rlx-ffi

C ABI for the RLX distributed node — link a mesh worker into an iOS app, an embedded host, or anything else that speaks C.

A node is not necessarily a workstation. The RLX transport layer has no platform gating and no default C dependencies, so anything with a CPU, a TCP stack and std can be a rank. This crate is the seam that lets a non-Rust shell be one.

#include "rlx_node.h"

/* Join a 2-rank mesh as worker 1, serving inference. */
rlx_node_start(1, 2, "192.168.1.10:29500", "auto", "infer");

char buf[256];
rlx_node_status(buf, sizeof buf);   /* "running", then "ok: …" */
rlx_node_stop();

Every entry point returns immediately — a serving loop parks in recv between activations, and a UI thread must not block on it. Poll rlx_node_status.

Modes

• "infer" — serve a stage the coordinator ships. Stoppable between activations.
• "train" — join a data-parallel training run. Not stoppable partway: the gradient reduce is a barrier, so a rank that leaves stalls every other rank.

Building

staticlib only. A shared library must resolve every symbol at link time and zstd-sys (pulled in by the GGUF reader) does not on iOS; a static archive links happily on both iOS and embedded Linux.

cargo build -p rlx-ffi --release --target aarch64-apple-ios

A static archive carries no link directives, so the consuming target must name what the Rust code needs itself. On Apple platforms that is -framework Accelerate (rlx-cpu's BLAS / LAPACK / vForce), plus -lresolv and -lc++ for Rust std. See ios/Demo/project.yml in the repository for a worked set.

Features: apple (Metal + ANE), gpu (portable wgpu). Both off by default so the archive cross-compiles to targets with no such frameworks.

iOS

ios/build-xcframework.sh wraps this crate into RlxNode.xcframework, with a Swift wrapper and a demo app. iOS needs NSLocalNetworkUsageDescription in Info.plist for any LAN peer traffic — without it the node reaches nothing and says nothing. See ios/README.md.

Trying it

The desktop side has a self-test that needs no device:

cargo run -p rlx-ffi --example node_coordinator -- --world 2 --self-test
cargo run -p rlx-ffi --example node_coordinator -- --world 2 --self-test --train

License

MIT OR Apache-2.0.