rlx-optim 0.2.16

RLX training-step optimizers — Adam, AdamW, NAdamW, RAdam, QHAdamW, LAMB, Adafactor, Lion, SOAP, Kron-PSGD, Muon, Sophia, MARS
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12

// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
// SPDX-License-Identifier: MIT OR Apache-2.0
// Link Apple's Accelerate framework on macOS so Muon's Newton–Schulz
// orthogonalization can call `cblas_sgemm` (AMX-backed BLAS) instead of the
// portable hand-rolled matmul — the dominant cost of Muon on large models.
// No-op on other platforms (the naive parallel path is used there).
fn main() {
    if std::env::var("CARGO_CFG_TARGET_OS").as_deref() == Ok("macos") {
        println!("cargo:rustc-link-lib=framework=Accelerate");
    }
}