Introduction

VRAM HUB is a decentralized training coordination layer that enables large-scale LLM pre-training across heterogeneous compute resources over the internet. Built on the Sui blockchain, it replaces trusted coordinators with verifiable on-chain incentives, Seal IBE credential encryption, and Nitro TEE-attested loss evaluation.

What Problem Does VRAM HUB Solve?

Large-scale LLM pre-training traditionally requires:

1. Centralized orchestration — a trusted coordinator routes data, collects gradients, and distributes rewards
2. Homogeneous compute — training clusters assume identical hardware and network conditions
3. Trust in evaluators — a centralized party determines who contributed useful work

VRAM HUB eliminates all three requirements:

• Every coordination event (window, score, reward) is anchored on Sui — verifiable by anyone
• Heterogeneous compute is first-class: miners can run any hardware in any geography
• Nitro TEE attestation makes loss evaluation unforgeable — the enclave's binary is committed to on-chain PCR values; a compromised scorer fails registration

Core Participants

How Windows Work

The system advances in synchronized windows (default: 10 minutes each):

Window N begins
  ├── Miners load checkpoint from window N-1
  ├── Miners train on their assigned data batch
  ├── Miners upload compressed gradient to storage backend (Walrus / R2 / local)
  │     └── ref: walrus:{blob_id} or gradient-{window}-{uid}-v{version}.pt
  ├── Validators discover gradient refs via on-chain gradient_registry
  ├── Validators download gradients → send to Nautilus enclave
  ├── Enclave computes loss delta → signs result with Ed25519
  ├── Validators submit signed scores to score_ledger.move
  └── reward_distributor.move emits tokens weighted by OpenSkill rating
Window N+1 begins

Key Properties

• Permissionless — any node with stake can register as miner or validator; no whitelist
• Sybil-resistant — rewards require stake; fake miners without stake earn nothing
• Verifiable — enclave PCR values committed on-chain; any observer can verify scorer integrity
• Pluggable — bring your own training framework via the TrainingFrameworkAdapter trait

Technology Stack