Enum QTensor 

pub enum QTensor {
    F32 {
        data: Vec<f32>,
        rows: usize,
        cols: usize,
    },
    Mapped {
        model: Arc<CmfModel>,
        idx: usize,
        dtype: TensorDtype,
        rows: usize,
        cols: usize,
        row_scale: Vec<f32>,
        col_field: Vec<f32>,
        vbit_offsets: Vec<usize>,
        repack: Vec<u8>,
    },
}

Variants

F32

Fields

data: Vec<f32>

rows: usize

cols: usize

Mapped

Fields

model: Arc<CmfModel>

idx: usize

Index into the model’s tensor directory.

dtype: TensorDtype

rows: usize

cols: usize

row_scale: Vec<f32>

Per-row scales, dequantized to f32 up front (tiny).

col_field: Vec<f32>

q8_2f column field (θ), dequantized up front; empty for q8_row.

vbit_offsets: Vec<usize>

Vbit only: byte offset of each row’s packed data within the tensor blob ([rows + 1], computed once at load — the per- matvec prefix scan over row bit-widths was O(rows) each call).

repack: Vec<u8>

q8-family decode repack (load-time, optional): rows in groups of 4, interleaved in 16-byte units — one 64-byte line per iteration feeds all 4 sdot lanes, ONE sequential weight stream per worker instead of four (this is where llama.cpp’s repacked Q8 kernels get their bandwidth). Empty = off (CMF_REPACK=0, non-SDOT arch, or an ineligible shape). Trades an anonymous copy of the quants for mmap pages that go cold.

Implementations

impl QTensor

pub fn from_f32(data: Vec<f32>, rows: usize, cols: usize) -> Self

pub fn from_model(model: &Arc<CmfModel>, name: &str) -> Result<Self, String>

Wrap a directory tensor without dequantizing the payload. Falls back to dequantized f32 for dtypes without a fused kernel.

Examples found in repository

examples/matvec_bw.rs (line 55)

fn main() {
    let mut args = std::env::args().skip(1);
    let path = args
        .next()
        .expect("usage: matvec_bw <model.cmf> [sweep|one <tensor>]");
    let mode = args.next().unwrap_or_else(|| "sweep".to_string());
    let model = Arc::new(cortiq_core::CmfModel::open(&path).expect("open model"));

    // Real LM activations carry a few heavy channels (>8·rms); measured
    // mean on this model is ~3.7. NOUT models that distribution.
    let nout: usize = std::env::var("NOUT")
        .ok()
        .and_then(|v| v.parse().ok())
        .unwrap_or(4);
    let mk_x = |cols: usize| -> Vec<f32> {
        let mut x: Vec<f32> = (0..cols).map(|i| ((i % 17) as f32 - 8.0) / 8.0).collect();
        for k in 0..nout {
            x[k * 37 % cols] = 40.0;
        }
        x
    };

    if mode == "one" {
        let name = args
            .next()
            .unwrap_or_else(|| "model.embed_tokens.weight".to_string());
        let entry = model.tensor(&name).expect("tensor not found");
        let (rows, cols) = (entry.shape[0], entry.shape[1]);
        let nbytes = entry.nbytes as f64;
        println!(
            "tensor {name}: {rows}x{cols} {:?} = {:.1} MB, NOUT={nout}",
            entry.dtype,
            nbytes / 1e6
        );
        let t = QTensor::from_model(&model, &name).expect("wrap");
        let x = mk_x(cols);
        let mut out = vec![0f32; rows];
        t.matvec(&x, &mut out, None);
        for nt in [1usize, 2, 4, 6, 8, 10] {
            let pool = if nt == 1 { None } else { Some(Pool::new(nt)) };
            let iters = 8;
            t.matvec(&x, &mut out, pool.as_ref());
            let t0 = Instant::now();
            for _ in 0..iters {
                t.matvec(&x, &mut out, pool.as_ref());
            }
            let el = t0.elapsed().as_secs_f64();
            println!(
                "threads={nt:2}  {:6.2} ms/matvec  {:6.1} GB/s (sink {:.3})",
                el / iters as f64 * 1e3,
                nbytes * iters as f64 / el / 1e9,
                out[0]
            );
        }
        return;
    }

    // sweep: every 2-D q8_2f tensor once = one decode's worth of weights.
    let names: Vec<String> = model
        .tensors
        .iter()
        .filter(|t| t.dtype == TensorDtype::Q8_2f && t.shape.len() == 2)
        .map(|t| t.name.clone())
        .collect();
    let total_bytes: f64 = model
        .tensors
        .iter()
        .filter(|t| t.dtype == TensorDtype::Q8_2f && t.shape.len() == 2)
        .map(|t| t.nbytes as f64)
        .sum();
    println!(
        "sweep: {} q8_2f tensors, {:.2} GB total (= weights streamed per decode token), NOUT={nout}",
        names.len(),
        total_bytes / 1e9
    );

    let tensors: Vec<(QTensor, Vec<f32>, Vec<f32>)> = names
        .iter()
        .map(|n| {
            let e = model.tensor(n).unwrap();
            let (rows, cols) = (e.shape[0], e.shape[1]);
            (
                QTensor::from_model(&model, n).expect("wrap"),
                mk_x(cols),
                vec![0f32; rows],
            )
        })
        .collect();
    let mut tensors = tensors;

    // Whole-model residency pass BEFORE any timing: the first touch of a
    // 4.2 GB mmap faults ~260k pages, which would otherwise be charged to
    // whichever thread count happens to run first. REVERSE=1 flips the
    // order as a check that no first-touch cost is left in the table.
    for _ in 0..2 {
        for (t, x, out) in tensors.iter_mut() {
            t.matvec(x, out, None);
        }
    }
    let mut counts = vec![1usize, 2, 4, 6, 8, 10];
    if std::env::var("REVERSE").is_ok() {
        counts.reverse();
    }
    for nt in counts {
        let pool = if nt == 1 { None } else { Some(Pool::new(nt)) };
        // one warm pass, then two measured
        for (t, x, out) in tensors.iter_mut() {
            t.matvec(x, out, pool.as_ref());
        }
        let iters = 2;
        let t0 = Instant::now();
        for _ in 0..iters {
            for (t, x, out) in tensors.iter_mut() {
                t.matvec(x, out, pool.as_ref());
            }
        }
        let el = t0.elapsed().as_secs_f64();
        let per_tok = el / iters as f64;
        println!(
            "threads={nt:2}  {:7.1} ms/sweep  {:6.1} GB/s  -> weight-path-only ceiling {:5.1} tok/s",
            per_tok * 1e3,
            total_bytes * iters as f64 / el / 1e9,
            1.0 / per_tok
        );
    }
}

pub fn rows(&self) -> usize

pub fn cols(&self) -> usize

pub fn mapped_q1(&self) -> Option<(&Arc<CmfModel>, usize)>

(model, tensor idx) for a q1 mapped weight — the wgpu token graph keys its resident VRAM cache by idx. None for any other dtype/kind.

pub fn graph_weight(&self) -> Option<(&Arc<CmfModel>, usize, u8, &[f32])>

(model, idx, kind, row_scale) for a graph-capable mapped weight. kind: 0=q8_row (per-row scales), 1=q1, 2=q4_tiled, 3=q1t (tile-embedded, no rs). None for dtypes the token graph does not handle (q8_2f/q4_block/vbit).

pub fn as_f32(&self) -> Option<&[f32]>

Dense f32 view — only for owned tensors. Masked/sparse execution paths require it; quantized weights don’t support masks yet.

pub fn row_f32(&self, r: usize, dst: &mut [f32])

Dequantize one row into dst (embedding lookup).

pub fn sparse_col_ok(&self) -> bool

Can this tensor’s columns be read cheaply (for sparse down_proj)? True for F32/Q8Row/Q8_2f (per-row scale, direct strided access); false for group-packed q4/vbit (column access would unpack whole groups — sparse execution falls back to f32 for those).

pub fn add_col_scaled(&self, c: usize, w: f32, out: &mut [f32])

down_proj [hidden, inter]: accumulate w · col(c) into out [hidden] — reads ONLY column c (one neuron) from the mmap, no full-matrix dequant. out[k] += w · down[k, c].

pub fn row_dot(&self, r: usize, x: &[f32], scratch: &mut [f32]) -> f32

Dot of row r with x (gate/up active-neuron path). Reads only row r from the mmap — no full dequant. q4/vbit dequant the row into scratch first (rare for active-FFN weights).

pub fn matvec(&self, x: &[f32], out: &mut [f32], pool: Option<&Pool>)

out = W · x (row-major). F32 delegates to the historical bit-exact path; Mapped runs the fused int8 kernel.

Examples found in repository

examples/matvec_bw.rs (line 58)

fn main() {
    let mut args = std::env::args().skip(1);
    let path = args
        .next()
        .expect("usage: matvec_bw <model.cmf> [sweep|one <tensor>]");
    let mode = args.next().unwrap_or_else(|| "sweep".to_string());
    let model = Arc::new(cortiq_core::CmfModel::open(&path).expect("open model"));

    // Real LM activations carry a few heavy channels (>8·rms); measured
    // mean on this model is ~3.7. NOUT models that distribution.
    let nout: usize = std::env::var("NOUT")
        .ok()
        .and_then(|v| v.parse().ok())
        .unwrap_or(4);
    let mk_x = |cols: usize| -> Vec<f32> {
        let mut x: Vec<f32> = (0..cols).map(|i| ((i % 17) as f32 - 8.0) / 8.0).collect();
        for k in 0..nout {
            x[k * 37 % cols] = 40.0;
        }
        x
    };

    if mode == "one" {
        let name = args
            .next()
            .unwrap_or_else(|| "model.embed_tokens.weight".to_string());
        let entry = model.tensor(&name).expect("tensor not found");
        let (rows, cols) = (entry.shape[0], entry.shape[1]);
        let nbytes = entry.nbytes as f64;
        println!(
            "tensor {name}: {rows}x{cols} {:?} = {:.1} MB, NOUT={nout}",
            entry.dtype,
            nbytes / 1e6
        );
        let t = QTensor::from_model(&model, &name).expect("wrap");
        let x = mk_x(cols);
        let mut out = vec![0f32; rows];
        t.matvec(&x, &mut out, None);
        for nt in [1usize, 2, 4, 6, 8, 10] {
            let pool = if nt == 1 { None } else { Some(Pool::new(nt)) };
            let iters = 8;
            t.matvec(&x, &mut out, pool.as_ref());
            let t0 = Instant::now();
            for _ in 0..iters {
                t.matvec(&x, &mut out, pool.as_ref());
            }
            let el = t0.elapsed().as_secs_f64();
            println!(
                "threads={nt:2}  {:6.2} ms/matvec  {:6.1} GB/s (sink {:.3})",
                el / iters as f64 * 1e3,
                nbytes * iters as f64 / el / 1e9,
                out[0]
            );
        }
        return;
    }

    // sweep: every 2-D q8_2f tensor once = one decode's worth of weights.
    let names: Vec<String> = model
        .tensors
        .iter()
        .filter(|t| t.dtype == TensorDtype::Q8_2f && t.shape.len() == 2)
        .map(|t| t.name.clone())
        .collect();
    let total_bytes: f64 = model
        .tensors
        .iter()
        .filter(|t| t.dtype == TensorDtype::Q8_2f && t.shape.len() == 2)
        .map(|t| t.nbytes as f64)
        .sum();
    println!(
        "sweep: {} q8_2f tensors, {:.2} GB total (= weights streamed per decode token), NOUT={nout}",
        names.len(),
        total_bytes / 1e9
    );

    let tensors: Vec<(QTensor, Vec<f32>, Vec<f32>)> = names
        .iter()
        .map(|n| {
            let e = model.tensor(n).unwrap();
            let (rows, cols) = (e.shape[0], e.shape[1]);
            (
                QTensor::from_model(&model, n).expect("wrap"),
                mk_x(cols),
                vec![0f32; rows],
            )
        })
        .collect();
    let mut tensors = tensors;

    // Whole-model residency pass BEFORE any timing: the first touch of a
    // 4.2 GB mmap faults ~260k pages, which would otherwise be charged to
    // whichever thread count happens to run first. REVERSE=1 flips the
    // order as a check that no first-touch cost is left in the table.
    for _ in 0..2 {
        for (t, x, out) in tensors.iter_mut() {
            t.matvec(x, out, None);
        }
    }
    let mut counts = vec![1usize, 2, 4, 6, 8, 10];
    if std::env::var("REVERSE").is_ok() {
        counts.reverse();
    }
    for nt in counts {
        let pool = if nt == 1 { None } else { Some(Pool::new(nt)) };
        // one warm pass, then two measured
        for (t, x, out) in tensors.iter_mut() {
            t.matvec(x, out, pool.as_ref());
        }
        let iters = 2;
        let t0 = Instant::now();
        for _ in 0..iters {
            for (t, x, out) in tensors.iter_mut() {
                t.matvec(x, out, pool.as_ref());
            }
        }
        let el = t0.elapsed().as_secs_f64();
        let per_tok = el / iters as f64;
        println!(
            "threads={nt:2}  {:7.1} ms/sweep  {:6.1} GB/s  -> weight-path-only ceiling {:5.1} tok/s",
            per_tok * 1e3,
            total_bytes * iters as f64 / el / 1e9,
            1.0 / per_tok
        );
    }
}

pub fn matvec2( &self, x1: &[f32], x2: &[f32], o1: &mut [f32], o2: &mut [f32], pool: Option<&Pool>, )

Fused two-input matvec (MTP verify pair): weights streamed once.

impl QTensor

pub fn matmat( &self, xs_all: &[f32], b: usize, out: &mut [f32], pool: Option<&Pool>, )

Batched matvec (prefill-GEMM): xs — row-major [b, cols], out — row-major [b, rows]. Element-wise semantics are IDENTICAL to b matvec calls (same dot kernels in the same order); the win — the weight row streams from DRAM once per batch, not b times.

impl QTensor

pub fn matvec_many<const N: usize>( ts: [&QTensor; N], x: &[f32], outs: [&mut [f32]; N], pool: Option<&Pool>, )

Multi-matrix job (roadmap §3 P0): N tensors sharing one input run under a SINGLE pool dispatch — QKV or gate+up cost one barrier instead of N. Per-row math is the exact same kernel as matvec (bit-identical outputs); only the dispatch is fused. Falls back to N sequential matvecs when the set is not a uniform q8-family/F32 group or there is no pool.

impl QTensor

pub fn matvec2_many<const N: usize>( ts: [&QTensor; N], x1: &[f32], x2: &[f32], o1s: [&mut [f32]; N], o2s: [&mut [f32]; N], pool: Option<&Pool>, )

Pair-input multi-matrix job: N tensors × 2 shared inputs under a single pool dispatch — the MTP/pair decode path publishes one job for Q/K/V (and one for gate+up) instead of one per tensor. Per-row math is exactly matvec2’s kernels; bit-identical.

pub fn matvec_silu_mul( gate: &QTensor, up: &QTensor, x: &[f32], out: &mut [f32], pool: Option<&Pool>, ) -> bool

Fused gate+up matvec with SiLU·mul: for each row r, computes silu(gate·x) * (up·x) and writes to out[r]. ONE pool dispatch, no intermediate g/u buffers, no separate silu pass. Falls back (returns false) for unsupported dtype combos.