crusoe-estimator 0.1.0

PyTorch Training Cost & Carbon Estimator for Crusoe Cloud

⚡ crusoe-estimator

PyTorch Training Cost & Carbon Estimator for Crusoe Cloud

Benchmark your PyTorch training loop locally on a small sample, then get instant estimates for:

• ⏱️ Training time on every Crusoe Cloud GPU (A100, H100, H200, MI300X)
• 💰 Cost on Crusoe Cloud
• 🌍 CO₂ savings — your local emissions vs Crusoe's near-zero carbon footprint

Installation

pip install -e .

Or with PyTorch:

pip install -e ".[torch]"

Quick Start

Pattern 1: Training Loop Wrapper (Recommended)

from crusoe_estimator import CrusoeEstimator

estimator = CrusoeEstimator(total_epochs=100, location="HR")

for epoch in estimator.epochs(sample=3):
    # Your normal training code — runs only 3 epochs
    for batch_x, batch_y in train_loader:
        optimizer.zero_grad()
        output = model(batch_x)
        loss = criterion(output, batch_y)
        loss.backward()
        optimizer.step()

# Generate a beautiful HTML report
estimator.generate_report("training_report.html")

# Or print a text summary
print(estimator.summary())

Pattern 2: Decorator

estimator = CrusoeEstimator(total_epochs=100, location="DE")

@estimator.watch(sample_epochs=3)
def train_epoch(model, dataloader, optimizer, criterion):
    for x, y in dataloader:
        optimizer.zero_grad()
        loss = criterion(model(x), y)
        loss.backward()
        optimizer.step()

# Calling the decorated function runs 3 epochs and benchmarks
train_epoch(model, train_loader, optimizer, criterion)
estimator.generate_report("report.html")

Pattern 3: Context Manager

estimator = CrusoeEstimator(total_epochs=100, sample_epochs=5, location="US")

with estimator.track():
    for epoch in range(5):
        train_one_epoch(model, train_loader, optimizer, criterion)

estimator.generate_report("report.html")

Pattern 4: Function Wrapper

estimator = CrusoeEstimator(total_epochs=100, location="HR")

result = estimator.estimate(
    train_fn=lambda: train_one_epoch(model, loader, optimizer, criterion),
    sample_epochs=3,
)

estimator.generate_report("report.html")

Dataset Scaling

If you benchmark with a smaller dataset:

estimator = CrusoeEstimator(
    total_epochs=100,
    total_dataset_size=50000,
    sample_dataset_size=5000,  # Using 10% of data
    location="HR",
)

Supported GPUs

Crusoe Cloud

GPU	FP16 TFLOPS	Memory	Price/hr
A100 80GB	312	80 GB	$1.47
H100 SXM	989	80 GB	$2.25
H200 SXM	989	141 GB	$3.25
MI300X	1307	192 GB	$2.49

Local GPU Detection

Automatically detects 50+ GPUs:

• NVIDIA: RTX 20/30/40 series, A100, H100, V100, T4
• AMD: MI300X, MI250X
• Apple Silicon: M1–M4 (all variants)

How It Works

1. Detect local GPU and its specifications (FLOPS, memory, TDP)
2. Benchmark your training loop for N sample epochs
3. Extrapolate total training time using average epoch time
4. Scale to Crusoe GPUs using FLOPS ratio (70% compute + 30% memory bandwidth weighting)
5. Calculate cost based on Crusoe pricing
6. Estimate CO₂ using regional carbon intensity (local) vs near-zero (Crusoe)
7. Generate an HTML report with tables, charts, and recommendations

Carbon Intensity

Uses real-world carbon intensity data (kg CO₂/kWh):

• 🇭🇷 Croatia: 0.170
• 🇩🇪 Germany: 0.340
• 🇫🇷 France: 0.056
• 🇺🇸 USA: 0.390
• 🇨🇳 China: 0.560
• ☁️ Crusoe: ≈ 0.000 (renewable energy)

API Reference

CrusoeEstimator

Main class. Constructor parameters:

• total_epochs — how many epochs the full training would run
• sample_epochs — how many epochs to actually benchmark (default: 3)
• total_dataset_size — full dataset size (optional, for scaling)
• sample_dataset_size — dataset size used in sample (optional)
• location — country code for CO₂ calculation (e.g., "HR", "US", "DE")
• power_watts_override — override auto-detected GPU power draw

Reports

• estimator.generate_report("report.html") — HTML report
• estimator.summary() — text summary

---

Built with 💚 at HackEurope 2026 • Powered by Crusoe Cloud