cortiq_engine/
lib.rs

1//! Cortiq inference engine — sparse forward pass, attention, tokenization, sampling.
2
3pub mod attention;
4pub mod audiovae;
5pub mod dit;
6pub mod dsv4;
7pub mod fcd;
8pub mod fcd_ops;
9pub mod g3n;
10pub mod gptq_capture;
11pub mod gpu;
12#[cfg(target_os = "macos")]
13pub mod gpu_metal;
14#[cfg(feature = "gpu")]
15pub mod gpu_wgpu;
16/// The native Vulkan lane — an accelerator behind a capability probe,
17/// present only where Vulkan is.
18#[cfg(all(
19    feature = "gpu",
20    any(target_os = "linux", target_os = "windows", target_os = "android")
21))]
22pub mod vulkan;
23pub mod imagegen;
24pub mod inference;
25pub mod kv_cache;
26pub mod linear_core;
27pub mod loader;
28pub mod mmh3;
29pub mod mm_ab;
30pub mod music3;
31pub mod nystrom;
32pub mod pin;
33pub mod pipeline;
34pub mod pool;
35pub mod qtensor;
36pub mod qwen3te;
37pub mod qwen3vis;
38pub mod router;
39pub mod runtime;
40pub mod sampler;
41pub mod skillbake;
42pub mod swarm;
43pub mod textenc;
44pub mod tokenizer;
45pub mod vae;
46pub mod vae3d;
47pub mod videogen;
48
49pub use nystrom::NystromState;
50pub use pipeline::{GenerateResult, Pipeline, TokenCallback, TokenTrace};
51pub use runtime::CortiqRuntime;
52
53/// Test-only: N empty Metal command-buffer round trips, total seconds.
54#[doc(hidden)]
55#[cfg(target_os = "macos")]
56pub fn gpu_empty_submit_for_test(n: usize) -> f64 {
57    gpu_metal::empty_submit_bench(n)
58}
59
60/// Test-only: N pipelined empty submits, one final wait.
61#[doc(hidden)]
62#[cfg(target_os = "macos")]
63pub fn gpu_pipelined_submit_for_test(n: usize) -> f64 {
64    gpu_metal::pipelined_submit_bench(n)
65}
66
67/// Test-only: build a q1 MoeJob trio (weight 1.0).
68#[doc(hidden)]
69#[cfg(target_os = "macos")]
70pub fn gpu_moe_job_for_test(
71    gi: usize,
72    ui: usize,
73    di: usize,
74    inter: usize,
75    hidden: usize,
76    x: Vec<f32>,
77) -> gpu::MoeJob<'static> {
78    gpu::MoeJob {
79        gate: (gi, inter, hidden, &[]),
80        up: (ui, inter, hidden, &[]),
81        down: (di, hidden, inter, &[]),
82        xs_gate: x.clone(),
83        xs_up: x,
84        down_col: &[],
85        w: 1.0,
86        q1: true,
87        q4t: false,
88        q4tp: false,
89        gu_q2: false,
90            swiglu_limit: 0.0,
91    }
92}
93
94/// Test-only: run the metal moe_block on one job.
95#[doc(hidden)]
96#[cfg(target_os = "macos")]
97pub fn gpu_moe_block_for_test(
98    model: &std::sync::Arc<cortiq_core::CmfModel>,
99    job: gpu::MoeJob<'_>,
100    out: &mut [f32],
101) -> bool {
102    gpu_metal::moe_block(model, &[job], out)
103}
104
105/// Test-only: q1 matvec_batch — jobs (idx, rows, cols); first two share
106/// x, the third takes xi.
107#[doc(hidden)]
108#[cfg(target_os = "macos")]
109pub fn gpu_batch_q1_for_test(
110    model: &std::sync::Arc<cortiq_core::CmfModel>,
111    shapes: &[(usize, usize, usize)],
112    x: &[f32],
113    xi: &[f32],
114    outs: &mut [&mut [f32]],
115) -> bool {
116    let jobs: Vec<gpu::BatchJob> = shapes
117        .iter()
118        .enumerate()
119        .map(|(k, &(idx, rows, cols))| gpu::BatchJob {
120            idx,
121            rows,
122            cols,
123            row_scale: &[],
124            xs: if k < 2 { x.to_vec() } else { xi.to_vec() },
125            layout: gpu::BatchLayout::Q1,
126        })
127        .collect();
128    gpu_metal::matvec_batch(model, &jobs, outs)
129}
130
131/// Test-only direct handle to the Metal q1 matvec (micro-benchmarks).
132#[doc(hidden)]
133#[cfg(target_os = "macos")]
134pub fn gpu_q1_matvec_for_test(
135    model: &std::sync::Arc<cortiq_core::CmfModel>,
136    idx: usize,
137    xs: &[f32],
138    rows: usize,
139    cols: usize,
140    out: &mut [f32],
141) -> bool {
142    gpu_metal::q1_matvec(model, idx, xs, rows, cols, out)
143}
144pub use sampler::SamplerConfig;
cortiq_engine/lib.rs

cortiq_engine/
lib.rs
