torch2bt

The bridge between PyTorch research and the Bittensor decentralized intelligence network.

Turn any torch.nn.Module into a revenue-generating Bittensor miner — zero boilerplate.

How it works

1. Inspect — Analyzes your model's forward() signature via reflection
2. Synthesize — Generates protocol.py, miner.py, and Dockerfile using Python 3.14 t-strings
3. Deploy — Drop the output into any GPU host and start mining

Install

uv add torch2bt

Usage

import torch2bt as t2b
from my_models import SuperNeuralNet

t2b.package(
    model=SuperNeuralNet(),
    target_subnet=18,
    optimization="fp16",
    wallet_name="mining_key",
)

Output: torch2bt_output/protocol.py, miner.py, Dockerfile, pyproject.toml

Supported subnets

NetUID	Name	Optimizations
1	Text Prompting	FP32/FP16/BF16/INT8/INT4
18	Cortex	FP16/BF16

Local testing

from torch2bt.testing import MockValidator

validator = MockValidator(MySynapse, subnet_id=18, forward_fn=my_forward)
result = validator.query({"prompt": "a red cat"})

Examples

See examples/ for full runnable scripts:

Script	Subnet	Description
sn1_text_prompting.py	SN1	Transformer LM → Text Prompting miner
sn18_image_generation.py	SN18	Diffusion model → Cortex image miner

TODO

Phase A — Alpha (current)

• inspector.py — extract model forward() signature via reflection
• codegen.py — generate protocol.py, miner.py, Dockerfile, pyproject.toml using Python 3.14 t-strings
• subnets/ — protocol registry for SN1 (Text Prompting) and SN18 (Cortex)
• testing/ — MockValidator + MockSynapse for offline miner testing
• t2b.package() — end-to-end packaging API
• CI — ruff lint/format, zuban type check, pytest
• PyPI metadata — version 0.1.0a1, classifiers, license, URLs
• Publish 0.1.0a1 to PyPI

Phase B — Beta

• t2b.deploy(platform="runpod") — provision GPU instance via RunPod API
• t2b.deploy(platform="lambda") — Lambda Labs GPU support
• Auto-register hotkey with btcli post-deploy
• Dynamic TAO (dTAO / BIT001) profitability dashboard integration
• Auto-quantization — convert FP32 models to INT4/INT8 on the fly with bitsandbytes
• uv.lock generation for deterministic miner environments

Phase C — Production

• Multi-subnet mining — host multiple models on a single Axon
• Self-healing miners — auto-restart on OOM or network failure
• Expand subnet registry beyond SN1 + SN18
• t2b.benchmark() — measure model latency vs subnet timeout requirements