memra-engine 0.83.0

// Focused FA-prefill sanitizer harness: exercises fa_prefill_f32 AND fa_prefill_q
// (the quant twin) with the SAME shapes/tolerances as kernel_check's FA section,
// WITHOUT the unrelated SSM/other kernels that abort the full kernel_check under
// compute-sanitizer. Run under: compute-sanitizer --tool memcheck|racecheck.
use memra_engine::Engine;
use memra_validate::maxdiff;

fn pr(i: usize) -> f32 {
    (((i.wrapping_mul(2654435761)) >> 8) & 0xffff) as f32 / 32768.0 - 1.0
}

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let e = Engine::new(0)?;
    let (hd, nh, nhkv) = (256usize, 16usize, 4usize);
    let scale = 1.0 / (hd as f32).sqrt();
    let cpu_sdpa = |q: &[f32], k: &[f32], v: &[f32], t: usize, tkv: usize| -> Vec<f32> {
        let mut o = vec![0f32; hd * nh * t];
        for head in 0..nh {
            let kvh = head / (nh / nhkv);
            for qt in 0..t {
                let q_pos = (tkv - t) + qt;
                let qv = &q[(qt * nh + head) * hd..][..hd];
                let mut sc = vec![0f32; tkv];
                for tk in 0..tkv {
                    let kv = &k[(tk * nhkv + kvh) * hd..][..hd];
                    let mut a = 0.0;
                    for d in 0..hd {
                        a += qv[d] * kv[d];
                    }
                    a *= scale;
                    if tk > q_pos {
                        a = -1e30;
                    }
                    sc[tk] = a;
                }
                let mx = sc.iter().cloned().fold(-1e30f32, f32::max);
                let mut sum = 0.0;
                for s in sc.iter_mut() {
                    *s = (*s - mx).exp();
                    sum += *s;
                }
                for s in sc.iter_mut() {
                    *s /= sum;
                }
                let ov = &mut o[(qt * nh + head) * hd..][..hd];
                for d in 0..hd {
                    let mut a = 0.0;
                    for tk in 0..tkv {
                        a += sc[tk] * v[(tk * nhkv + kvh) * hd + d];
                    }
                    ov[d] = a;
                }
            }
        }
        o
    };
    let mut fails = 0;
    // f32 prefill (fa_prefill_f32)
    for (t, tkv) in [(16usize, 16usize), (64, 64), (100, 100), (256, 256)] {
        let q: Vec<f32> = (0..hd * nh * t).map(|i| pr(i) * 0.2).collect();
        let k: Vec<f32> = (0..hd * nhkv * tkv).map(|i| pr(i + 7) * 0.2).collect();
        let v: Vec<f32> = (0..hd * nhkv * tkv).map(|i| pr(i + 11) * 0.2).collect();
        let cpu = cpu_sdpa(&q, &k, &v, t, tkv);
        let qd = e.htod(&q)?;
        let kd = e.htod(&k)?;
        let vd = e.htod(&v)?;
        let mut od = e.zeros(hd * nh * t)?;
        e.fa_prefill(&qd, &kd, &vd, &mut od, hd, nh, nhkv, t, tkv, scale, true)?;
        let g = e.dtoh(&od)?;
        let d = maxdiff(&cpu, &g);
        let sc = cpu.iter().map(|v| v.abs()).fold(0.0, f32::max).max(1e-3);
        let rel = d / sc;
        println!(
            "fa_prefill_f32 T={t} Tkv={tkv}: rel={rel:.2e} {}",
            if rel < 2e-2 {
                "OK"
            } else {
                fails += 1;
                "FAIL"
            }
        );
    }
    // quant prefill twin (fa_prefill_q via fa_prefill_view) — quantize K/V then run.
    let kv_dim_k = hd * nhkv;
    let kv_dim_v = hd * nhkv;
    let (kbb, vbb) = memra_engine::kv_blk_bytes(); // env-selected KV formats
    let k_tok_bytes = (kv_dim_k / 32) * kbb;
    let v_tok_bytes = (kv_dim_v / 32) * vbb;
    for (t, tkv) in [(16usize, 16usize), (64, 64), (100, 100), (256, 256)] {
        let q: Vec<f32> = (0..hd * nh * t).map(|i| pr(i) * 0.2).collect();
        let k: Vec<f32> = (0..hd * nhkv * tkv).map(|i| pr(i + 7) * 0.2).collect();
        let v: Vec<f32> = (0..hd * nhkv * tkv).map(|i| pr(i + 11) * 0.2).collect();
        let cpu = cpu_sdpa(&q, &k, &v, t, tkv);
        let qd = e.htod(&q)?;
        let kd = e.htod(&k)?;
        let vd = e.htod(&v)?;
        let mut kc = e.alloc_u8(tkv * k_tok_bytes)?;
        let mut vc = e.alloc_u8(tkv * v_tok_bytes)?;
        for tok in 0..tkv {
            let k_row = kd.slice(tok * kv_dim_k..(tok + 1) * kv_dim_k);
            let v_row = vd.slice(tok * kv_dim_v..(tok + 1) * kv_dim_v);
            e.append_kv_quantized_view(
                &k_row,
                &v_row,
                &mut kc,
                &mut vc,
                tok,
                kv_dim_k,
                kv_dim_v,
                k_tok_bytes,
                v_tok_bytes,
                false,
            )?;
        }
        let kview = e.view_u8(&kc, tkv * k_tok_bytes);
        let vview = e.view_u8(&vc, tkv * v_tok_bytes);
        let mut od = e.zeros(hd * nh * t)?;
        e.fa_prefill_view(
            &qd,
            &kview,
            &vview,
            &mut od,
            hd,
            nh,
            nhkv,
            t,
            tkv,
            scale,
            true,
            k_tok_bytes,
            v_tok_bytes,
            false,
        )?;
        let g = e.dtoh(&od)?;
        let d = maxdiff(&cpu, &g);
        let sc = cpu.iter().map(|v| v.abs()).fold(0.0, f32::max).max(1e-3);
        let rel = d / sc;
        // quant twin: looser (q8_0 K / q5_1 V) — gate at q5_1 noise floor
        println!(
            "fa_prefill_q   T={t} Tkv={tkv}: rel={rel:.2e} {}",
            if rel < 6e-2 {
                "OK"
            } else {
                fails += 1;
                "FAIL"
            }
        );
    }
    if fails == 0 {
        println!("FA-SANITIZE GREEN");
    } else {
        println!("FA-SANITIZE FAIL ({fails})");
        std::process::exit(1);
    }
    Ok(())
}