# GPU Requirements for OWSM v3.1 Space

## Recommended Hardware

### Minimum (Inference Only)
- **GPU**: T4 (16GB) - Available on HF Spaces
- **Use Case**: Running inference on test set
- **Performance**: ~2-5x real-time factor

### Recommended (Training + Inference)
- **GPU**: A10 Small (24GB) or A10 Large (40GB)
- **Use Case**: Fine-tuning OWSM v3.1 model
- **Performance**: Training ~4-8 hours for 3 epochs

### CPU Fallback
- **Possible**: Yes, but very slow
- **Inference**: ~10-20x real-time factor
- **Training**: Not practical (days/weeks)

## Model Sizes

- **OWSM v3.1 Small**: 367M parameters (~1.5GB)
- **OWSM v3.1 Medium**: ~1B parameters (~4GB)
- **Memory Usage**: ~2-4GB GPU memory for inference
- **Training Memory**: ~8-16GB GPU memory (with batch size 4)

## Space Configuration

The `README.md` includes hardware specification:
```yaml
hardware: gpu-a10g-large
```

### Available GPU Options in HF Spaces:

1. **gpu-t4** (Recommended minimum)
   - 16GB GPU memory
   - Good for inference
   - Can handle small batch training

2. **gpu-a10-small**
   - 24GB GPU memory
   - Better for training
   - Faster inference

3. **gpu-a10-large**
   - 40GB GPU memory
   - Best for training
   - Can use larger batch sizes

### To Change GPU:

1. Edit `README.md` hardware line
2. Or use Space Settings → Hardware → Select GPU

## Pricing (Hourly Rates)

| Hardware | GPU | VRAM | Hourly Price | Daily (24h) | Weekly (7d) |
|----------|-----|------|--------------|-------------|-------------|
| **T4 - small** | T4 | 16 GB | **$0.40** | $9.60 | $67.20 |
| **T4 - medium** | T4 | 16 GB | **$0.60** | $14.40 | $100.80 |
| **A10G - small** | A10G | 24 GB | **$1.00** | $24.00 | $168.00 |
| **A10G - large** | A10G | 24 GB | **$1.50** | $36.00 | $252.00 |

**Note**: You only pay for time the Space is running with GPU active.

## Performance Estimates

### Inference (8,510 test files):

| Hardware | Time Estimate | Cost Estimate | Real-time Factor |
|----------|---------------|---------------|------------------|
| CPU | ~4-8 hours | FREE | 10-20x |
| T4 GPU | ~30-60 minutes | **$0.20-$0.60** | 2-5x |
| A10 Small | ~15-30 minutes | **$0.25-$0.50** | 1-2x |
| A10 Large | ~15-30 minutes | **$0.38-$0.75** | 1-2x |

### Training (19,856 samples, 3 epochs):

| Hardware | Time Estimate | Cost Estimate | Batch Size |
|----------|---------------|---------------|------------|
| T4 GPU | ~6-8 hours | **$2.40-$4.80** | 4 |
| A10 Small | ~4-6 hours | **$4.00-$6.00** | 8 |
| A10 Large | ~3-4 hours | **$4.50-$6.00** | 16 |

## Current Configuration

- **Hardware**: `gpu-a10g-large` (set in README.md)
- **Model**: OWSM v3.1 Small (367M)
- **Batch Size**: 8-16 (training), dynamic (inference)

## Notes

- GPU is **required** for practical training
- GPU is **highly recommended** for inference (much faster)
- CPU will work but be very slow
- T4 is sufficient for inference and small-scale training