flow-g

A typed FunctionGraph IR for heterogeneous inference.

flowG is the compiler IR + runtime substrate behind the openIE-dev ecosystem. It takes typed function graphs and dispatches them across CPU, Apple-silicon Metal, WGPU, AMX, and the OpDispatch backend family — with picojoule-budgeted per-op energy measurement built in.

This is the public release surface. Source is private at openIE-dev/flow-g-core — 99 crates spanning the IR, codegen passes, OpDispatch leaves, model runtimes, and benchmarks.

Status

Release binaries + examples + documentation. flowG is free to use software, not an open-source project. See LICENSE for the Business Source License 1.1 terms — converts to Apache-2.0 four years after each binary's release date.

What you get

Plus a Rust library, flowg, for embedding.

Install

# via cargo
cargo install flowg

# via cargo-binstall (prebuilt binary)
cargo binstall flowg

# direct download
curl -fsSL https://github.com/openIE-dev/flow-g/releases/latest/download/flowg-$(uname -s)-$(uname -m).tar.gz | tar xz

Platform support:

Hello, flow-g

# compile a graph from .fg to executable
flowg compile examples/hello-graph/add.fg -o add

# lower to WGSL for WebGPU
flowg emit --target wgsl examples/hello-graph/add.fg

# run with energy receipts
flowg run examples/hello-graph/add.fg --record-energy receipt.json
cat receipt.json | jq '.ops[] | {name, picojoules}'

See examples/ for runnable programs spanning every backend.

What makes flow-g different

• Typed FunctionGraph IR — not an AST, not a SSA, not a tensor program. Functions are first-class graph nodes with typed signatures, designed for energy-cost-aware dispatch.
• OpDispatch backend — kernels are addressed by (op, dtype, layout, target) tuples. The runtime picks the lowest-joule kernel that matches.
• Heterogeneous by default — CPU + Metal + WGPU + AMX share the same IR. No "GPU port" — just dispatch the same graph at a different target.
• Energy substrate — every op has measured picojoule cost on Apple Silicon (IOReport-backed), NVML for Linux NVIDIA, RAPL for x86_64 Linux. Energy receipts are first-class output.
• 70+ model runtimes — Llama family, Qwen family, Mamba2, Bamba, Jamba, Hunyuan, Gemma 4, LFM2 family, Whisper, Moonshine, V-JEPA, and more — all using the same IR.

How it fits

flowG is the substrate. The openIE-dev language family compiles to it:

• Lux — general-purpose reactive language → flowG
• JMax — math-native language → flowG
• Joule — energy-aware self-hosted language → flowG

And the substrate is metered by JouleDB — the energy-metered HDC database that records inference receipts.

Documentation

• Getting started → https://openie-dev.github.io/flow-g
• Examples → examples/
• API reference → https://docs.rs/flowg
• Source mirror → openIE-dev/flow-g-core

Releases

GitHub Releases — tagged versions with prebuilt binaries for every supported platform, plus a SHA-256 checksums file per release.

Community

• Discussions — Q&A, design discussion
• Issues — bug reports
• Security: see SECURITY.md

License

• Binaries — Business Source License 1.1; see LICENSE. Free for non-commercial use, internal use by orgs under $1M revenue, security/academic research. Converts to Apache-2.0 four years after each release.
• Documentation — CC-BY-4.0
• Examples — Apache-2.0