Architecture · 11-crate Rust workspaceLive on mainnet

Built from the bottom up

OmniNode is a single Rust workspace on edition 2024. Networking and storage sit at the base; a zero-copy bridge and pipeline coordinator sit above; the proof, chain, and operator surfaces sit at the top. Default builds pull zero chain and zero proof dependencies — capabilities are opt-in cargo features.

The workspace

Eleven crates, one dependency spine

Each crate has one job and depends only on the layers beneath it.

Protocol & operator surface

omni-zkml

Proof-artifact substrate, verifier attestation envelope, and the chain-client abstraction.

omni-sumchain

SUM Chain JSON-RPC adapter — attestation submit + settlement read / claim / dispute.

omni-contributor

Off-chain contributor inference node — jobs, sessions, integrity-evidence chains.

omni-node

The operator CLI binary — listen, shard, fetch, verify-proof, settlement.

omni-proofs-halo2-reference

4→8→4 int16 Halo2 reference proof (halo2-mlp-v1).

omni-proofs-halo2-production-mlp

16→32→16→8 int16 Halo2 production proof (production-fixedpoint-mlp-v1).

Bridge & coordination

omni-bridge

PyO3 native extension — zero-copy shard access exposed to NumPy / MLX.

omni-pipeline

Pipeline-parallel coordination — GPipe scheduling, hidden-state tensor transport.

Networking & storage

omni-net

P2P mesh on libp2p 0.55 — QUIC, mDNS discovery, Gossipsub, shard & tensor request-response.

omni-store

Zero-copy GGUF parsing, layer-wise chunking, BLAKE3 → CIDv1 addressing, SNIP V2 store.

Shared foundation

omni-types

Shared types, errors, config, and the Phase-5 attestation / commitment types.

Networking

A mesh on libp2p 0.55

Peers move two kinds of bytes: content-addressed model weights, and the hidden-state tensors that flow between pipeline stages.

  • QUIC and TCP transports, Noise encryption, Yamux multiplexing.
  • Peers discover each other over mDNS; Gossipsub carries pub/sub with strict validation.
  • Two request-response codecs: /omni/shard-xfer for weights, /omni/tensor-xfer for tensors.

LAN is the proven path

mDNS discovery and the two-machine LAN demo are what run end-to-end today. WAN traversal (AutoNAT, relay, DCUtR) is present in the code but deferred — treat internet-scale meshing as roadmap, not a shipped feature.

omni-net.spec

transports
QUIC · TCP
security
Noise
multiplexing
Yamux
discovery
mDNS (LAN)
pub/sub
Gossipsub
shard_window
64 MiB
Apple Silicon

A weight is never copied

On unified-memory hardware, a shard travels from disk to the GPU without a single copy. The mmap the operating system makes is the same memory Metal computes on.

  1. Rust mmap — GGUF shardFile mapped into unified memory; CPU and GPU share the same physical RAM.
  2. PyO3 __getbuffer__The raw pointer is exposed to Python over the buffer protocol — no copy.
  3. NumPy frombuffer()Wraps the pointer as an ndarray in place.
  4. MLX mx.arrayMetal reads tensors directly from the same physical memory.
  5. GPU inferenceZero memory copies from disk to compute.

Grounded in