inferencelayer 0.2.4

//! GGUF container reading + the BINARY-weight ("Q1") pack — the Bonsai/BitNet class.
//!
//! Two things live here, deliberately together and deliberately OUTSIDE `weights.rs` (which
//! carries another session's in-flight work): the minimal GGUF v2/v3 parser (header, metadata,
//! tensor table, on-demand tensor reads — enough to ingest `prism-ml/Bonsai-27B-gguf` and its
//! relatives), and the Q1 weight codec the engine's kernels consume.
//!
//! ## The Q1_0 disk format (reverse-engineered from the `bonsai-webgpu-kernels` repack shader)
//!
//! Per 128 weights along a row: an **18-byte block** = 2-byte f16 scale + 16 bytes (128 bits)
//! of signs, bit `i` = sign of weight `i` (1 → +1, 0 → −1); `w = scale · sign`. That is
//! 1.125 bits/weight — a 27B model in 3.8 GB, exactly what the space's loader reports.
//!
//! ## The engine's GPU layout ([`quantize_q1`] / [`gemv_q1_k_lcpp_src`](crate::gemv_q1_k_lcpp_src))
//!
//! Mirrors the Q4_0 family: `scales` f16 `[m, n/128]`, `bits` u32 `[m, n/32]` row-major —
//! per 32-weight block one sign word (4.5 bytes amortized vs Q4_0's 18). The GEMV decodes
//! with a `select`, no weight-side multiplies; at the decode bandwidth wall the byte ratio is
//! the speed ratio (`tests/q1_kernel.rs` measures it).

use anyhow::{Context, Result, bail, ensure};
use std::io::{Read, Seek, SeekFrom};

/// Pack an `[m, n]` row-major f32 weight into the engine's Q1 layout: per-128-block scale =
/// `mean(|w|)` (the BitNet convention — for a checkpoint that is ALREADY binary-valued this
/// recovers its scale exactly), bit = 1 where `w ≥ 0`. `n` must be a multiple of 128 (every
/// production projection is; the GEMV's block stride assumes it).
pub fn quantize_q1(w: &[f32], m: usize, n: usize) -> (Vec<f32>, Vec<u32>) {
    assert_eq!(w.len(), m * n, "weight is not [m, n]");
    assert_eq!(n % 128, 0, "Q1 rows pack in 128-weight blocks");
    let nsc = n / 128;
    let nblk = n / 32;
    let mut scales = vec![0f32; m * nsc];
    let mut bits = vec![0u32; m * nblk];
    for r in 0..m {
        let row = &w[r * n..(r + 1) * n];
        for sb in 0..nsc {
            let blk = &row[sb * 128..(sb + 1) * 128];
            let mean_abs = blk.iter().map(|v| v.abs()).sum::<f32>() / 128.0;
            scales[r * nsc + sb] = if mean_abs > 0.0 { mean_abs } else { 1.0 };
            for (j, &v) in blk.iter().enumerate() {
                if v >= 0.0 {
                    let word = sb * 4 + j / 32;
                    bits[r * nblk + word] |= 1u32 << (j % 32);
                }
            }
        }
    }
    (scales, bits)
}

/// Dequantize the engine layout back to f32 — the CPU reference the parity gate compares
/// against (and the fallback a non-Q1 device would use).
pub fn dequantize_q1(scales: &[f32], bits: &[u32], m: usize, n: usize) -> Vec<f32> {
    let nsc = n / 128;
    let nblk = n / 32;
    let mut w = vec![0f32; m * n];
    for r in 0..m {
        for j in 0..n {
            let s = scales[r * nsc + j / 128];
            let b = (bits[r * nblk + j / 32] >> (j % 32)) & 1;
            w[r * n + j] = if b == 1 { s } else { -s };
        }
    }
    w
}

/// Decode one Q1_0 DISK block run (`nblocks` × 18 bytes) into the engine layout for one row
/// segment: returns (scales f32 per 128, sign words). The GGUF tensor data for a `[rows, k]`
/// Q1_0 tensor is exactly `rows · k/128` such blocks, row-major.
pub fn q1_0_disk_to_engine(raw: &[u8], nblocks: usize) -> Result<(Vec<f32>, Vec<u32>)> {
    ensure!(
        raw.len() >= nblocks * 18,
        "Q1_0 data short: {} < {}",
        raw.len(),
        nblocks * 18
    );
    let mut scales = Vec::with_capacity(nblocks);
    let mut bits = Vec::with_capacity(nblocks * 4);
    for b in 0..nblocks {
        let base = b * 18;
        let sc = u16::from_le_bytes([raw[base], raw[base + 1]]);
        scales.push(f16_bits_to_f32(sc));
        for w in 0..4 {
            let o = base + 2 + w * 4;
            bits.push(u32::from_le_bytes([
                raw[o],
                raw[o + 1],
                raw[o + 2],
                raw[o + 3],
            ]));
        }
    }
    Ok((scales, bits))
}

/// Decode a run of ggml-quantized bytes to row-major f32, for the block types an Unsloth Dynamic
/// GGUF actually mixes: `Q8_0`(8), `Q4_0`(2), `Q4_K`(12), `Q6_K`(14). `count` is the element
/// count (a whole number of blocks). Every decoder mirrors llama.cpp's own dequantization and is
/// gated bit-for-bit against it (`tests/gguf_kquant.rs`, fixtures from the reference reader).
///
/// This is the ingestion half of loading an Unsloth artifact: decode-to-f32, then repack per the
/// engine's precision policy. The double-quantization cost (e.g. Q4_K → engine-Q4_0) is what the
/// KLD ruler measures before anyone trusts it.
pub fn dequant_ggml(raw: &[u8], ggml_type: u32, count: usize) -> Result<Vec<f32>> {
    match ggml_type {
        8 => dequant_q8_0(raw, count),
        2 => dequant_q4_0(raw, count),
        6 => dequant_q5_0(raw, count),
        10 => dequant_q2_k(raw, count),
        11 => dequant_q3_k(raw, count),
        12 => dequant_q4_k(raw, count),
        13 => dequant_q5_k(raw, count),
        14 => dequant_q6_k(raw, count),
        16 => dequant_iq2_xxs(raw, count),
        17 => dequant_iq2_xs(raw, count),
        18 => dequant_iq3_xxs(raw, count),
        19 => dequant_iq1_s(raw, count),
        20 => dequant_iq4_nl(raw, count),
        21 => dequant_iq3_s(raw, count),
        22 => dequant_iq2_s(raw, count),
        23 => dequant_iq4_xs(raw, count),
        29 => dequant_iq1_m(raw, count),
        30 => {
            // BF16: the upper 16 bits of an f32 — widen by shifting left.
            ensure!(raw.len() >= count * 2, "BF16 data short");
            Ok(raw[..count * 2]
                .chunks_exact(2)
                .map(|c| f32::from_bits(u32::from(u16::from_le_bytes([c[0], c[1]])) << 16))
                .collect())
        }
        other => bail!(
            "dequant_ggml: ggml type {other} ({}) not decodable",
            GgufFile::type_name(other)
        ),
    }
}

/// Q5_0: 32 weights/block, 22 bytes = f16 `d` + `u32 qh` (the 5th bit of each weight) + 16 nibble
/// bytes. `q = low_nibble | (qh_bit << 4)`; `w = d · (q − 16)`. Nibble order matches Q4_0 (byte
/// `l` holds weight `l` low, `l+16` high).
fn dequant_q5_0(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(32), "Q5_0 count {count} not /32");
    let (nb, bs) = (count / 32, 22);
    ensure!(raw.len() >= nb * bs, "Q5_0 data short");
    let mut out = vec![0f32; count];
    for b in 0..nb {
        let o = b * bs;
        let d = f16_bits_to_f32(u16::from_le_bytes([raw[o], raw[o + 1]]));
        let qh = u32::from_le_bytes([raw[o + 2], raw[o + 3], raw[o + 4], raw[o + 5]]);
        let qs = &raw[o + 6..o + 22];
        for l in 0..16 {
            let byte = qs[l];
            let hlo = ((qh >> l) & 1) as u8;
            let hhi = ((qh >> (l + 16)) & 1) as u8;
            let qlo = (byte & 0x0F) | (hlo << 4);
            let qhi = (byte >> 4) | (hhi << 4);
            out[b * 32 + l] = d * (qlo as i32 - 16) as f32;
            out[b * 32 + l + 16] = d * (qhi as i32 - 16) as f32;
        }
    }
    Ok(out)
}

/// Q8_0: 32 weights/block, 34 bytes = f16 scale `d` + 32 `i8`. `w = d · q`.
fn dequant_q8_0(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(32), "Q8_0 count {count} not /32");
    let (nb, bs) = (count / 32, 34);
    ensure!(raw.len() >= nb * bs, "Q8_0 data short");
    let mut out = Vec::with_capacity(count);
    for b in 0..nb {
        let o = b * bs;
        let d = f16_bits_to_f32(u16::from_le_bytes([raw[o], raw[o + 1]]));
        for i in 0..32 {
            out.push(d * (raw[o + 2 + i] as i8 as f32));
        }
    }
    Ok(out)
}

/// Q4_0: 32 weights/block, 18 bytes = f16 scale `d` + 16 nibble-pairs. Byte `l` holds weight `l`
/// (low nibble) and `l+16` (high nibble); `w = d · (nibble − 8)`.
fn dequant_q4_0(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(32), "Q4_0 count {count} not /32");
    let (nb, bs) = (count / 32, 18);
    ensure!(raw.len() >= nb * bs, "Q4_0 data short");
    let mut out = vec![0f32; count];
    for b in 0..nb {
        let o = b * bs;
        let d = f16_bits_to_f32(u16::from_le_bytes([raw[o], raw[o + 1]]));
        for l in 0..16 {
            let byte = raw[o + 2 + l];
            out[b * 32 + l] = d * (((byte & 0x0F) as i32 - 8) as f32);
            out[b * 32 + l + 16] = d * (((byte >> 4) as i32 - 8) as f32);
        }
    }
    Ok(out)
}

/// The Q4_K/Q5_K 6-bit scale/min unpack (mirrors `gguf.quants.Q4_K.get_scale_min`): 12 packed
/// bytes → eight 6-bit `sc` and eight 6-bit `min`.
fn q4k_scale_min(s: &[u8]) -> ([u8; 8], [u8; 8]) {
    let (mut sc, mut mn) = ([0u8; 8], [0u8; 8]);
    for i in 0..4 {
        sc[i] = s[i] & 0x3F;
        mn[i] = s[i + 4] & 0x3F;
        sc[i + 4] = (s[i + 8] & 0x0F) | ((s[i] >> 2) & 0x30);
        mn[i + 4] = (s[i + 8] >> 4) | ((s[i + 4] >> 2) & 0x30);
    }
    (sc, mn)
}

/// Q4_K: 256-weight superblock, 144 bytes = f16 `d` + f16 `dmin` + 12 scale bytes + 128 nibble
/// bytes. Eight 32-weight sub-blocks; sub-block `s` uses byte-group `s/2`, low nibble if `s` even
/// else high. `w = d·sc[s]·nibble − dmin·min[s]`.
fn dequant_q4_k(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256), "Q4_K count {count} not /256");
    let (nb, bs) = (count / 256, 144);
    ensure!(raw.len() >= nb * bs, "Q4_K data short");
    let mut out = vec![0f32; count];
    for b in 0..nb {
        let o = b * bs;
        let d = f16_bits_to_f32(u16::from_le_bytes([raw[o], raw[o + 1]]));
        let dmin = f16_bits_to_f32(u16::from_le_bytes([raw[o + 2], raw[o + 3]]));
        let (sc, mn) = q4k_scale_min(&raw[o + 4..o + 16]);
        let qs = &raw[o + 16..o + 144];
        for s in 0..8 {
            let (g, p) = (s / 2, s % 2);
            let ds = d * sc[s] as f32;
            let ms = dmin * mn[s] as f32;
            for l in 0..32 {
                let nib = (qs[g * 32 + l] >> (4 * p)) & 0x0F;
                out[b * 256 + s * 32 + l] = ds * nib as f32 - ms;
            }
        }
    }
    Ok(out)
}

/// Q5_K: 256-weight superblock, 176 bytes = f16 `d` + f16 `dmin` + 12 scale bytes (the SAME
/// 6-bit pack as Q4_K) + 32 `qh` bytes (the 5th bit) + 128 nibble bytes. Sub-block `s` takes its
/// 5th bit from `qh[l] >> s`; `w = d·sc[s]·q − dmin·min[s]` with `q = nibble | (bit << 4)`.
fn dequant_q5_k(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256), "Q5_K count {count} not /256");
    let (nb, bs) = (count / 256, 176);
    ensure!(raw.len() >= nb * bs, "Q5_K data short");
    let mut out = vec![0f32; count];
    for b in 0..nb {
        let o = b * bs;
        let d = f16_bits_to_f32(u16::from_le_bytes([raw[o], raw[o + 1]]));
        let dmin = f16_bits_to_f32(u16::from_le_bytes([raw[o + 2], raw[o + 3]]));
        let (sc, mn) = q4k_scale_min(&raw[o + 4..o + 16]);
        let qh = &raw[o + 16..o + 48];
        let qs = &raw[o + 48..o + 176];
        for s in 0..8 {
            let (g, p) = (s / 2, s % 2);
            let ds = d * sc[s] as f32;
            let ms = dmin * mn[s] as f32;
            for l in 0..32 {
                let nib = (qs[g * 32 + l] >> (4 * p)) & 0x0F;
                let hi = (qh[l] >> s) & 1;
                out[b * 256 + s * 32 + l] = ds * (nib | (hi << 4)) as f32 - ms;
            }
        }
    }
    Ok(out)
}

/// Q2_K: 256-weight superblock, 84 bytes = 16 scale bytes (low nibble = scale, high = min, one
/// per 16-weight sub-block) + 64 two-bit bytes + f16 `d` + f16 `dmin`. Element `e`: byte group
/// `e/128`, shift `2·((e%128)/32)`, byte `e%32`; `w = d·(sc&0xF)·q − dmin·(sc>>4)`.
fn dequant_q2_k(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256), "Q2_K count {count} not /256");
    let (nb, bs) = (count / 256, 84);
    ensure!(raw.len() >= nb * bs, "Q2_K data short");
    let mut out = vec![0f32; count];
    for b in 0..nb {
        let o = b * bs;
        let scales = &raw[o..o + 16];
        let qs = &raw[o + 16..o + 80];
        let d = f16_bits_to_f32(u16::from_le_bytes([raw[o + 80], raw[o + 81]]));
        let dmin = f16_bits_to_f32(u16::from_le_bytes([raw[o + 82], raw[o + 83]]));
        for e in 0..256usize {
            let (g, w) = (e / 128, e % 128);
            let (sh, l) = (w / 32, w % 32);
            let q = (qs[g * 32 + l] >> (2 * sh)) & 3;
            let sc = scales[e / 16];
            out[b * 256 + e] = d * (sc & 0x0F) as f32 * q as f32 - dmin * (sc >> 4) as f32;
        }
    }
    Ok(out)
}

/// Q3_K: 256-weight superblock, 110 bytes = 32 `hmask` bytes + 64 two-bit bytes + 12 packed
/// 6-bit scale bytes + f16 `d`. Sixteen signed 6-bit scales (−32-centered); element `e`'s value
/// is `2-bit − 4·(1 − hmask_bit)` — the high bit REMOVES the −4 offset (llama.cpp convention).
fn dequant_q3_k(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256), "Q3_K count {count} not /256");
    let (nb, bs) = (count / 256, 110);
    ensure!(raw.len() >= nb * bs, "Q3_K data short");
    let mut out = vec![0f32; count];
    for b in 0..nb {
        let o = b * bs;
        let hmask = &raw[o..o + 32];
        let qs = &raw[o + 32..o + 96];
        let scb = &raw[o + 96..o + 108];
        let d = f16_bits_to_f32(u16::from_le_bytes([raw[o + 108], raw[o + 109]]));
        // 16 six-bit scales: low nibbles from bytes 0-7 (j%8, shifted by 4·(j/8)), high 2 bits
        // from bytes 8-11 (j%4, shifted by 2·(j/4)).
        let mut dl = [0f32; 16];
        for (j, slot) in dl.iter_mut().enumerate() {
            let lo = (scb[j % 8] >> (4 * (j / 8))) & 0x0F;
            let hi = (scb[8 + (j % 4)] >> (2 * (j / 4))) & 0x03;
            *slot = d * (((lo | (hi << 4)) as i8 as i32) - 32) as f32;
        }
        for e in 0..256usize {
            let (g, w) = (e / 128, e % 128);
            let (sh, l) = (w / 32, w % 32);
            let q = ((qs[g * 32 + l] >> (2 * sh)) & 3) as i32;
            let hbit = (hmask[e % 32] >> (e / 32)) & 1;
            let q = q - if hbit == 0 { 4 } else { 0 };
            out[b * 256 + e] = dl[e / 16] * q as f32;
        }
    }
    Ok(out)
}

/// Q6_K: 256-weight superblock, 210 bytes = 128 `ql` + 64 `qh` + 16 `i8` scales + f16 `d`. Two
/// 128-weight halves; each 6-bit weight = 4 low bits (`ql`) + 2 high bits (`qh`), centered by −32.
/// `w = d·scale·(q − 32)` (llama.cpp `dequantize_row_q6_K`).
fn dequant_q6_k(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256), "Q6_K count {count} not /256");
    let (nb, bs) = (count / 256, 210);
    ensure!(raw.len() >= nb * bs, "Q6_K data short");
    let mut out = vec![0f32; count];
    for b in 0..nb {
        let o = b * bs;
        let ql = &raw[o..o + 128];
        let qh = &raw[o + 128..o + 192];
        let sc = &raw[o + 192..o + 208];
        let d = f16_bits_to_f32(u16::from_le_bytes([raw[o + 208], raw[o + 209]]));
        for half in 0..2 {
            let (qlh, qhh, sch) = (&ql[half * 64..], &qh[half * 32..], &sc[half * 8..]);
            let base = b * 256 + half * 128;
            for l in 0..32 {
                let is = l / 16;
                let q1 = ((qlh[l] & 0x0F) | (((qhh[l] >> 0) & 3) << 4)) as i32 - 32;
                let q2 = ((qlh[l + 32] & 0x0F) | (((qhh[l] >> 2) & 3) << 4)) as i32 - 32;
                let q3 = ((qlh[l] >> 4) | (((qhh[l] >> 4) & 3) << 4)) as i32 - 32;
                let q4 = ((qlh[l + 32] >> 4) | (((qhh[l] >> 6) & 3) << 4)) as i32 - 32;
                out[base + l] = d * sch[is] as i8 as f32 * q1 as f32;
                out[base + l + 32] = d * sch[is + 2] as i8 as f32 * q2 as f32;
                out[base + l + 64] = d * sch[is + 4] as i8 as f32 * q3 as f32;
                out[base + l + 96] = d * sch[is + 6] as i8 as f32 * q4 as f32;
            }
        }
    }
    Ok(out)
}

/// IEEE binary16 bits → f32, exact. A private twin of `cpu_gemm::f16_to_f32` (that module is
/// feature-gated out of wasm builds, and this reader must build everywhere); the canonical
/// copy carries the exhaustive round-trip test.
fn f16_bits_to_f32(bits: u16) -> f32 {
    let sign = u32::from(bits >> 15) << 31;
    let exp = u32::from(bits >> 10) & 0x1f;
    let man = u32::from(bits) & 0x3ff;
    if exp == 0x1f {
        return f32::from_bits(sign | 0x7f80_0000 | (man << 13));
    }
    if exp == 0 {
        if man == 0 {
            return f32::from_bits(sign);
        }
        let shift = man.leading_zeros() - 21;
        let man = (man << shift) & 0x3ff;
        let exp = 127 - 15 + 1 - shift;
        return f32::from_bits(sign | (exp << 23) | (man << 13));
    }
    f32::from_bits(sign | ((exp + 127 - 15) << 23) | (man << 13))
}

/// One tensor's table entry: dims are in GGUF order (fastest-varying first — the REVERSE of
/// the row-major `[out, in]` the engine writes).
#[derive(Debug, Clone)]
pub struct GgufTensor {
    pub name: String,
    pub dims: Vec<u64>,
    /// Raw ggml type id (0 = F32, 1 = F16, … — Q1-class ids vary by fork; callers match on
    /// what they find and [`GgufFile::type_name`] labels the known ones).
    pub ggml_type: u32,
    /// Byte offset within the DATA section (already aligned by the writer).
    pub offset: u64,
}

/// A parsed GGUF header: metadata subset + tensor table + where the data section starts.
/// Tensor payloads are read on demand ([`GgufFile::read_tensor_raw`]) — a 3.8 GB file costs
/// only its header until asked.
pub struct GgufFile {
    file: std::fs::File,
    pub version: u32,
    pub kvs: std::collections::BTreeMap<String, GgufValue>,
    pub tensors: Vec<GgufTensor>,
    pub data_start: u64,
}

/// The metadata value subset the loaders consult. Arrays keep only their element count unless
/// they are string arrays (tokenizer vocab) or numeric arrays small enough to matter.
#[derive(Debug, Clone)]
pub enum GgufValue {
    U64(u64),
    I64(i64),
    F64(f64),
    Bool(bool),
    Str(String),
    /// (element type id, length) — large arrays are skipped, not stored.
    Array(u32, u64),
    StrArray(Vec<String>),
}

impl GgufFile {
    pub fn open(path: &std::path::Path) -> Result<Self> {
        let mut f = std::fs::File::open(path).with_context(|| format!("open {path:?}"))?;
        let mut r = std::io::BufReader::new(&mut f);
        let mut magic = [0u8; 4];
        r.read_exact(&mut magic)?;
        ensure!(&magic == b"GGUF", "not a GGUF file (magic {magic:?})");
        let version = read_u32(&mut r)?;
        ensure!(
            (2..=3).contains(&version),
            "GGUF v{version} unsupported (v2/v3 only)"
        );
        let n_tensors = read_u64(&mut r)?;
        let n_kvs = read_u64(&mut r)?;
        let mut kvs = std::collections::BTreeMap::new();
        for _ in 0..n_kvs {
            let key = read_string(&mut r)?;
            let ty = read_u32(&mut r)?;
            let val = read_value(&mut r, ty, &key)?;
            kvs.insert(key, val);
        }
        let mut tensors = Vec::with_capacity(n_tensors as usize);
        for _ in 0..n_tensors {
            let name = read_string(&mut r)?;
            let n_dims = read_u32(&mut r)?;
            ensure!(n_dims <= 8, "tensor {name}: {n_dims} dims");
            let mut dims = Vec::with_capacity(n_dims as usize);
            for _ in 0..n_dims {
                dims.push(read_u64(&mut r)?);
            }
            let ggml_type = read_u32(&mut r)?;
            let offset = read_u64(&mut r)?;
            tensors.push(GgufTensor {
                name,
                dims,
                ggml_type,
                offset,
            });
        }
        // Data section starts at the next alignment boundary after the header.
        let align = match kvs.get("general.alignment") {
            Some(GgufValue::U64(a)) if *a > 0 => *a,
            _ => 32,
        };
        let here = r.stream_position()?;
        let data_start = here.div_ceil(align) * align;
        drop(r);
        Ok(Self {
            file: f,
            version,
            kvs,
            tensors,
            data_start,
        })
    }

    /// Raw bytes of one tensor's payload (caller sizes it from dims × type).
    pub fn read_tensor_raw(&mut self, t: &GgufTensor, len: usize) -> Result<Vec<u8>> {
        self.file
            .seek(SeekFrom::Start(self.data_start + t.offset))?;
        let mut buf = vec![0u8; len];
        self.file.read_exact(&mut buf)?;
        Ok(buf)
    }

    /// Row-major logical shape `[out, in]` of a tensor — GGUF stores dims fastest-first, so this
    /// reverses them. 1-D tensors report `[n]`.
    pub fn shape(&self, name: &str) -> Result<Vec<usize>> {
        let t = self.tensor(name)?;
        Ok(t.dims.iter().rev().map(|&d| d as usize).collect())
    }

    fn tensor(&self, name: &str) -> Result<&GgufTensor> {
        self.tensors
            .iter()
            .find(|t| t.name == name)
            .with_context(|| format!("GGUF missing tensor {name}"))
    }

    /// Decode any supported tensor to row-major f32: F32 pass-through, F16 widen, and the
    /// Bonsai Q1_0 (type 41) sign-block format. The result is `[out · in]` row-major — exactly
    /// what `weights::quantize_q1`/`quantize_q4_0` consume.
    pub fn tensor_f32(&mut self, name: &str) -> Result<Vec<f32>> {
        let t = self.tensor(name)?.clone();
        let count: usize = t.dims.iter().map(|&d| d as usize).product();
        match t.ggml_type {
            0 => {
                let raw = self.read_tensor_raw(&t, count * 4)?;
                Ok(raw
                    .chunks_exact(4)
                    .map(|c| f32::from_le_bytes(c.try_into().unwrap()))
                    .collect())
            }
            1 => {
                let raw = self.read_tensor_raw(&t, count * 2)?;
                Ok(raw
                    .chunks_exact(2)
                    .map(|c| f16_bits_to_f32(u16::from_le_bytes([c[0], c[1]])))
                    .collect())
            }
            41 => {
                // Q1_0: `count` weights = `count/128` sign blocks (18 B each), row-major.
                anyhow::ensure!(
                    count.is_multiple_of(128),
                    "{name}: Q1_0 count {count} not /128"
                );
                let nblocks = count / 128;
                let raw = self.read_tensor_raw(&t, nblocks * 18)?;
                let (scales, bits) = q1_0_disk_to_engine(&raw, nblocks)?;
                // The logical shape is [out, in] (dims reversed); dequant over the inner (in) dim.
                let inner: usize = t.dims[0] as usize; // fastest dim = in_features
                let outer = count / inner;
                Ok(dequantize_q1(&scales, &bits, outer, inner))
            }
            // The block-quantized types an Unsloth Dynamic GGUF mixes. `(block_elems, block_bytes)`
            // per type; `dequant_ggml` mirrors llama.cpp (gated bit-for-bit in tests/gguf_kquant.rs).
            2 | 6 | 8 | 10 | 11 | 12 | 13 | 14 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 29
            | 30 => {
                let (be, bb) = match t.ggml_type {
                    8 => (32usize, 34usize),  // Q8_0
                    2 => (32, 18),            // Q4_0
                    6 => (32, 22),            // Q5_0
                    10 => (256, 84),          // Q2_K
                    11 => (256, 110),         // Q3_K
                    12 => (256, 144),         // Q4_K
                    13 => (256, 176),         // Q5_K
                    14 => (256, 210),         // Q6_K
                    16 => (256, 66),          // IQ2_XXS
                    17 => (256, 74),          // IQ2_XS
                    18 => (256, 98),          // IQ3_XXS
                    19 => (256, 50),          // IQ1_S
                    20 => (32, 18),           // IQ4_NL
                    21 => (256, 110),         // IQ3_S
                    22 => (256, 82),          // IQ2_S
                    23 => (256, 136),         // IQ4_XS
                    29 => (256, 56),          // IQ1_M
                    30 => (1, 2),             // BF16
                    _ => unreachable!(),
                };
                anyhow::ensure!(
                    count.is_multiple_of(be),
                    "{name}: {} count {count} not /{be}",
                    Self::type_name(t.ggml_type)
                );
                let raw = self.read_tensor_raw(&t, (count / be) * bb)?;
                dequant_ggml(&raw, t.ggml_type, count)
            }
            other => bail!(
                "{name}: ggml type {other} ({}) not decodable",
                Self::type_name(other)
            ),
        }
    }

    /// Human label for the ggml type ids this module knows about.
    pub fn type_name(ty: u32) -> &'static str {
        match ty {
            0 => "F32",
            1 => "F16",
            2 => "Q4_0",
            8 => "Q8_0",
            16 => "IQ2_XXS",
            17 => "IQ2_XS",
            18 => "IQ3_XXS",
            19 => "IQ1_S",
            20 => "IQ4_NL",
            21 => "IQ3_S",
            22 => "IQ2_S",
            23 => "IQ4_XS",
            29 => "IQ1_M",
            30 => "BF16",
            34 => "TQ1_0",
            35 => "TQ2_0",
            // The prism-ml Bonsai fork's binary type (18-byte/128-weight blocks — see the
            // module docs); `Bonsai-27B-Q1_0.gguf` carries 498 of these under arch "qwen35".
            41 => "Q1_0",
            other => {
                // Fork-specific ids (the Bonsai Q1_0 among them) land here; the caller sees
                // the raw id alongside.
                let _ = other;
                "unknown"
            }
        }
    }
}

fn read_u32(r: &mut impl Read) -> Result<u32> {
    let mut b = [0u8; 4];
    r.read_exact(&mut b)?;
    Ok(u32::from_le_bytes(b))
}

fn read_u64(r: &mut impl Read) -> Result<u64> {
    let mut b = [0u8; 8];
    r.read_exact(&mut b)?;
    Ok(u64::from_le_bytes(b))
}

fn read_string(r: &mut impl Read) -> Result<String> {
    let len = read_u64(r)?;
    ensure!(len <= 1 << 24, "unreasonable GGUF string length {len}");
    let mut b = vec![0u8; len as usize];
    r.read_exact(&mut b)?;
    Ok(String::from_utf8_lossy(&b).into_owned())
}

/// Scalar type sizes for skipping array payloads (GGUF metadata type ids).
fn scalar_size(ty: u32) -> Result<u64> {
    Ok(match ty {
        0 | 1 | 7 => 1, // u8, i8, bool
        2 | 3 => 2,     // u16, i16
        4..=6 => 4,     // u32, i32, f32
        10..=12 => 8,   // u64, i64, f64
        other => bail!("non-scalar GGUF type {other} where scalar expected"),
    })
}

fn read_value(r: &mut (impl Read + Seek), ty: u32, key: &str) -> Result<GgufValue> {
    Ok(match ty {
        0 => GgufValue::U64(u64::from({
            let mut b = [0u8; 1];
            r.read_exact(&mut b)?;
            b[0]
        })),
        1 => GgufValue::I64(i64::from({
            let mut b = [0u8; 1];
            r.read_exact(&mut b)?;
            b[0] as i8
        })),
        2 => GgufValue::U64(u64::from(read_u16(r)?)),
        3 => GgufValue::I64(i64::from(read_u16(r)? as i16)),
        4 => GgufValue::U64(u64::from(read_u32(r)?)),
        5 => GgufValue::I64(i64::from(read_u32(r)? as i32)),
        6 => GgufValue::F64(f64::from(f32::from_bits(read_u32(r)?))),
        7 => GgufValue::Bool({
            let mut b = [0u8; 1];
            r.read_exact(&mut b)?;
            b[0] != 0
        }),
        8 => GgufValue::Str(read_string(r)?),
        9 => {
            // Array: (elem type, count, payload). Keep string arrays under 1M entries (the
            // tokenizer tables a loader needs); skip everything else by size.
            let elem = read_u32(r)?;
            let count = read_u64(r)?;
            if elem == 8 {
                ensure!(count <= 1 << 20, "{key}: string array of {count}");
                let mut v = Vec::with_capacity(count as usize);
                for _ in 0..count {
                    v.push(read_string(r)?);
                }
                GgufValue::StrArray(v)
            } else {
                let sz = scalar_size(elem)?;
                r.seek(SeekFrom::Current((sz * count) as i64))?;
                GgufValue::Array(elem, count)
            }
        }
        10 => GgufValue::U64(read_u64(r)?),
        11 => GgufValue::I64({
            let mut b = [0u8; 8];
            r.read_exact(&mut b)?;
            i64::from_le_bytes(b)
        }),
        12 => GgufValue::F64({
            let mut b = [0u8; 8];
            r.read_exact(&mut b)?;
            f64::from_le_bytes(b)
        }),
        other => bail!("{key}: unknown GGUF value type {other}"),
    })
}

fn read_u16(r: &mut impl Read) -> Result<u16> {
    let mut b = [0u8; 2];
    r.read_exact(&mut b)?;
    Ok(u16::from_le_bytes(b))
}

/// Load a Bonsai-class GGUF (`general.architecture = "qwen35"`, binary Q1_0 weights) into the
/// engine's [`weights::Weights`], keeping the weights BINARY (packed into the `Q4` containers
/// the plan already binds — the M=1 plan runs the Q1 kernel twins under `OSFKB_Q1_WEIGHTS=1`).
///
/// This is deliberately built from the PUBLIC weights API only, so it never touches
/// `weights.rs`. The layer/embed mapping is the GGUF twin of `build_qwen35_layer`:
/// DeltaNet `attn_qkv|attn_gate|ssm_beta|ssm_alpha` → the fused in_proj; attention's doubled
/// `attn_q` de-interleaves into q- and gate-rows; norms are used DIRECTLY (llama.cpp folds the
/// `1+w` delta at conversion, unlike the safetensors path's [`get_p1_at`]).
#[cfg(not(target_arch = "wasm32"))]
pub fn load_bonsai(ctx: &crate::GpuCtx, path: &std::path::Path) -> Result<crate::weights::Weights> {
    use crate::weights::{Arch, EdgeCfg, Layer, Lfm2Config, Op, Q4, WDtype, Weights, f32_to_f16_bytes};
    let mut g = GgufFile::open(path)?;

    let ku = |k: &str| -> Result<usize> {
        match g.kvs.get(k) {
            Some(GgufValue::U64(v)) => Ok(*v as usize),
            _ => bail!("GGUF missing u64 metadata {k}"),
        }
    };
    let kf = |k: &str| -> Result<f32> {
        match g.kvs.get(k) {
            Some(GgufValue::F64(v)) => Ok(*v as f32),
            _ => bail!("GGUF missing f32 metadata {k}"),
        }
    };
    let arch = match g.kvs.get("general.architecture") {
        Some(GgufValue::Str(s)) if s == "qwen35" => Arch::Qwen35,
        other => bail!("load_bonsai: architecture {other:?} != qwen35"),
    };
    let hidden = ku("qwen35.embedding_length")?;
    let n_layers = ku("qwen35.block_count")?;
    let n_heads = ku("qwen35.attention.head_count")?;
    let n_kv_heads = ku("qwen35.attention.head_count_kv")?;
    let head_dim = ku("qwen35.attention.key_length")?;
    let intermediate = ku("qwen35.feed_forward_length")?;
    let eps = kf("qwen35.attention.layer_norm_rms_epsilon")?;
    let rope_theta = kf("qwen35.rope.freq_base")?;
    let rotary_dim = ku("qwen35.rope.dimension_count")?;
    let full_iv = ku("qwen35.full_attention_interval")?;
    let vocab = g.shape("token_embd.weight")?[0];
    // DeltaNet geometry from the ssm metadata: nv = time_step_rank, dv = state_size (inner =
    // nv·dv), nk = group_count, dk = (in_proj_qkv_width − nv·dv) / (2·nk).
    let dn_nv = ku("qwen35.ssm.time_step_rank")?;
    let dn_dv = ku("qwen35.ssm.state_size")?;
    let dn_kernel = ku("qwen35.ssm.conv_kernel")?;
    let dn_nk = ku("qwen35.ssm.group_count")?;
    let inner = ku("qwen35.ssm.inner_size")?; // nv·dv
    anyhow::ensure!(inner == dn_nv * dn_dv, "ssm inner {inner} != nv·dv");
    let qkv_w = g.shape("blk.0.attn_qkv.weight")?[0]; // 2·nk·dk + nv·dv
    let dn_dk = (qkv_w - inner) / (2 * dn_nk);
    anyhow::ensure!(2 * dn_nk * dn_dk + inner == qkv_w, "DeltaNet dk derivation");
    let layer_is_attn: Vec<bool> = (0..n_layers).map(|i| (i + 1) % full_iv == 0).collect();

    let cfg = Lfm2Config {
        arch,
        hidden,
        vocab,
        n_layers,
        intermediate,
        n_heads,
        n_kv_heads,
        head_dim,
        conv_l: 0,
        eps,
        rope_theta,
        rope_local_theta: rope_theta,
        sliding_window: 0,
        layer_is_attn: layer_is_attn.clone(),
        layer_is_sliding: vec![false; n_layers],
        layer_k_eq_v: vec![false; n_layers],
        act_gelu: false, // Qwen3.5 SwiGLU is SiLU-gated
        num_experts: 0,
        top_k_experts: 0,
        moe_intermediate: 0,
        stage_first: true,
        stage_last: true,
        rotary_dim,
        mrope_section: None, // text path — M-RoPE degenerates to 1-D
        dn_nk,
        dn_nv,
        dn_dk,
        dn_dv,
        dn_kernel,
        // The Bonsai/BitNet path packs BINARY weights into the Q4 containers and swaps the
        // pipelines via `OSFKB_Q1_WEIGHTS`; that predates this field and is not a `WDtype` arm,
        // so the config records Q4 (the container shape) and the Q1 flag still selects kernels.
        wdtype: WDtype::Q4,
            head_wdtype: WDtype::Q4,
            layer_wdtype: Vec::new(),
            layer_kinds: Vec::new(),
            head_kind: None,
            edge: EdgeCfg::default(),
    };

    // Q1-in-Q4 container: pack an `[m, n]` f32 (decoded from GGUF) into sign bits + f16 scales.
    let mk_q1 = |g: &mut GgufFile, name: &str| -> Result<Q4> {
        let sh = g.shape(name)?;
        let (m, n) = (sh[0], sh[1]);
        let f = g.tensor_f32(name)?;
        let (scales, bits) = quantize_q1(&f, m, n);
        Ok(Q4 {
            scales: ctx.storage_bytes(&f32_to_f16_bytes(&scales)),
            quants: ctx.storage(bytemuck::cast_slice::<u32, f32>(&bits)),
        })
    };
    // Q1 container from an already-assembled `[m, n]` f32 (concatenated / de-interleaved rows).
    let mk_q1_f = |f: &[f32], m: usize, n: usize| -> Q4 {
        let (scales, bits) = quantize_q1(f, m, n);
        Q4 {
            scales: ctx.storage_bytes(&f32_to_f16_bytes(&scales)),
            quants: ctx.storage(bytemuck::cast_slice::<u32, f32>(&bits)),
        }
    };
    let norm = |g: &mut GgufFile, name: &str| -> Result<wgpu::Buffer> {
        Ok(ctx.storage(&g.tensor_f32(name)?)) // GGUF folds the 1+w delta already
    };

    let (nk, nv, dk, dv) = (dn_nk, dn_nv, dn_dk, dn_dv);
    let (qd, kd) = (n_heads * head_dim, n_kv_heads * head_dim);
    let mut layers = Vec::with_capacity(n_layers);
    for li in 0..n_layers {
        let p = format!("blk.{li}");
        let op = if layer_is_attn[li] {
            // Doubled q_proj: per head [q(head_dim) | gate(head_dim)] interleaved.
            let qp = g.tensor_f32(&format!("{p}.attn_q.weight"))?;
            let hd2 = 2 * head_dim;
            let mut qrows = vec![0f32; qd * hidden];
            let mut grows = vec![0f32; qd * hidden];
            for h in 0..n_heads {
                let src = h * hd2 * hidden;
                let dst = h * head_dim * hidden;
                let half = head_dim * hidden;
                qrows[dst..dst + half].copy_from_slice(&qp[src..src + half]);
                grows[dst..dst + half].copy_from_slice(&qp[src + half..src + 2 * half]);
            }
            let kp = g.tensor_f32(&format!("{p}.attn_k.weight"))?;
            let vp = g.tensor_f32(&format!("{p}.attn_v.weight"))?;
            let mut qkv_f = qrows;
            qkv_f.extend_from_slice(&kp);
            qkv_f.extend_from_slice(&vp);
            Op::Attn {
                qkv: mk_q1_f(&qkv_f, qd + 2 * kd, hidden),
                o: mk_q1(&mut g, &format!("{p}.attn_output.weight"))?,
                q_norm: norm(&mut g, &format!("{p}.attn_q_norm.weight"))?,
                k_norm: norm(&mut g, &format!("{p}.attn_k_norm.weight"))?,
                window: 0,
                local_rope: false,
                attn_gate: Some(mk_q1_f(&grows, qd, hidden)),
                qkv_bias: None,
            }
        } else {
            // DeltaNet fused in_proj: [attn_qkv | attn_gate(=z) | ssm_beta(=b) | ssm_alpha(=a)].
            let mut in_f = g.tensor_f32(&format!("{p}.attn_qkv.weight"))?;
            in_f.extend_from_slice(&g.tensor_f32(&format!("{p}.attn_gate.weight"))?);
            in_f.extend_from_slice(&g.tensor_f32(&format!("{p}.ssm_beta.weight"))?);
            in_f.extend_from_slice(&g.tensor_f32(&format!("{p}.ssm_alpha.weight"))?);
            let rows = 2 * nk * dk + nv * dv + nv * dv + 2 * nv;
            let conv = g.tensor_f32(&format!("{p}.ssm_conv1d.weight"))?;
            let mut gpar = g.tensor_f32(&format!("{p}.ssm_a"))?;
            gpar.extend_from_slice(&g.tensor_f32(&format!("{p}.ssm_dt.bias"))?);
            gpar.extend_from_slice(&g.tensor_f32(&format!("{p}.ssm_norm.weight"))?);
            Op::DeltaNet {
                in_proj: mk_q1_f(&in_f, rows, hidden),
                conv_w: ctx.storage(&conv),
                gpar: ctx.storage(&gpar),
                out_proj: mk_q1(&mut g, &format!("{p}.ssm_out.weight"))?,
            }
        };
        let layer = Layer {
            operator_norm: norm(&mut g, &format!("{p}.attn_norm.weight"))?,
            ffn_norm: norm(&mut g, &format!("{p}.post_attention_norm.weight"))?,
            op,
            w1: mk_q1(&mut g, &format!("{p}.ffn_gate.weight"))?,
            w2: mk_q1(&mut g, &format!("{p}.ffn_down.weight"))?,
            w3: mk_q1(&mut g, &format!("{p}.ffn_up.weight"))?,
            post_op_norm: None,
            post_ffn_norm: None,
            moe: None,
            edge: None,
        };
        layers.push(layer);
        ctx.queue.submit(std::iter::empty());
        let _ = ctx.device.poll(wgpu::PollType::wait_indefinitely());
        eprintln!("bonsai load: layer {li} done");
    }

    // TWO copies of the gather table, each for what only it can do: the f32 `embed` feeds the
    // batch/spec staging paths, which COPY rows out of it (copies have no size cap anywhere);
    // the f16 `embed_f16` feeds the gather KERNELS, whose storage bindings cap at 2 GiB-4 on
    // wgpu-Vulkan — 2.5 GB of f16 binds as two row-aligned halves, 5 GB of f32 cannot.
    let ef32 = g.tensor_f32("token_embd.weight")?;
    // f32 halves for the staging COPIES (Vulkan max_buffer_size ~4 GiB < the 5 GB table)...
    let split_row = (cfg.vocab / 2) as u32;
    let split_at = (split_row as usize) * cfg.hidden;
    let embed = ctx.storage(&ef32[..split_at]);
    let embed_split = Some((ctx.storage(&ef32[split_at..]), split_row));
    // ...and the f16 twin for the gather KERNELS (a storage BINDING caps at 2 GiB-4).
    let embed_f16 = Some(ctx.storage_bytes(&f32_to_f16_bytes(&ef32)));
    drop(ef32);
    let embed_q4 = mk_q1(&mut g, "output.weight")?;
    let embedding_norm = norm(&mut g, "output_norm.weight")?;
    eprintln!("bonsai load: embed + head done");

    Ok(Weights {
        cfg,
        embed,
        embed_q4,
        embedding_norm,
        layers,
        mtp: None,
        embed_f16,
        embed_split,
        ple: None,
    })
}

/// Load a mainstream **llama-architecture** GGUF (`general.architecture = "llama"`, e.g. a
/// Llama/Mistral/SmolLM checkpoint, INCLUDING an Unsloth Dynamic K-quant mix) into the engine's
/// [`weights::Weights`], REPACKING every projection to `wdtype` (`Q4` or `F16`).
///
/// This is the payoff of the K-quant decoders: each tensor is decoded to f32 by
/// [`GgufFile::tensor_f32`] (which now handles Q4_0/Q5_0/Q8_0/Q4_K/Q6_K), then repacked. Loading
/// to `F16` is a near-lossless repack, so the KLD ruler comparing a GGUF-`F16` arm to the
/// safetensors-`F16` arm isolates **the GGUF file's own quantization damage** — the number that
/// answers "is an Unsloth Dynamic artifact better than our Q4 at equal bytes?".
///
/// llama specifics vs the Bonsai/qwen35 path: separate `attn_q|k|v` (concatenated here into one
/// GEMV), no qk-norm (ones buffers satisfy the bind-group layout; the llama plan skips the norm),
/// no DeltaNet, raw RMSNorm weights (llama does not fold `1+w`).
#[cfg(not(target_arch = "wasm32"))]
pub fn load_gguf(
    ctx: &crate::GpuCtx,
    path: &std::path::Path,
    wdtype: crate::weights::WDtype,
) -> Result<crate::weights::Weights> {
    use crate::weights::{Arch, EdgeCfg, Layer, Lfm2Config, Op, Weights, pack_weight};
    let mut g = GgufFile::open(path)?;
    let arch_name = match g.kvs.get("general.architecture") {
        Some(GgufValue::Str(s)) => s.clone(),
        other => bail!("load_gguf: missing general.architecture ({other:?})"),
    };
    ensure!(
        arch_name == "llama" || arch_name == "qwen2",
        "load_gguf handles the `llama`/`qwen2` arches (got {arch_name:?}); load_bonsai for qwen35"
    );
    // llama.cpp PERMUTES q/k at conversion for the llama arch (its rope rotates adjacent pairs);
    // the qwen2 arch uses neox-mode rope and is stored in HF order — no un-permute, and the
    // q/k/v BIAS vectors concatenate in the same row order as the fused qkv weight.
    let is_llama = arch_name == "llama";
    let ku = |g: &GgufFile, k: &str| -> Result<usize> {
        match g.kvs.get(k) {
            Some(GgufValue::U64(v)) => Ok(*v as usize),
            _ => bail!("GGUF missing u64 metadata {k}"),
        }
    };
    let kf = |g: &GgufFile, k: &str| -> Result<f32> {
        match g.kvs.get(k) {
            Some(GgufValue::F64(v)) => Ok(*v as f32),
            _ => bail!("GGUF missing f32 metadata {k}"),
        }
    };
    let pre = arch_name.clone();
    let hidden = ku(&g, &format!("{pre}.embedding_length"))?;
    let n_layers = ku(&g, &format!("{pre}.block_count"))?;
    let n_heads = ku(&g, &format!("{pre}.attention.head_count"))?;
    let n_kv_heads = ku(&g, &format!("{pre}.attention.head_count_kv"))?;
    let intermediate = ku(&g, &format!("{pre}.feed_forward_length"))?;
    let eps = kf(&g, &format!("{pre}.attention.layer_norm_rms_epsilon"))?;
    let rope_theta = kf(&g, &format!("{pre}.rope.freq_base")).unwrap_or(10_000.0);
    // head_dim: prefer key_length, else rope.dimension_count, else hidden/heads.
    let head_dim = ku(&g, &format!("{pre}.attention.key_length"))
        .or_else(|_| ku(&g, &format!("{pre}.rope.dimension_count")))
        .unwrap_or(hidden / n_heads);
    let rotary_dim = ku(&g, &format!("{pre}.rope.dimension_count")).unwrap_or(head_dim);
    let vocab = g.shape("token_embd.weight")?[0];

    let cfg = Lfm2Config {
        arch: if is_llama { Arch::Llama } else { Arch::Qwen2 },
        hidden,
        vocab,
        n_layers,
        intermediate,
        n_heads,
        n_kv_heads,
        head_dim,
        conv_l: 0,
        eps,
        rope_theta,
        rope_local_theta: rope_theta,
        sliding_window: 0,
        layer_is_attn: vec![true; n_layers],
        layer_is_sliding: vec![false; n_layers],
        layer_k_eq_v: vec![false; n_layers],
        act_gelu: false, // llama SwiGLU is SiLU-gated
        num_experts: 0,
        top_k_experts: 0,
        moe_intermediate: 0,
        stage_first: true,
        stage_last: true,
        rotary_dim,
        mrope_section: None,
        dn_nk: 0,
        dn_nv: 0,
        dn_dk: 0,
        dn_dv: 0,
        dn_kernel: 0,
        wdtype,
        head_wdtype: wdtype,
        layer_wdtype: Vec::new(),
            layer_kinds: Vec::new(),
            head_kind: None,
            edge: EdgeCfg::default(),
    };

    // Decode a GGUF tensor to f32 and repack to the target dtype.
    let pk = |g: &mut GgufFile, name: &str| -> Result<crate::weights::Q4> {
        let sh = g.shape(name)?;
        let (rows, cols) = (sh[0], sh[1]);
        let f = g.tensor_f32(name)?;
        pack_weight(ctx, &f, rows, cols, wdtype)
    };
    // llama.cpp's converter PERMUTES q/k rows (interleaves the two rotary halves) so its RoPE
    // rotates adjacent pairs; the engine uses HF-style RoPE (rotate-halves), so q/k must be
    // permuted BACK to HF order or every attention score is wrong. `v` is never permuted.
    // Inverse of convert_hf_to_gguf's `permute`: for GGUF within-head index g, a=g&1, b=g>>1,
    // the HF index is a·(hd/2)+b.
    let un_permute = |w: &[f32], n_head: usize| -> Vec<f32> {
        if !is_llama {
            return w.to_vec(); // qwen2: stored in HF order already
        }
        let hd = head_dim;
        let mut out = vec![0f32; w.len()];
        for h in 0..n_head {
            for g in 0..hd {
                let (a, b) = (g & 1, g >> 1);
                let f = a * (hd / 2) + b;
                let (src, dst) = ((h * hd + g) * hidden, (h * hd + f) * hidden);
                out[dst..dst + hidden].copy_from_slice(&w[src..src + hidden]);
            }
        }
        out
    };
    // Concatenate q|k|v (q,k un-permuted), then repack as one — the shape the qkv GEMV binds.
    let pk_qkv = |g: &mut GgufFile, li: usize| -> Result<crate::weights::Q4> {
        let (qd, kd) = (n_heads * head_dim, n_kv_heads * head_dim);
        let qf = un_permute(&g.tensor_f32(&format!("blk.{li}.attn_q.weight"))?, n_heads);
        let kf = un_permute(&g.tensor_f32(&format!("blk.{li}.attn_k.weight"))?, n_kv_heads);
        let mut f = qf;
        f.extend_from_slice(&kf);
        f.extend_from_slice(&g.tensor_f32(&format!("blk.{li}.attn_v.weight"))?);
        pack_weight(ctx, &f, qd + 2 * kd, hidden, wdtype)
    };
    let norm = |g: &mut GgufFile, name: &str| -> Result<wgpu::Buffer> {
        Ok(ctx.storage(&g.tensor_f32(name)?)) // llama uses raw RMSNorm weights
    };

    let mut layers = Vec::with_capacity(n_layers);
    for li in 0..n_layers {
        // Qwen2: fused q|k|v bias in the SAME row order as the fused qkv weight.
        let qkv_bias = if !is_llama {
            let mut b = g.tensor_f32(&format!("blk.{li}.attn_q.bias"))?;
            b.extend_from_slice(&g.tensor_f32(&format!("blk.{li}.attn_k.bias"))?);
            b.extend_from_slice(&g.tensor_f32(&format!("blk.{li}.attn_v.bias"))?);
            Some(ctx.storage(&b))
        } else {
            None
        };
        let op = Op::Attn {
            qkv: pk_qkv(&mut g, li)?,
            o: pk(&mut g, &format!("blk.{li}.attn_output.weight"))?,
            // llama/qwen2 have no qk-norm; the plan skips it, buffers satisfy the layout.
            q_norm: ctx.storage(&vec![1.0f32; head_dim]),
            k_norm: ctx.storage(&vec![1.0f32; head_dim]),
            window: 0,
            local_rope: false,
            attn_gate: None,
            qkv_bias,
        };
        layers.push(Layer {
            operator_norm: norm(&mut g, &format!("blk.{li}.attn_norm.weight"))?,
            ffn_norm: norm(&mut g, &format!("blk.{li}.ffn_norm.weight"))?,
            op,
            w1: pk(&mut g, &format!("blk.{li}.ffn_gate.weight"))?,
            w2: pk(&mut g, &format!("blk.{li}.ffn_down.weight"))?,
            w3: pk(&mut g, &format!("blk.{li}.ffn_up.weight"))?,
            post_op_norm: None,
            post_ffn_norm: None,
            moe: None,
            edge: None,
        });
    }

    let embed = ctx.storage(&g.tensor_f32("token_embd.weight")?);
    // Untied head if `output.weight` is present, else tied to the embedding table.
    let head_name = if g.tensors.iter().any(|t| t.name == "output.weight") {
        "output.weight"
    } else {
        "token_embd.weight"
    };
    let embed_q4 = pk(&mut g, head_name)?;
    let embedding_norm = norm(&mut g, "output_norm.weight")?;

    Ok(Weights {
        cfg,
        embed,
        embed_q4,
        embedding_norm,
        layers,
        mtp: None,
        embed_f16: None,
        embed_split: None,
        ple: None,
    })
}

/// Load a llama-arch GGUF **natively**: every projection keeps the artifact's OWN block
/// formats (padded to word alignment; values untouched), so it serves at the artifact's bytes
/// with the artifact's fidelity — the whole point of Unsloth-style dynamic quants.
///
/// Per-site container policy (v1):
/// * `qkv` (concatenated q|k|v, q/k rotary-UN-PERMUTED at the raw-row level — blocks never
///   straddle rows, so whole-row reorder is value-exact): if all three are legacy 32-block
///   types → **Q8_0N**, via the bitwise-lossless upcast (`d8=d, q8=q−zp`). Mixed/K-quant → f16.
/// * `o` / `down`: Q4_K → native Q4K; Q6_K → native Q6K (padded); Q5_0 → native Q5_0N;
///   Q8_0 → native Q8_0N; Q4_0 → Q8_0N upcast; anything else → f16 repack.
/// * `gate_up`: both legacy → both Q8_0N (one fused kernel); else → f16 repack.
/// * `head`: legacy → Q8_0N (the q8_0n lmhead); else → f16 repack. The gather table stays f32.
#[cfg(not(target_arch = "wasm32"))]
pub fn load_gguf_native(
    ctx: &crate::GpuCtx,
    path: &std::path::Path,
) -> Result<crate::weights::Weights> {
    use crate::weights::{Arch, EdgeCfg, Layer, LayerKinds, Lfm2Config, Op, Q4, WDtype, WKind, Weights};
    let mut g = GgufFile::open(path)?;
    let arch_name = match g.kvs.get("general.architecture") {
        Some(GgufValue::Str(s)) => s.clone(),
        other => bail!("load_gguf_native: missing general.architecture ({other:?})"),
    };
    ensure!(
        arch_name == "llama" || arch_name == "qwen2",
        "load_gguf_native handles the `llama`/`qwen2` arches"
    );
    let is_llama = arch_name == "llama";
    let pre = arch_name.clone();
    let ku = |g: &GgufFile, k: &str| -> Result<usize> {
        match g.kvs.get(k) {
            Some(GgufValue::U64(v)) => Ok(*v as usize),
            _ => bail!("GGUF missing u64 metadata {k}"),
        }
    };
    let kf = |g: &GgufFile, k: &str| -> Result<f32> {
        match g.kvs.get(k) {
            Some(GgufValue::F64(v)) => Ok(*v as f32),
            _ => bail!("GGUF missing f32 metadata {k}"),
        }
    };
    let hidden = ku(&g, &format!("{pre}.embedding_length"))?;
    let n_layers = ku(&g, &format!("{pre}.block_count"))?;
    let n_heads = ku(&g, &format!("{pre}.attention.head_count"))?;
    let n_kv_heads = ku(&g, &format!("{pre}.attention.head_count_kv"))?;
    let intermediate = ku(&g, &format!("{pre}.feed_forward_length"))?;
    let eps = kf(&g, &format!("{pre}.attention.layer_norm_rms_epsilon"))?;
    let rope_theta = kf(&g, &format!("{pre}.rope.freq_base")).unwrap_or(10_000.0);
    let head_dim = ku(&g, &format!("{pre}.attention.key_length"))
        .or_else(|_| ku(&g, &format!("{pre}.rope.dimension_count")))
        .unwrap_or(hidden / n_heads);
    let rotary_dim = ku(&g, &format!("{pre}.rope.dimension_count")).unwrap_or(head_dim);
    let vocab = g.shape("token_embd.weight")?[0];
    ensure!(hidden.is_multiple_of(32), "native serving needs hidden % 32 == 0");

    let mut cfg = Lfm2Config {
        arch: if is_llama { Arch::Llama } else { Arch::Qwen2 },
        hidden,
        vocab,
        n_layers,
        intermediate,
        n_heads,
        n_kv_heads,
        head_dim,
        conv_l: 0,
        eps,
        rope_theta,
        rope_local_theta: rope_theta,
        sliding_window: 0,
        layer_is_attn: vec![true; n_layers],
        layer_is_sliding: vec![false; n_layers],
        layer_k_eq_v: vec![false; n_layers],
        act_gelu: false,
        num_experts: 0,
        top_k_experts: 0,
        moe_intermediate: 0,
        stage_first: true,
        stage_last: true,
        rotary_dim,
        mrope_section: None,
        dn_nk: 0,
        dn_nv: 0,
        dn_dk: 0,
        dn_dv: 0,
        dn_kernel: 0,
        wdtype: WDtype::Q4,
        head_wdtype: WDtype::Q4,
        layer_wdtype: Vec::new(),
        layer_kinds: Vec::new(),
        head_kind: None,
        edge: EdgeCfg::default(),
    };

    // Raw bytes + geometry of one tensor. `rows` = out features, `cols` = in features (row len).
    let raw_of = |g: &mut GgufFile, name: &str| -> Result<(u32, usize, usize, Vec<u8>)> {
        let t = g
            .tensors
            .iter()
            .find(|t| t.name == name)
            .with_context(|| format!("GGUF missing tensor {name}"))?
            .clone();
        let cols = t.dims[0] as usize;
        let rows: usize = t.dims.iter().skip(1).map(|&d| d as usize).product();
        let (be, bb) = match t.ggml_type {
            2 => (32usize, 18usize),
            6 => (32, 22),
            8 => (32, 34),
            12 => (256, 144),
            13 => (256, 176), // Q5_K — MUST be here: single_native serves it natively
            14 => (256, 210),
            20 => (32, 18),   // IQ4_NL
            23 => (256, 136), // IQ4_XS
            16 => (256, 66),  // IQ2_XXS
            17 => (256, 74),  // IQ2_XS
            22 => (256, 82),  // IQ2_S
            18 => (256, 98),  // IQ3_XXS
            21 => (256, 110), // IQ3_S
            19 => (256, 50),  // IQ1_S
            29 => (256, 56),  // IQ1_M
            other => {
                // Not a block type we serve natively — signal with be=0; caller falls back.
                let _ = other;
                (0, 0)
            }
        };
        if be == 0 {
            return Ok((t.ggml_type, rows, cols, Vec::new()));
        }
        ensure!(cols.is_multiple_of(be), "{name}: cols {cols} not /{be}");
        let raw = g.read_tensor_raw(&t, rows * (cols / be) * bb)?;
        Ok((t.ggml_type, rows, cols, raw))
    };
    // Legacy 32-block row run → padded Q8_0N row run (bitwise-lossless).
    let to_q8n = |ty: u32, raw: &[u8], nblocks: usize| -> Result<Vec<u8>> {
        Ok(match ty {
            8 => crate::q8_0_pad_blocks(raw, nblocks),
            6 => crate::q5_0_to_q8_0n(raw, nblocks),
            2 => crate::q4_0_to_q8_0n(raw, nblocks),
            other => bail!("no lossless Q8_0N upcast for ggml type {other}"),
        })
    };
    // Rotary un-permutation as a whole-ROW reorder of raw block runs (value-exact).
    let unpermute_raw = |raw: &[u8], n_head: usize, hd: usize, row_bytes: usize| -> Vec<u8> {
        if !is_llama {
            return raw.to_vec(); // qwen2: HF row order already
        }
        let mut out = vec![0u8; raw.len()];
        for h in 0..n_head {
            for gidx in 0..hd {
                let (a, b) = (gidx & 1, gidx >> 1);
                let f = a * (hd / 2) + b;
                let (src, dst) = ((h * hd + gidx) * row_bytes, (h * hd + f) * row_bytes);
                out[dst..dst + row_bytes].copy_from_slice(&raw[src..src + row_bytes]);
            }
        }
        out
    };
    // Native container: raw padded blocks ride `quants`; `scales` is the 4-byte stub.
    let native = |ctx: &crate::GpuCtx, padded: &[u8]| -> Q4 {
        Q4 {
            scales: ctx.storage_bytes(&[0u8; 4]),
            quants: ctx.storage_bytes(padded),
        }
    };
    // f16 fallback: decode to f32 (any supported type) and repack as the engine's f16.
    let f16_fallback = |g: &mut GgufFile, name: &str| -> Result<Q4> {
        let sh = g.shape(name)?;
        let f = g.tensor_f32(name)?;
        crate::weights::pack_weight(ctx, &f, sh[0], sh[1], WDtype::F16)
    };
    // Single-tensor site (o / down / head body): pick the native kernel for the type, else f16.
    let single_native = |ctx: &crate::GpuCtx,
                         g: &mut GgufFile,
                         name: &str|
     -> Result<(Q4, WKind)> {
        let (ty, rows, cols, raw) = raw_of(g, name)?;
        // A native arm below with EMPTY raw = raw_of's block table lost a type this match
        // claims — that shipped once as a zero-size bind group (Q5_K). Fail load, not dispatch.
        ensure!(
            raw.is_empty()
                == !matches!(ty, 2 | 6 | 8 | 12 | 13 | 14 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 29),
            "{name}: ggml type {ty} claimed native but raw_of returned {} bytes",
            raw.len()
        );
        Ok(match ty {
            12 => (native(ctx, &raw), WKind::Q4K),
            13 => (native(ctx, &raw), WKind::Q5K),
            14 => (
                native(ctx, &crate::q6k_pad_blocks(&raw, rows * cols / 256)),
                WKind::Q6K,
            ),
            6 => (
                native(ctx, &crate::q5_0_pad_blocks(&raw, rows * cols / 32)),
                WKind::Q5_0N,
            ),
            8 => (
                native(ctx, &crate::q8_0_pad_blocks(&raw, rows * cols / 32)),
                WKind::Q8_0N,
            ),
            2 => (
                native(ctx, &crate::q4_0_to_q8_0n(&raw, rows * cols / 32)),
                WKind::Q8_0N,
            ),
            20 => (
                native(ctx, &crate::iq4_nl_pad_blocks(&raw, rows * cols / 32)),
                WKind::Iq4Nl,
            ),
            23 => (native(ctx, &raw), WKind::Iq4Xs),
            // Grid-codebook family: padded blocks in `quants`, the CODEBOOK in `scales`
            // (the dispatch binds it at slot 0 — the Q4_0 5-binding shape).
            16 | 17 | 18 | 19 | 21 | 22 | 29 => {
                let (words, table, bb) = crate::iq_grid_geom(ty);
                let nb = rows * cols / 256;
                let padded = if bb % 4 == 0 && bb == words * 4 {
                    raw.clone()
                } else {
                    crate::iq_pad_blocks(&raw, nb, bb, words * 4)
                };
                let tb: &[u8] = bytemuck::cast_slice(table);
                let kind = match ty {
                    16 => WKind::Iq2Xxs,
                    17 => WKind::Iq2Xs,
                    22 => WKind::Iq2S,
                    18 => WKind::Iq3Xxs,
                    21 => WKind::Iq3S,
                    19 => WKind::Iq1S,
                    _ => WKind::Iq1M,
                };
                (
                    Q4 {
                        scales: ctx.storage_bytes(tb),
                        quants: ctx.storage_bytes(&padded),
                    },
                    kind,
                )
            }
            _ => (f16_fallback(g, name)?, WKind::F16),
        })
    };

    let (qd, kd) = (n_heads * head_dim, n_kv_heads * head_dim);
    let norm = |g: &mut GgufFile, name: &str| -> Result<wgpu::Buffer> {
        Ok(ctx.storage(&g.tensor_f32(name)?))
    };
    let mut layers = Vec::with_capacity(n_layers);
    let mut kinds = Vec::with_capacity(n_layers);
    for li in 0..n_layers {
        // qkv: all-legacy → un-permute q/k rows raw, upcast all to Q8_0N, concatenate.
        let (qt, _, _, qraw) = raw_of(&mut g, &format!("blk.{li}.attn_q.weight"))?;
        let (kt, _, _, kraw) = raw_of(&mut g, &format!("blk.{li}.attn_k.weight"))?;
        let (vt, _, _, vraw) = raw_of(&mut g, &format!("blk.{li}.attn_v.weight"))?;
        let legacy = |t: u32| matches!(t, 2 | 6 | 8);
        let nb_row = hidden / 32;
        let (qkv, qkv_kind) = if legacy(qt) && legacy(kt) && legacy(vt) {
            let rb = |t: u32| nb_row * match t {
                2 => 18,
                6 => 22,
                _ => 34,
            };
            let qp = unpermute_raw(&qraw, n_heads, head_dim, rb(qt));
            let kp = unpermute_raw(&kraw, n_kv_heads, head_dim, rb(kt));
            let mut blocks = to_q8n(qt, &qp, qd * nb_row)?;
            blocks.extend_from_slice(&to_q8n(kt, &kp, kd * nb_row)?);
            blocks.extend_from_slice(&to_q8n(vt, &vraw, kd * nb_row)?);
            (native(ctx, &blocks), WKind::Q8_0N)
        } else {
            // K-quant or f-type q/k/v: decode + f16 repack (with the f32-level un-permute).
            let un = |w: &[f32], nh: usize| -> Vec<f32> {
                if !is_llama {
                    return w.to_vec();
                }
                let hd = head_dim;
                let mut out = vec![0f32; w.len()];
                for h in 0..nh {
                    for gi in 0..hd {
                        let (a, b) = (gi & 1, gi >> 1);
                        let f = a * (hd / 2) + b;
                        let (src, dst) = ((h * hd + gi) * hidden, (h * hd + f) * hidden);
                        out[dst..dst + hidden].copy_from_slice(&w[src..src + hidden]);
                    }
                }
                out
            };
            let qf = un(&g.tensor_f32(&format!("blk.{li}.attn_q.weight"))?, n_heads);
            let kf2 = un(&g.tensor_f32(&format!("blk.{li}.attn_k.weight"))?, n_kv_heads);
            let mut f = qf;
            f.extend_from_slice(&kf2);
            f.extend_from_slice(&g.tensor_f32(&format!("blk.{li}.attn_v.weight"))?);
            (
                crate::weights::pack_weight(ctx, &f, qd + 2 * kd, hidden, WDtype::F16)?,
                WKind::F16,
            )
        };
        let (o, o_kind) = single_native(ctx, &mut g, &format!("blk.{li}.attn_output.weight"))?;
        // gate/up: both legacy → both Q8_0N (the one fused native MLP); else both f16.
        let (gt, _, _, graw) = raw_of(&mut g, &format!("blk.{li}.ffn_gate.weight"))?;
        let (ut, _, _, uraw) = raw_of(&mut g, &format!("blk.{li}.ffn_up.weight"))?;
        let gu_blocks = intermediate * (hidden / 32);
        let (w1, w3, gu_kind) = if legacy(gt) && legacy(ut) {
            (
                native(ctx, &to_q8n(gt, &graw, gu_blocks)?),
                native(ctx, &to_q8n(ut, &uraw, gu_blocks)?),
                WKind::Q8_0N,
            )
        } else {
            (
                f16_fallback(&mut g, &format!("blk.{li}.ffn_gate.weight"))?,
                f16_fallback(&mut g, &format!("blk.{li}.ffn_up.weight"))?,
                WKind::F16,
            )
        };
        let (w2, w2_kind) = single_native(ctx, &mut g, &format!("blk.{li}.ffn_down.weight"))?;
        kinds.push(LayerKinds {
            qkv: qkv_kind,
            o: o_kind,
            gate_up: gu_kind,
            down: w2_kind,
        });
        let qkv_bias = if !is_llama {
            let mut bv = g.tensor_f32(&format!("blk.{li}.attn_q.bias"))?;
            bv.extend_from_slice(&g.tensor_f32(&format!("blk.{li}.attn_k.bias"))?);
            bv.extend_from_slice(&g.tensor_f32(&format!("blk.{li}.attn_v.bias"))?);
            Some(ctx.storage(&bv))
        } else {
            None
        };
        layers.push(Layer {
            operator_norm: norm(&mut g, &format!("blk.{li}.attn_norm.weight"))?,
            ffn_norm: norm(&mut g, &format!("blk.{li}.ffn_norm.weight"))?,
            op: Op::Attn {
                qkv,
                o,
                q_norm: ctx.storage(&vec![1.0f32; head_dim]),
                k_norm: ctx.storage(&vec![1.0f32; head_dim]),
                window: 0,
                local_rope: false,
                attn_gate: None,
                qkv_bias,
            },
            w1,
            w2,
            w3,
            post_op_norm: None,
            post_ffn_norm: None,
            moe: None,
            edge: None,
        });
    }

    let embed = ctx.storage(&g.tensor_f32("token_embd.weight")?);
    let head_name = if g.tensors.iter().any(|t| t.name == "output.weight") {
        "output.weight"
    } else {
        "token_embd.weight"
    };
    let (embed_q4, hk) = {
        let (ty, rows, cols, raw) = raw_of(&mut g, head_name)?;
        if matches!(ty, 2 | 6 | 8) {
            (native(ctx, &to_q8n(ty, &raw, rows * cols / 32)?), WKind::Q8_0N)
        } else {
            (f16_fallback(&mut g, head_name)?, WKind::F16)
        }
    };
    cfg.layer_kinds = kinds;
    cfg.head_kind = Some(hk);
    let embedding_norm = norm(&mut g, "output_norm.weight")?;

    Ok(Weights {
        cfg,
        embed,
        embed_q4,
        embedding_norm,
        layers,
        mtp: None,
        embed_f16: None,
        embed_split: None,
        ple: None,
    })
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn q1_round_trips_binary_valued_weights() {
        // A weight that IS binary×scale per 128-block must survive pack → dequant exactly.
        let (m, n) = (3usize, 256usize);
        let mut w = vec![0f32; m * n];
        for r in 0..m {
            for j in 0..n {
                let s = 0.02 + 0.01 * (r as f32) + 0.005 * ((j / 128) as f32);
                w[r * n + j] = if (r + j) % 3 == 0 { s } else { -s };
            }
        }
        let (scales, bits) = quantize_q1(&w, m, n);
        let back = dequantize_q1(&scales, &bits, m, n);
        for (a, b) in w.iter().zip(&back) {
            assert!((a - b).abs() < 1e-6, "{a} vs {b}");
        }
    }

    #[test]
    fn disk_blocks_decode_to_the_engine_layout() {
        // Build one synthetic 18-byte disk block and check scale + sign extraction.
        let mut raw = vec![0u8; 18];
        let scale_bits: u16 = 0x2800; // f16(0.03125) — exact power of two
        raw[..2].copy_from_slice(&scale_bits.to_le_bytes());
        raw[2] = 0b1010_0101; // weights 0..7: +,-,+,-,-,+,-,+ (bit i = weight i)
        let (scales, bits) = q1_0_disk_to_engine(&raw, 1).unwrap();
        assert!((scales[0] - 0.03125).abs() < 1e-6);
        assert_eq!(bits[0] & 0xFF, 0b1010_0101);
        assert_eq!(bits.len(), 4);
    }
}

// ─── IQ codebook decoders ───────────────────────────────────────────────────────────────────
// Bit-exact ports of llama.cpp's IQ dequantization (via gguf-python's reference, which the
// fixture tests in tests/gguf_iq.rs pin bit-for-bit). Grids live in crate::iq_tables
// (generated). f32 op ORDER mirrors the reference so equality is exact, not approximate.

use crate::iq_tables as iqt;

fn iq_f16(b: &[u8]) -> f32 {
    half::f16::from_le_bytes([b[0], b[1]]).to_f32()
}

/// llama.cpp's ksigns table, computed: entry i = i with an odd-parity bit in the MSB.
fn ksign(i: u32) -> u8 {
    (i as u8) | (((i.count_ones() & 1) as u8) << 7)
}

fn sgn(byte: u8, bit: usize) -> f32 {
    if (byte >> bit) & 1 == 0 { 1.0 } else { -1.0 }
}

/// IQ2_XXS: 256 elems / 66 bytes — f16 d + 8 groups of (u32 grid bytes, u32 signs+scale).
fn dequant_iq2_xxs(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256) && raw.len() >= count / 256 * 66);
    let mut out = Vec::with_capacity(count);
    for blk in raw[..count / 256 * 66].chunks_exact(66) {
        let d = iq_f16(blk);
        let w32 = |i: usize| {
            u32::from_le_bytes([blk[2 + i * 4], blk[3 + i * 4], blk[4 + i * 4], blk[5 + i * 4]])
        };
        for g in 0..8 {
            let (w0, w1) = (w32(2 * g), w32(2 * g + 1));
            let db = d * (0.5 + (w1 >> 28) as f32) * 0.25;
            for j in 0..4 {
                let row = ((w0 >> (8 * j)) & 0xFF) as usize;
                let sb = ksign((w1 >> (7 * j)) & 0x7F);
                for k in 0..8 {
                    out.push(db * iqt::IQ2_XXS_GRID[row * 8 + k] as f32 * sgn(sb, k));
                }
            }
        }
    }
    Ok(out)
}

/// IQ2_XS: 256 / 74 — f16 d + 32 u16 (9-bit grid | 7-bit sign idx) + 16 scale nibbles.
fn dequant_iq2_xs(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256) && raw.len() >= count / 256 * 74);
    let mut out = Vec::with_capacity(count);
    for blk in raw[..count / 256 * 74].chunks_exact(74) {
        let d = iq_f16(blk);
        for w in 0..32 {
            let v = u16::from_le_bytes([blk[2 + w * 2], blk[3 + w * 2]]);
            let sc = (blk[66 + w / 4] >> (4 * ((w / 2) % 2))) & 0x0F;
            let db = d * (0.5 + sc as f32) * 0.25;
            let row = (v & 511) as usize;
            let sb = ksign(u32::from(v >> 9));
            for k in 0..8 {
                out.push(db * iqt::IQ2_XS_GRID[row * 8 + k] as f32 * sgn(sb, k));
            }
        }
    }
    Ok(out)
}

/// IQ2_S: 256 / 82 — f16 d + 32 grid bytes + 32 sign bytes + 8 qh + 16 scale nibbles.
fn dequant_iq2_s(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256) && raw.len() >= count / 256 * 82);
    let (qs0, sg0, qh0, sc0) = (2usize, 34usize, 66usize, 74usize);
    let mut out = Vec::with_capacity(count);
    for blk in raw[..count / 256 * 82].chunks_exact(82) {
        let d = iq_f16(blk);
        for b8 in 0..32 {
            let sc = (blk[sc0 + b8 / 4] >> (4 * ((b8 / 2) % 2))) & 0x0F;
            let db = d * (0.5 + sc as f32) * 0.25;
            let row =
                blk[qs0 + b8] as usize | ((((blk[qh0 + b8 / 4] >> (2 * (b8 % 4))) & 3) as usize) << 8);
            let sb = blk[sg0 + b8];
            for k in 0..8 {
                out.push(db * iqt::IQ2_S_GRID[row * 8 + k] as f32 * sgn(sb, k));
            }
        }
    }
    Ok(out)
}

/// IQ3_XXS: 256 / 98 — f16 d + 64 grid bytes (4 elems each) + 8 u32 (signs+scale).
fn dequant_iq3_xxs(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256) && raw.len() >= count / 256 * 98);
    let mut out = Vec::with_capacity(count);
    for blk in raw[..count / 256 * 98].chunks_exact(98) {
        let d = iq_f16(blk);
        let w32 = |i: usize| {
            u32::from_le_bytes([blk[66 + i * 4], blk[67 + i * 4], blk[68 + i * 4], blk[69 + i * 4]])
        };
        for g in 0..8 {
            let w = w32(g);
            let db = d * (0.5 + (w >> 28) as f32) * 0.5;
            for j in 0..4 {
                let sb = ksign((w >> (7 * j)) & 0x7F);
                for t in 0..2 {
                    let row = blk[2 + g * 8 + j * 2 + t] as usize;
                    for k in 0..4 {
                        out.push(db * iqt::IQ3_XXS_GRID[row * 4 + k] as f32 * sgn(sb, t * 4 + k));
                    }
                }
            }
        }
    }
    Ok(out)
}

/// IQ3_S: 256 / 110 — f16 d + 64 grid bytes + 8 qh + 32 sign bytes + 8 scale nibbles.
fn dequant_iq3_s(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256) && raw.len() >= count / 256 * 110);
    let (qs0, qh0, sg0, sc0) = (2usize, 66usize, 74usize, 106usize);
    let mut out = Vec::with_capacity(count);
    for blk in raw[..count / 256 * 110].chunks_exact(110) {
        let d = iq_f16(blk);
        for i in 0..64 {
            let e0 = i * 4; // first of this row's 4 elements
            let sc = (blk[sc0 + e0 / 64] >> (4 * ((e0 / 32) % 2))) & 0x0F;
            let db = d * (1 + 2 * sc as i32) as f32;
            let row = blk[qs0 + i] as usize | ((((blk[qh0 + i / 8] >> (i % 8)) & 1) as usize) << 8);
            let sb = blk[sg0 + e0 / 8];
            for k in 0..4 {
                out.push(db * iqt::IQ3_S_GRID[row * 4 + k] as f32 * sgn(sb, (e0 + k) % 8));
            }
        }
    }
    Ok(out)
}

/// IQ1_S: 256 / 50 — f16 d + 32 grid bytes + 8 u16 (3-bit highs ×4, 3-bit scale, delta sign).
fn dequant_iq1_s(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256) && raw.len() >= count / 256 * 50);
    let mut out = Vec::with_capacity(count);
    for blk in raw[..count / 256 * 50].chunks_exact(50) {
        let d = iq_f16(blk);
        for g in 0..8 {
            let h = u16::from_le_bytes([blk[34 + g * 2], blk[35 + g * 2]]);
            let dl = d * (2 * ((h >> 12) & 7) + 1) as f32;
            let delta = if h & 0x8000 == 0 { 0.125f32 } else { -0.125 };
            for j in 0..4 {
                let row = blk[2 + g * 4 + j] as usize | ((((h >> (3 * j)) & 7) as usize) << 8);
                for k in 0..8 {
                    out.push(dl * (iqt::IQ1_GRID[row * 8 + k] as f32 + delta));
                }
            }
        }
    }
    Ok(out)
}

/// IQ1_M: 256 / 56 — 32 grid bytes + 16 qh nibble-bytes + 4 u16 scales (f16 d packed in
/// their top nibbles).
fn dequant_iq1_m(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256) && raw.len() >= count / 256 * 56);
    let mut out = Vec::with_capacity(count);
    for blk in raw[..count / 256 * 56].chunks_exact(56) {
        let sc16 = |i: usize| u16::from_le_bytes([blk[48 + i * 2], blk[49 + i * 2]]);
        let dbits = ((sc16(0) & 0xF000) >> 12)
            | ((sc16(1) & 0xF000) >> 8)
            | ((sc16(2) & 0xF000) >> 4)
            | (sc16(3) & 0xF000);
        let d = half::f16::from_bits(dbits).to_f32();
        for s in 0..16 {
            // scale s covers 16 elems = 2 grid rows; 3 bits at position s%4 of u16 s/4
            let sc = (sc16(s / 4) >> (3 * (s % 4))) & 0x07;
            let dl = d * (2 * sc + 1) as f32;
            for half_i in 0..2 {
                let b8 = s * 2 + half_i; // 8-elem group index (0..32)
                let nib = (blk[32 + b8 / 2] >> (4 * (b8 % 2))) & 0x0F;
                let row = blk[b8] as usize | (((nib & 7) as usize) << 8);
                let delta = if nib & 8 == 0 { 0.125f32 } else { -0.125 };
                for k in 0..8 {
                    out.push(dl * (iqt::IQ1_GRID[row * 8 + k] as f32 + delta));
                }
            }
        }
    }
    Ok(out)
}

/// IQ4_NL: 32 / 18 — f16 d + 16 nibble bytes through the 16-entry nonlinear LUT
/// (Q4_0's layout, a codebook instead of the linear ramp).
fn dequant_iq4_nl(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(32) && raw.len() >= count / 32 * 18);
    let mut out = Vec::with_capacity(count);
    for blk in raw[..count / 32 * 18].chunks_exact(18) {
        let d = iq_f16(blk);
        for l in 0..16 {
            out.push(d * iqt::IQ4_KVALUES[(blk[2 + l] & 0x0F) as usize] as f32);
        }
        for l in 0..16 {
            out.push(d * iqt::IQ4_KVALUES[(blk[2 + l] >> 4) as usize] as f32);
        }
    }
    Ok(out)
}

/// IQ4_XS: 256 / 136 — f16 d + u16 scales_h + 4 scale-nibble bytes + 128 nibble bytes,
/// 8 sub-blocks of 32 through the same LUT with 6-bit (−32-biased) sub-scales.
fn dequant_iq4_xs(raw: &[u8], count: usize) -> Result<Vec<f32>> {
    ensure!(count.is_multiple_of(256) && raw.len() >= count / 256 * 136);
    let mut out = Vec::with_capacity(count);
    for blk in raw[..count / 256 * 136].chunks_exact(136) {
        let d = iq_f16(blk);
        let sh = u16::from_le_bytes([blk[2], blk[3]]);
        for sb in 0..8 {
            let ls = ((blk[4 + sb / 2] >> (4 * (sb % 2))) & 0x0F) | ((((sh >> (2 * sb)) & 3) as u8) << 4);
            let dl = d * (ls as i8 as i32 - 32) as f32;
            let q = &blk[8 + sb * 16..8 + sb * 16 + 16];
            for l in 0..16 {
                out.push(dl * iqt::IQ4_KVALUES[(q[l] & 0x0F) as usize] as f32);
            }
            for l in 0..16 {
                out.push(dl * iqt::IQ4_KVALUES[(q[l] >> 4) as usize] as f32);
            }
        }
    }
    Ok(out)
}