mixed_precision():
        train()  # your existing training code, unchanged

Output:

  +======================================+
  |        GreenTensor Report            |
  +======================================+
  Runtime          : 12.34 s
  Energy Used      : 0.000412 kWh
  CO2 Emissions    : 0.000096 kg
  ======================================

---

Compare against a baseline

import pickle
from greentensor import GreenTensor
from greentensor.core.tracker import Tracker
from greentensor.report.metrics import RunMetrics
import time

# 1. Run baseline (no optimizations)
tracker = Tracker()
tracker.start()
t0 = time.perf_counter()
train()
duration = time.perf_counter() - t0
emissions_kg, energy_kwh = tracker.stop()
baseline = RunMetrics(duration_s=duration, energy_kwh=energy_kwh, emissions_kg=emissions_kg)

# 2. Run optimized — report shows real savings
with GreenTensor(baseline=baseline) as gt:
    with gt.mixed_precision():
        train()

---

Batch size optimizer

from greentensor import optimize_batch_size

batch_size = optimize_batch_size(32)  # auto-scales based on available GPU memory

---

Dashboard

streamlit run dashboard/app.py

Enter baseline and optimized metrics manually, or upload .pkl files saved from your runs.

---

Configuration

from greentensor import GreenTensor, Config

config = Config(
    enable_cudnn_benchmark=True,
    enable_mixed_precision=True,
    idle_threshold_pct=10.0,   # GPU util % below which idle is detected
    max_batch_size=512,
)

with GreenTensor(config=config) as gt:
    train()

---

Run tests

pip install pytest
pytest tests/ -v

---

License

MIT