Metadata-Version: 2.1
Name: quarterbit
Version: 20.0.1
Summary: Memory-efficient LLM training. AXIOM enables 70B on single H100, 7B on consumer GPUs.
Home-page: https://quarterbit.dev
Author: Clouthier Simulation Labs
Author-email: Clouthier Simulation Labs <info@quarterbit.dev>
Project-URL: Homepage, https://quarterbit.dev
Project-URL: Documentation, https://quarterbit.dev/docs
Keywords: optimizer,adam,deep-learning,pytorch,gpu,memory-efficient,compression,axiom
Classifier: Development Status :: 5 - Production/Stable
Classifier: Intended Audience :: Science/Research
Classifier: Operating System :: Microsoft :: Windows
Classifier: Operating System :: POSIX :: Linux
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.11
Description-Content-Type: text/markdown

# QuarterBit AXIOM

**Train 70B models on a single GPU. Train 13B models for FREE on Kaggle.**

AXIOM is a memory-efficient training system that achieves **15-17x memory compression** through VLA streaming weights and a built-in optimizer that matches or beats AdamW.

## What's New in v20.0.0

- **70B on Single GPU**: Train Llama-2 70B on one A100/H100 (was impossible before)
- **13B on FREE Kaggle T4**: Full training on free 16GB GPU
- **15.7x Memory Compression**: 840 GB → 53 GB for 70B model
- **90% Energy Savings**: Same model, 11x fewer GPUs

## The Numbers

| Model | Standard FP16+AdamW | AXIOM | Compression |
|-------|---------------------|-------|-------------|
| Llama-2 7B | 84 GB | 5.5 GB | 15x |
| Llama-2 13B | 156 GB | 9 GB | 17x |
| Llama-2 70B | 840 GB | 53 GB | 16x |

## Installation

```bash
pip install quarterbit
```

**Requirements:**
- Python 3.11+
- PyTorch 2.0+ with CUDA
- NVIDIA GPU (Pascal or newer)

## Quick Start

### Option 1: VLA Streaming (Largest Models)

For 70B+ models on single GPU:

```python
from transformers import AutoModelForCausalLM
from quarterbit.vla_trainable import make_vla_trainable

# Load model
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-70b-hf", torch_dtype=torch.float16)

# Convert to AXIOM training (17x compression)
model = make_vla_trainable(model)
model = model.cuda()

# Train with standard PyTorch
optimizer = torch.optim.AdamW(model.parameters(), lr=5e-5)
for batch in dataloader:
    loss = model(**batch).loss
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()
```

### Option 2: AXIOM Trainer (Simplest)

For smaller models with zero configuration:

```python
from quarterbit import AXIOM_Trainer

trainer = AXIOM_Trainer(model, train_loader, val_loader)
results = trainer.fit(steps=2000)

print(f"Val PPL: {results['final_val_ppl']:.1f}")
print(f"Peak VRAM: {results['peak_vram_gb']:.1f} GB")
```

## What Can You Train?

| Hardware | Standard Max | AXIOM Max | Improvement |
|----------|--------------|-----------|-------------|
| RTX 4070 (8GB) | 0.6B | 9B | 15x |
| RTX 4090 (24GB) | 1.8B | 26B | 14x |
| Kaggle T4 FREE (16GB) | 1.2B | 17B | 14x |
| A100 (80GB) | 6B | 88B | 15x |
| H100 (80GB) | 6B | 88B | 15x |

## Verified Results

**Llama-2 70B on Single GPU** (March 2026):

| Metric | Value |
|--------|-------|
| GPU Memory | 53 GB |
| Compression | 15.7x |
| Energy Saved | 90.9% |
| PPL Improvement | 100% |
| Full Training | Yes (all 70B params) |

## CLI Commands

```bash
# Create free account
quarterbit login

# Check license status
quarterbit status

# Deactivate machine
quarterbit deactivate
```

## Supported Models

- LLaMA 3, LLaMA 2, LLaMA
- Mistral, Mixtral
- Yi-34B, Yi-1.5
- Phi-3, Phi-2
- Gemma, Gemma 2
- Qwen, Qwen2
- Any HuggingFace CausalLM

## License

**Free account required** - Sign up at [quarterbit.dev](https://quarterbit.dev)

| Tier | Price | GPU Hours | Features |
|------|-------|-----------|----------|
| Free | $0 | 5/month | Personal use, community support |
| Academic | $0 | 10/month | University email required |
| Pro | $49/mo | Unlimited | Commercial license, priority support |
| Team | $299/mo | Unlimited | Up to 10 members, direct engineer access |
| Enterprise | Custom | Unlimited | Custom SLA, dedicated support |

## Links

- **Website**: [quarterbit.dev](https://quarterbit.dev)
- **Documentation**: [quarterbit.dev/docs](https://quarterbit.dev/docs)
- **Support**: info@quarterbit.dev

---

**Clouthier Simulation Labs** | Y Combinator Spring 2026 Applicant

Copyright 2026 Clouthier Simulation Labs
