Skip to content

VRAM-AI/vram-network

Repository files navigation

Vram Network

Sui Testnet Move tests License Apache 2.0

Trustless decentralised machine learning on Sui. GPU miners train models; AWS Nitro Enclave validators score every gradient and sign the proof on-chain; rewards settle in VRAM every 10 minutes — no trusted coordinator, no whitelist.

Block explorer → · Live demo · Sui Explorer


What's in this repo

This is the open-source SDK for Vram Network — the contracts, on-chain client, adapter framework, operator CLI, and Python sidecar trainer. Everything here is open and auditable.

The miner and validator binaries are closed-source and distributed separately (see below). This repo contains the protocol they speak.

Component Path What it is
Smart contracts contracts/sources/ 13 Sui Move modules — PeerRegistry, ScoreLedger, RewardDistributor, TrainingJobBoard, and more
Core types + OpenSkill vramhub-core/ Shared types, Bayesian rating system, error types — no I/O
Sui RPC client vramhub-chain/ All on-chain calls: register, submit scores, claim rewards, post training jobs
Training adapters vramhub-adapter/ Pluggable adapter trait + impls: Candle nano-GPT, Python sidecar, custom
Operator CLI vramhub-cli/ vram-cli register-miner, register-validator, status, scores
Python sidecar python/ Run any HuggingFace model as a VRAM miner via HTTP
Docs docs/ Architecture, incentives, contract reference, env-var guide

Run a miner

curl -sSL https://www.vram.network/install.sh | bash

Set your wallet mnemonic in ~/.vramhub/.env and run vram-miner. Full guide: ONBOARDING.md — Option B.

Network Sui Testnet
Reward pool 6,000,000 VRAM ready to distribute
Emission 70 VRAM / 10-minute window
Min stake 1 SUI (refundable)

Run a validator

The validator installer lives at VRAM-AI/vram-validator.

curl -sSf https://raw.githubusercontent.com/VRAM-AI/vram-validator/main/install.sh | bash

Installs vram-validator + vram-cli. Full setup guide: ONBOARDING.md — Option C1/C2.

  • Test mode — any Linux VPS, simulated scoring, no AWS required
  • Nitro enclave — AWS c5.xlarge, hardware-attested scoring (~$36/mo spot)

Train a model

Post a training job on vram.network/training. The reference dataset (Sui Move / DeepBook instruction pairs for google/gemma-4-E2B-it) is at VRAM-AI/training.

# Start the Python sidecar on your GPU with the Move dataset
python python/vram_trainer.py \
  --model google/gemma-4-E2B-it \
  --dataset https://raw.githubusercontent.com/VRAM-AI/training/main/data/train_clean.jsonl \
  --device cuda --dtype bfloat16 --batch 2 --seqlen 2048

v0.7 testnet contracts

Object ID
Package 0xaff18bf6286047126901610d758d8fd111c9215a6e46abc704b6a0be838badd5
PeerRegistry 0x50a9982f6a3d6c1e6674f0fb4fa8b985007dbe19fc797abc691579be1f6493df
ValidatorRegistry 0x438d0ce63d40210d4e621cca6aaaf5d5438adaa54dfd71383fe41a82692a2561
EnclaveRegistry 0x442b82e471c1ee4577ea1f2168deb1f0b04fcc861ab79edb4b9c7d7738bf7f9f
ScoreLedger 0x0d2594727abeb45a13763baf8801ae765fbe41d147b28916ca78a0d08f73223a
RoundState 0xc1f18dc92629907641bc3176449af39738d2d8a93b4ad6b22548f4aed91d2611
Hparams 0x18b884530033f9b3e449b898c540ee5d3a25c4cab0abcf4843ef8e86e12adbfc
RewardPool 0x576ebeb78449ad46ef70dc3c5ca4e38d178846610bd7cf9f0764ae2f1dc0fe93
TrainingJobBoard 0xb481254350087569f904fe6fc45c337c0905651040791e532e0f044b9fc7474c

Chain-id 4c78adac · Deployer 0xb7aaeb31…c772c3

# Verify contracts locally — no wallet needed
git clone https://github.com/VRAM-AI/vram-network.git
cd vram-network/contracts && sui move test

How to contribute

The quickest entry points depending on what you want to do:

Audit / improve the contracts

All 13 Move modules are in contracts/sources/. Run sui move test — 101 tests, all must pass. Good first areas: training_jobs.move (job lifecycle), score_ledger.move (signature verification), reward_distributor.move (halving math).

Add a training adapter

Implement TrainingFrameworkAdapter in vramhub-adapter/src/. The trait has five methods: train_step, forward_loss, compress_gradient, load_checkpoint, save_checkpoint. See vramhub-adapter/src/sidecar.rs for the simplest reference (~150 lines — it just calls a local HTTP server). See CONTRIBUTING.md for the full walkthrough.

Improve the on-chain client

vramhub-chain/src/ has one file per object type. If a new contract method is added, the corresponding Rust function goes here. The client is pure async Rust with no hidden state — easy to extend.

Write a Python sidecar for a new model family

python/vram_trainer.py supports any AutoModelForCausalLM model via --model. To add a new dataset format or training objective, extend _get_batch(). See python/README.md if it exists, otherwise read the docstring at the top of the file.

Fix docs

docs/ is CC BY 4.0. If something is wrong or out of date, open a PR directly — no issue needed for doc fixes.

Run a node and report issues

The fastest way to find real bugs is to actually run a miner or validator and report what breaks. See ONBOARDING.md.


Architecture

   GPU Miner ──── compressed gradient ───→  Walrus (Sui)
                                                 │
                                                 ▼
                                        Validator (AWS Nitro Enclave)
                                                 │
                                         attested loss-delta
                                                 ▼
                                        Sui ScoreLedger
                                                 │
                                      Bayesian OpenSkill update
                                                 ▼
                                        RewardDistributor
                                                 │
                                         VRAM per window
                                                 ▼
                                           Miner wallet

Full threat model and trust assumptions: docs/architecture.md


Tokenomics

  • Hard cap: 21,000,000 VRAM
  • Mining pool (50%): emitted per-window, supply-based halving — Phase 1: 70 VRAM/window → Phase 2: 35 VRAM/window
  • TGE pre-mint (50%): Treasury 25% (6m cliff, 48m vest) · Investors 5% (SAFT lock-up) · Team 8% (12m cliff, 36m vest) · Liquidity 7% (instant) · Airdrop 5% (converts from testnet points)
  • Validator bonding: 2,100 → 21,000 VRAM burned per slot (4 tiers, max 500 validators)

Full model: docs/incentives.md


Roadmap

Feature Status
Miner binary (testnet) ✅ Live
Validator — test mode ✅ Live
Validator — Nitro enclave (mainnet) ⏳ In progress
VRAM token payouts ⏳ Pending first Nitro validator
vram-sdk public release 🗓 Q3 2026
GPU marketplace / rental 🗓 Planned

License

  • Contracts (contracts/) — MIT
  • SDK crates (vramhub-*/) — Apache 2.0
  • Validator installer (VRAM-AI/vram-validator) — MIT
  • Miner / Validator binaries — proprietary; freely installable under LICENSE-BINARIES.md

Security issues: email security@vram.ai — see SECURITY.md. Do not open public issues for vulnerabilities.


Team: team@vram.ai · Builders: builders@vram.ai · © 2024-2026 VRAM AI Limited

About

No description, website, or topics provided.

Resources

License

Contributing

Security policy

Stars

Watchers

Forks

Packages

 
 
 

Contributors