1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
//! Accelerator backends: a trait the model's linear layers dispatch through, plus a global
//! registry chosen at start-up from the platform detection. Every backend is optional at run time;
//! the CPU kernels remain the reference path and the fallback.
use crate::ggml::GgufMatrix;
use crate::quant::{Q4Matrix, QMatrix};
use std::any::Any;
use std::sync::OnceLock;
use std::sync::atomic::{AtomicU8, Ordering};
/// A weight matrix in one of the supported storage formats (`[n_out, n_in]`).
pub enum WeightRef<'a> {
Bf16(&'a [u16]),
Q8(&'a QMatrix),
Q4(&'a Q4Matrix),
/// llama.cpp block formats: no device kernels yet, every backend declines them (CPU path).
Gguf(&'a GgufMatrix),
}
impl WeightRef<'_> {
pub fn bytes(&self) -> usize {
match self {
WeightRef::Bf16(w) => w.len() * 2,
WeightRef::Q8(m) => m.bytes(),
WeightRef::Q4(m) => m.bytes(),
WeightRef::Gguf(m) => m.bytes(),
}
}
}
/// Device-side state for one weight matrix (buffers / handles), owned by the layer.
pub type DeviceWeight = Box<dyn Any + Send + Sync>;
/// One matmul of a batch sharing the same input `x`: `y[t][n_out] = x[t][n_in] * W^T`.
pub struct Job<'a> {
pub w: WeightRef<'a>,
pub dev: &'a DeviceWeight,
pub n_out: usize,
pub y: &'a mut [f32],
}
/// Geometry of a k x k convolution with padding k/2 over channels-last input `[h*w, cin]`,
/// producing `[oh*ow, n_out]`; the weights are in the unfolded `[n_out][ky][kx][cin]` layout.
#[derive(Clone, Copy, Debug)]
pub struct ConvGeom {
pub h: usize,
pub w: usize,
pub cin: usize,
pub k: usize,
pub stride: usize,
pub oh: usize,
pub ow: usize,
}
/// A normalization applied to a convolution's input as it is loaded (so it runs on the device
/// without an extra pass): `GroupNorm(groups)` over the channels of all pixels, or `Layer` over each
/// pixel's (row's) channels; then `* gamma + beta`, then SiLU if `silu`.
#[derive(Clone, Copy, Debug, PartialEq)]
pub enum NormKind {
Group(usize),
Layer,
}
#[derive(Clone, Copy, Debug)]
pub struct NormSpec<'a> {
pub kind: NormKind,
pub eps: f32,
pub gamma: &'a [f32],
pub beta: &'a [f32],
pub silu: bool,
}
pub trait Backend: Send + Sync {
fn name(&self) -> &str;
/// Whether this backend should run a matmul of these dimensions. Dispatch has a fixed cost
/// (~0.15 ms on Metal), so small ops and single-token decode stay on the CPU.
fn wants(&self, n_out: usize, n_in: usize, t: usize) -> bool;
/// Prepares device-side state for a weight matrix (zero-copy where the API allows).
fn upload(&self, w: &WeightRef, n_out: usize, n_in: usize) -> Option<DeviceWeight>;
/// Runs a batch of matmuls over one shared input in a single submission. Returns false if
/// nothing was computed (the caller then runs the whole batch on the CPU).
fn matmul(&self, x: &[f32], t: usize, n_in: usize, jobs: &mut [Job]) -> bool;
/// Runs a convolution directly on the device (gathering the k x k neighbourhoods itself, so
/// the unfolded input never exists in host memory), optionally normalizing its input on the
/// way in. `job.y` is `[oh*ow, n_out]`. Returns false when the backend does not handle this
/// case (the caller normalizes and unfolds on the CPU).
fn conv2d(&self, _x: &[f32], _g: &ConvGeom, _norm: Option<&NormSpec>, _job: &mut Job) -> bool {
false
}
/// Multi-head attention: `q` `[tq, heads*dh]`, `k` and `v` `[tk, heads*dh]` into `out`
/// (`[tq, heads*dh]`), softmax(q k^T / sqrt(dh)) v, keys after each query masked if `causal`.
#[allow(clippy::too_many_arguments)]
fn attention(
&self,
_q: &[f32],
_k: &[f32],
_v: &[f32],
_tq: usize,
_tk: usize,
_heads: usize,
_dh: usize,
_causal: bool,
_out: &mut [f32],
) -> bool {
false
}
/// Device state for the routed experts of an MoE layer: expert `e` has the fused gate/up
/// projection `gate_up[e]` (`[2 * inter, h]`, gate rows then up rows) and the down projection
/// `down[e]` (`[h, inter]`). None if the backend does not run MoE layers.
fn upload_moe(&self, _gate_up: &[WeightRef], _down: &[WeightRef], _h: usize, _inter: usize) -> Option<DeviceWeight> {
None
}
/// The routed experts of an MoE layer (`dev` from `upload_moe`) in one submission: token `i`
/// of `x` (`[t, h]`) goes through the experts `routes[i]` lists (expert index, weight), and
/// `y[i] += sum of weight * down(silu(gate) * up)`. Returns false if nothing was computed
/// (the caller then runs the experts itself).
#[allow(clippy::too_many_arguments)]
fn moe(
&self,
_x: &[f32],
_t: usize,
_h: usize,
_inter: usize,
_routes: &[Vec<(usize, f32)>],
_dev: &DeviceWeight,
_y: &mut [f32],
) -> bool {
false
}
/// Cumulative timing summary for `--profile`.
fn stats(&self) -> Option<String> {
None
}
}
static ACTIVE: OnceLock<Option<Box<dyn Backend>>> = OnceLock::new();
/// Process-wide switches: 0 = unset (environment default), 1 = on, 2 = off.
static EXACT: AtomicU8 = AtomicU8::new(0);
static MOE: AtomicU8 = AtomicU8::new(0);
fn switch(a: &AtomicU8, env: &'static OnceLock<Option<String>>, var: &str, on_env: fn(&str) -> bool, default: bool) -> bool {
match a.load(Ordering::Relaxed) {
1 => true,
2 => false,
_ => env.get_or_init(|| std::env::var(var).ok()).as_deref().map_or(default, on_env),
}
}
/// Keeps GPU matmuls bit-identical to the CPU: no reduced-precision matrix units (Metal on
/// M5/A19 converts inputs to half with f32 accumulation, a KL divergence of 1e-7 to 1e-3 from
/// the CPU). Process-wide, takes effect on the next dispatch. Default: off, or
/// `OXINFER_METAL_MMA=0`.
pub fn set_exact(on: bool) {
EXACT.store(if on { 1 } else { 2 }, Ordering::Relaxed);
}
/// See [`set_exact`].
pub fn exact() -> bool {
static ENV: OnceLock<Option<String>> = OnceLock::new();
switch(&EXACT, &ENV, "OXINFER_METAL_MMA", |v| v == "0", false)
}
/// Runs MoE expert matmuls on the GPU when the backend can (prefill). Process-wide; affects
/// models loaded afterwards. Default: on, or `OXINFER_METAL_MOE=0` for off.
pub fn set_moe_offload(on: bool) {
MOE.store(if on { 1 } else { 2 }, Ordering::Relaxed);
}
/// See [`set_moe_offload`].
pub fn moe_offload() -> bool {
static ENV: OnceLock<Option<String>> = OnceLock::new();
switch(&MOE, &ENV, "OXINFER_METAL_MOE", |v| v != "0", true)
}
/// Installs the backend for this process (call once, before loading models).
pub fn install(b: Option<Box<dyn Backend>>) {
let _ = ACTIVE.set(b);
}
/// Whether a backend choice (possibly "none") was installed already.
pub fn initialized() -> bool {
ACTIVE.get().is_some()
}
pub fn active() -> Option<&'static dyn Backend> {
ACTIVE.get().and_then(|b| b.as_deref())
}
/// Picks and initialises a backend for the detected platform and the user's preference.
pub fn init(plat: &crate::platform::Platform, pref: crate::platform::DevicePref) -> Option<&'static dyn Backend> {
use crate::platform::DevicePref;
let chosen: Option<Box<dyn Backend>> = match plat.pick(pref) {
None => None,
Some(g) => {
let r: Result<Box<dyn Backend>, String> = match g.kind {
#[cfg(all(feature = "gpu", any(target_os = "macos", target_os = "ios"), target_arch = "aarch64"))]
crate::platform::GpuKind::Metal => crate::metal::MetalBackend::new()
.map(|b| Box::new(b) as Box<dyn Backend>)
.map_err(|e| e.to_string()),
#[cfg(all(feature = "gpu", not(any(target_os = "macos", target_os = "ios", target_arch = "wasm32"))))]
crate::platform::GpuKind::Cuda => crate::cuda::CudaBackend::new(g.index)
.map(|b| Box::new(b) as Box<dyn Backend>)
.map_err(|e| e.to_string()),
_ => Err(format!("{} compute is not implemented in this build", g.kind.as_str())),
};
match r {
Ok(b) => Some(b),
Err(e) => {
if pref != DevicePref::Auto || !e.contains("not implemented") {
eprintln!("[{}: {e}; using CPU]", g.kind.as_str());
}
None
}
}
}
};
install(chosen);
active()
}