rlx_cpu/
dequant_cache.rs

1// RLX — versatile ML compiler + runtime.
2// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
3//
4// This program is free software: you can redistribute it and/or modify
5// it under the terms of the GNU General Public License as published by
6// the Free Software Foundation, version 3.
7//
8// This program is distributed in the hope that it will be useful,
9// but WITHOUT ANY WARRANTY; without even the implied warranty of
10// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11// GNU General Public License for more details.
12//
13// You should have received a copy of the GNU General Public License
14// along with this program. If not, see <https://www.gnu.org/licenses/>.
15//! Cache dequantized GGUF weight bytes for static params.
16//!
17//! Qwen3.5 decode with `--packed` was re-dequantizing every K-quant
18//! weight on every matmul (hundreds of times per token). Keys are
19//! `(k, n, scheme, bytes_hash)` — stable for identical GGUF bytes regardless
20//! of arena offset (multiple compiled graphs reuse offsets).
21
22use rlx_ir::quant::QuantScheme;
23use std::collections::HashMap;
24use std::sync::{Arc, OnceLock, RwLock};
25
26#[derive(Clone, Copy, Debug, Eq, PartialEq, Hash)]
27struct DequantKey {
28    k: u32,
29    n: u32,
30    scheme: u8,
31    /// Content hash — `w_off` alone collides across compiled graphs.
32    bytes_hash: u64,
33}
34
35fn weight_bytes_hash(w_bytes: &[u8]) -> u64 {
36    use std::hash::{Hash, Hasher};
37    let mut hasher = std::collections::hash_map::DefaultHasher::new();
38    w_bytes.hash(&mut hasher);
39    hasher.finish()
40}
41
42fn scheme_tag(scheme: QuantScheme) -> u8 {
43    match scheme {
44        QuantScheme::GgufQ4K => 1,
45        QuantScheme::GgufQ5K => 2,
46        QuantScheme::GgufQ6K => 3,
47        QuantScheme::GgufQ8K => 4,
48        _ => 255,
49    }
50}
51
52fn dequant_gguf(w_bytes: &[u8], k: usize, n: usize, scheme: QuantScheme) -> Vec<f32> {
53    match scheme {
54        QuantScheme::GgufQ4K => rlx_gguf::dequant_q4_k(w_bytes, k * n),
55        QuantScheme::GgufQ5K => rlx_gguf::dequant_q5_k(w_bytes, k * n),
56        QuantScheme::GgufQ6K => rlx_gguf::dequant_q6_k(w_bytes, k * n),
57        QuantScheme::GgufQ8K => rlx_gguf::dequant_q8_k(w_bytes, k * n),
58        other => panic!("dequant_cache: unsupported GGUF scheme {other:?}"),
59    }
60    .expect("GGUF dequant failed")
61}
62
63static CACHE: OnceLock<RwLock<HashMap<DequantKey, Arc<[f32]>>>> = OnceLock::new();
64
65fn cache_enabled() -> bool {
66    !matches!(
67        rlx_ir::env::var("RLX_DEQUANT_CACHE").as_deref(),
68        Some("0") | Some("false") | Some("off")
69    )
70}
71
72/// Return dense `[k×n]` weights (GGUF row-major `[n,k]` layout) for `w_bytes`.
73pub fn gguf_weight_f32(
74    _w_off: usize,
75    w_bytes: &[u8],
76    k: usize,
77    n: usize,
78    scheme: QuantScheme,
79) -> Arc<[f32]> {
80    if !cache_enabled() {
81        return Arc::from(dequant_gguf(w_bytes, k, n, scheme).into_boxed_slice());
82    }
83    let key = DequantKey {
84        k: k as u32,
85        n: n as u32,
86        scheme: scheme_tag(scheme),
87        bytes_hash: weight_bytes_hash(w_bytes),
88    };
89    let cache = CACHE.get_or_init(|| RwLock::new(HashMap::new()));
90    if let Some(hit) = cache.read().expect("dequant cache poisoned").get(&key) {
91        return Arc::clone(hit);
92    }
93    let dense = dequant_gguf(w_bytes, k, n, scheme);
94    let arc: Arc<[f32]> = Arc::from(dense.into_boxed_slice());
95    cache
96        .write()
97        .expect("dequant cache poisoned")
98        .insert(key, Arc::clone(&arc));
99    arc
100}
101
102/// Drop cached dequantized weights (e.g. between model loads in tests).
103pub fn clear_dequant_cache() {
104    if let Some(c) = CACHE.get() {
105        c.write().expect("dequant cache poisoned").clear();
106    }
107}
108
109#[cfg(test)]
110mod tests {
111    use super::*;
112
113    #[test]
114    fn gguf_dequant_cache_hits_on_second_lookup() {
115        clear_dequant_cache();
116        const QK_K: usize = 256;
117        let mut packed = Vec::new();
118        packed.extend_from_slice(&half::f16::from_f32(1.0).to_le_bytes());
119        packed.extend_from_slice(&half::f16::from_f32(1.0).to_le_bytes());
120        let mut scales = [0u8; 12];
121        for s in &mut scales[0..4] {
122            *s = 0x01;
123        }
124        packed.extend_from_slice(&scales);
125        packed.extend(std::iter::repeat_n(0x77u8, QK_K / 2));
126        let k = 256;
127        let n = 1;
128        let w_off = 4096;
129        let hash = weight_bytes_hash(&packed);
130        let a = gguf_weight_f32(w_off, &packed, k, n, QuantScheme::GgufQ4K);
131        let b = gguf_weight_f32(w_off + 999, &packed, k, n, QuantScheme::GgufQ4K);
132        assert!(
133            Arc::ptr_eq(&a, &b),
134            "same bytes at different offsets should hit"
135        );
136        let mut other = packed.clone();
137        other[0] ^= 0x01;
138        let c = gguf_weight_f32(w_off, &other, k, n, QuantScheme::GgufQ4K);
139        assert!(!Arc::ptr_eq(&a, &c), "different bytes should miss: {hash}");
140    }
141}
rlx_cpu/dequant_cache.rs

rlx_cpu/
dequant_cache.rs
