matvec_bw/
matvec_bw.rs

1//! Weight-path bandwidth probe (scratch diagnostic, not a product).
2//!
3//! Two modes, both isolating the weight path from GDN/attention/sampling:
4//!
5//! - `one <tensor>`: repeat ONE tensor. A single dispatch amortized over a
6//!   large matrix — measures the kernel's ceiling. NOTE: tensors small
7//!   enough to sit in the SLC report inflated GB/s; only the ~636 MB head
8//!   is a trustworthy DRAM number here.
9//! - `sweep` (default): walk EVERY 2-D q8_2f tensor once, in directory
10//!   order — the real decode access pattern (cold weights, one dispatch
11//!   per tensor, ~200 dispatches). This is the honest weight-path number.
12//!
13//! Usage: cargo run --release --example matvec_bw -- <model.cmf> [sweep|one <tensor>]
14
15use cortiq_engine::pool::Pool;
16use cortiq_engine::qtensor::QTensor;
17use cortiq_core::TensorDtype;
18use std::sync::Arc;
19use std::time::Instant;
20
21fn main() {
22    let mut args = std::env::args().skip(1);
23    let path = args.next().expect("usage: matvec_bw <model.cmf> [sweep|one <tensor>]");
24    let mode = args.next().unwrap_or_else(|| "sweep".to_string());
25    let model = Arc::new(cortiq_core::CmfModel::open(&path).expect("open model"));
26
27    // Real LM activations carry a few heavy channels (>8·rms); measured
28    // mean on this model is ~3.7. NOUT models that distribution.
29    let nout: usize = std::env::var("NOUT").ok().and_then(|v| v.parse().ok()).unwrap_or(4);
30    let mk_x = |cols: usize| -> Vec<f32> {
31        let mut x: Vec<f32> = (0..cols).map(|i| ((i % 17) as f32 - 8.0) / 8.0).collect();
32        for k in 0..nout {
33            x[k * 37 % cols] = 40.0;
34        }
35        x
36    };
37
38    if mode == "one" {
39        let name = args.next().unwrap_or_else(|| "model.embed_tokens.weight".to_string());
40        let entry = model.tensor(&name).expect("tensor not found");
41        let (rows, cols) = (entry.shape[0], entry.shape[1]);
42        let nbytes = entry.nbytes as f64;
43        println!("tensor {name}: {rows}x{cols} {:?} = {:.1} MB, NOUT={nout}", entry.dtype, nbytes / 1e6);
44        let t = QTensor::from_model(&model, &name).expect("wrap");
45        let x = mk_x(cols);
46        let mut out = vec![0f32; rows];
47        t.matvec(&x, &mut out, None);
48        for nt in [1usize, 2, 4, 6, 8, 10] {
49            let pool = if nt == 1 { None } else { Some(Pool::new(nt)) };
50            let iters = 8;
51            t.matvec(&x, &mut out, pool.as_ref());
52            let t0 = Instant::now();
53            for _ in 0..iters {
54                t.matvec(&x, &mut out, pool.as_ref());
55            }
56            let el = t0.elapsed().as_secs_f64();
57            println!("threads={nt:2}  {:6.2} ms/matvec  {:6.1} GB/s (sink {:.3})",
58                el / iters as f64 * 1e3, nbytes * iters as f64 / el / 1e9, out[0]);
59        }
60        return;
61    }
62
63    // sweep: every 2-D q8_2f tensor once = one decode's worth of weights.
64    let names: Vec<String> = model
65        .tensors
66        .iter()
67        .filter(|t| t.dtype == TensorDtype::Q8_2f && t.shape.len() == 2)
68        .map(|t| t.name.clone())
69        .collect();
70    let total_bytes: f64 = model
71        .tensors
72        .iter()
73        .filter(|t| t.dtype == TensorDtype::Q8_2f && t.shape.len() == 2)
74        .map(|t| t.nbytes as f64)
75        .sum();
76    println!(
77        "sweep: {} q8_2f tensors, {:.2} GB total (= weights streamed per decode token), NOUT={nout}",
78        names.len(),
79        total_bytes / 1e9
80    );
81
82    let tensors: Vec<(QTensor, Vec<f32>, Vec<f32>)> = names
83        .iter()
84        .map(|n| {
85            let e = model.tensor(n).unwrap();
86            let (rows, cols) = (e.shape[0], e.shape[1]);
87            (QTensor::from_model(&model, n).expect("wrap"), mk_x(cols), vec![0f32; rows])
88        })
89        .collect();
90    let mut tensors = tensors;
91
92    // Whole-model residency pass BEFORE any timing: the first touch of a
93    // 4.2 GB mmap faults ~260k pages, which would otherwise be charged to
94    // whichever thread count happens to run first. REVERSE=1 flips the
95    // order as a check that no first-touch cost is left in the table.
96    for _ in 0..2 {
97        for (t, x, out) in tensors.iter_mut() {
98            t.matvec(x, out, None);
99        }
100    }
101    let mut counts = vec![1usize, 2, 4, 6, 8, 10];
102    if std::env::var("REVERSE").is_ok() {
103        counts.reverse();
104    }
105    for nt in counts {
106        let pool = if nt == 1 { None } else { Some(Pool::new(nt)) };
107        // one warm pass, then two measured
108        for (t, x, out) in tensors.iter_mut() {
109            t.matvec(x, out, pool.as_ref());
110        }
111        let iters = 2;
112        let t0 = Instant::now();
113        for _ in 0..iters {
114            for (t, x, out) in tensors.iter_mut() {
115                t.matvec(x, out, pool.as_ref());
116            }
117        }
118        let el = t0.elapsed().as_secs_f64();
119        let per_tok = el / iters as f64;
120        println!(
121            "threads={nt:2}  {:7.1} ms/sweep  {:6.1} GB/s  -> weight-path-only ceiling {:5.1} tok/s",
122            per_tok * 1e3,
123            total_bytes * iters as f64 / el / 1e9,
124            1.0 / per_tok
125        );
126    }
127}
matvec_bw/matvec_bw.rs

matvec_bw/
matvec_bw.rs
