rlx_optim/
qhadamw.rs

1// RLX — versatile ML compiler + runtime.
2// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
3// SPDX-License-Identifier: MIT OR Apache-2.0
4
5//! QHAdamW — Quasi-Hyperbolic Adam (Ma & Yarats, 2019) with decoupled
6//! weight decay.
7//!
8//! # Idea
9//!
10//! Adam can be viewed as "the second moment EMA scales the gradient,
11//! the first moment EMA *replaces* the gradient." The quasi-hyperbolic
12//! family says: don't *replace* — *interpolate*. Mix the EMA with the
13//! raw current gradient, controlled by per-moment scalars `ν₁, ν₂`.
14//!
15//! # Update rule
16//!
17//! ```text
18//! m_t = β₁·m_{t-1} + (1 − β₁)·g_t
19//! v_t = β₂·v_{t-1} + (1 − β₂)·g_t²
20//! num = (1 − ν₁)·g_t   + ν₁·m̂_t
21//! den = √((1 − ν₂)·g_t² + ν₂·v̂_t) + ε
22//! θ_t = θ_{t-1} − lr · ( num / den + λ·θ_{t-1} )
23//! ```
24//!
25//! Setting `ν₁ = ν₂ = 1` recovers standard AdamW; `ν₁ = β₁` recovers
26//! Nesterov-style behavior in the limit; `ν₁ < 1` injects more of the
27//! current gradient and tends to be more robust on noisy losses.
28//!
29//! # When to use
30//!
31//! When you've found AdamW too sluggish on noisy / heavy-tail
32//! gradients (RL, GAN training) and an LR sweep didn't help.
33
34use std::collections::HashMap;
35
36use crate::Optimizer;
37use crate::common::zeros_entry;
38
39/// Quasi-hyperbolic AdamW. Per-tensor state: two `f32` buffers.
40#[derive(Debug, Clone)]
41pub struct QHAdamW {
42    /// Learning rate.
43    pub lr: f32,
44    /// First-moment EMA decay β₁. Ma & Yarats recommend `0.995` (much
45    /// closer to 1 than vanilla Adam) — the QH interpolation already
46    /// keeps current-gradient weight in the numerator.
47    pub beta1: f32,
48    /// Second-moment EMA decay β₂. Default `0.999`.
49    pub beta2: f32,
50    /// First-moment QH interpolation coefficient ν₁ ∈ \[0, 1\].
51    /// `1.0` = pure EMA (standard Adam first moment); `0.0` = pure
52    /// current gradient (no momentum). Default `0.7`.
53    pub nu1: f32,
54    /// Second-moment QH interpolation coefficient ν₂. `1.0` = standard
55    /// Adam denominator. Default `1.0`.
56    pub nu2: f32,
57    /// Denominator stability constant. Default `1e-8`.
58    pub eps: f32,
59    /// Decoupled weight-decay coefficient λ. Default `0.01`.
60    pub weight_decay: f32,
61    step: u64,
62    m: HashMap<String, Vec<f32>>,
63    v: HashMap<String, Vec<f32>>,
64}
65
66impl QHAdamW {
67    /// Construct with `(β₁, β₂, ν₁, ν₂, ε, λ) = (0.995, 0.999, 0.7, 1.0, 1e-8, 0.01)`.
68    pub fn new(lr: f32) -> Self {
69        Self {
70            lr,
71            beta1: 0.995,
72            beta2: 0.999,
73            nu1: 0.7,
74            nu2: 1.0,
75            eps: 1e-8,
76            weight_decay: 0.01,
77            step: 0,
78            m: HashMap::new(),
79            v: HashMap::new(),
80        }
81    }
82
83    /// Override (β₁, β₂).
84    pub fn with_betas(mut self, b1: f32, b2: f32) -> Self {
85        self.beta1 = b1;
86        self.beta2 = b2;
87        self
88    }
89
90    /// Override the quasi-hyperbolic coefficients (ν₁, ν₂).
91    pub fn with_nus(mut self, n1: f32, n2: f32) -> Self {
92        self.nu1 = n1;
93        self.nu2 = n2;
94        self
95    }
96
97    /// Override the decoupled-decay coefficient.
98    pub fn with_weight_decay(mut self, wd: f32) -> Self {
99        self.weight_decay = wd;
100        self
101    }
102}
103
104impl Optimizer for QHAdamW {
105    fn set_lr(&mut self, lr: f32) {
106        self.lr = lr;
107    }
108
109    fn step(&mut self, name: &str, _shape: &[usize], param: &mut [f32], grad: &[f32]) {
110        debug_assert_eq!(param.len(), grad.len());
111        let t = (self.step + 1) as f64;
112        let b1 = self.beta1 as f64;
113        let b2 = self.beta2 as f64;
114        let bc1 = 1.0 - b1.powf(t);
115        let bc2 = 1.0 - b2.powf(t);
116        let n1 = self.nu1 as f64;
117        let n2 = self.nu2 as f64;
118        let eps = self.eps as f64;
119        let lr = self.lr as f64;
120        let wd = self.weight_decay as f64;
121        let m = zeros_entry(&mut self.m, name, param.len());
122        let v = zeros_entry(&mut self.v, name, param.len());
123        for i in 0..param.len() {
124            let g = grad[i] as f64;
125            let mi = b1 * m[i] as f64 + (1.0 - b1) * g;
126            let vi = b2 * v[i] as f64 + (1.0 - b2) * g * g;
127            m[i] = mi as f32;
128            v[i] = vi as f32;
129            let m_hat = mi / bc1;
130            let v_hat = vi / bc2;
131            // Quasi-hyperbolic numerator & denominator (Ma & Yarats Alg. 2).
132            let num = (1.0 - n1) * g + n1 * m_hat;
133            let den = ((1.0 - n2) * g * g + n2 * v_hat).sqrt() + eps;
134            let p = param[i] as f64;
135            param[i] = (p - lr * (num / den + wd * p)) as f32;
136        }
137    }
138
139    fn end_iteration(&mut self) {
140        self.step += 1;
141    }
142}
rlx_optim/qhadamw.rs

rlx_optim/
qhadamw.rs
