inferencelayer 0.2.9

Kortexya's engine-native inference layer — LLM generation + embedding/encoder family on wgpu (WGSL kernels, any adapter) with a pure-Rust CPU fallback
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178

//! `qwen35-cpu` — proves the pure-CPU Qwen3.5 decoder ([`inferencelayer::reference_qwen35`]) matches
//! the transformers gold for NuExtract-3, with NO wgpu/Vulkan involved.
//!
//! ```text
//! qwen35-cpu --model <dir> --gold <gold.json>
//! ```
//!
//! The gold (scratchpad/qwen35_gold.json + _last_logits.f32) is produced by gen_qwen35_gold.py
//! (transformers, float32/eager/CPU). Checks: prefill next-token argmax, top-8 overlap, full
//! last-position logit cosine + max-abs diff, per-decoder-layer last-hidden norm, and end-to-end
//! greedy continuation parity.

use anyhow::{Context, Result};
use inferencelayer::reference_qwen35::{Qwen35Ref, argmax};
use std::path::PathBuf;

fn read_f32_le(path: &str) -> Result<Vec<f32>> {
    let bytes = std::fs::read(path).with_context(|| format!("reading {path}"))?;
    Ok(bytes
        .chunks_exact(4)
        .map(|c| f32::from_le_bytes([c[0], c[1], c[2], c[3]]))
        .collect())
}

fn topk(v: &[f32], k: usize) -> Vec<usize> {
    let mut idx: Vec<usize> = (0..v.len()).collect();
    idx.sort_unstable_by(|&a, &b| v[b].partial_cmp(&v[a]).unwrap());
    idx.truncate(k);
    idx
}

fn main() -> Result<()> {
    let scratch = "/private/tmp/claude-501/-Users-dlo-code-crawl-bee/\
6a3b906d-8642-4873-90fc-2aab8224d53c/scratchpad";
    let mut model = PathBuf::from(format!("{scratch}/models/nuextract-3"));
    let mut gold = PathBuf::from(format!("{scratch}/qwen35_gold.json"));
    let mut args = std::env::args().skip(1);
    while let Some(a) = args.next() {
        match a.as_str() {
            "--model" => model = args.next().context("--model needs a path")?.into(),
            "--gold" => gold = args.next().context("--gold needs a path")?.into(),
            other => anyhow::bail!("unknown arg {other}"),
        }
    }

    let g: serde_json::Value = serde_json::from_slice(&std::fs::read(&gold)?)?;
    let prompt_ids: Vec<u32> = g["prompt_ids"]
        .as_array()
        .context("prompt_ids")?
        .iter()
        .map(|x| x.as_u64().unwrap() as u32)
        .collect();
    let gold_argmax = g["prefill_argmax"].as_u64().unwrap() as u32;
    let gold_cont: Vec<u32> = g["continuation"]
        .as_array()
        .context("continuation")?
        .iter()
        .map(|x| x.as_u64().unwrap() as u32)
        .collect();
    let gold_logits = read_f32_le(g["last_logits_bin"].as_str().context("last_logits_bin")?)?;
    let gold_layers = g["layer_hidden"].as_array().context("layer_hidden")?;

    eprintln!("loading CPU reference from {} ...", model.display());
    let t0 = std::time::Instant::now();
    let mut m = Qwen35Ref::load(&model)?;
    eprintln!(
        "loaded {} layers ({} full-attn) in {:.1}s",
        m.cfg.n_layers,
        m.cfg.layer_is_full.iter().filter(|&&x| x).count(),
        t0.elapsed().as_secs_f32()
    );

    // ---- Prefill: replay prompt tokens, capture the last-token layer trace + logits. ----
    let t1 = std::time::Instant::now();
    let mut last_logits = Vec::new();
    for (pos, &tok) in prompt_ids.iter().enumerate() {
        m.capture_trace = pos == prompt_ids.len() - 1;
        last_logits = m.forward(tok, pos);
    }
    let prefill_dt = t1.elapsed().as_secs_f32();

    // Per-layer last-hidden norm vs gold (localizes any divergence to a layer).
    println!("\n== per-layer last-token hidden (CPU vs gold) ==");
    let mut worst_layer = (0usize, 0f32);
    for (li, gl) in gold_layers.iter().enumerate() {
        let gnorm = gl["norm"].as_f64().unwrap() as f32;
        let h = &m.trace[li];
        let cnorm = h.iter().map(|v| v * v).sum::<f32>().sqrt();
        let rel = (cnorm - gnorm).abs() / gnorm.max(1e-6);
        if rel > worst_layer.1 {
            worst_layer = (li, rel);
        }
        if li < 2 || li >= gold_layers.len() - 2 || rel > 0.02 {
            println!(
                "  layer {li:2}: cpu_norm={cnorm:9.4} gold_norm={gnorm:9.4} rel={:.2e}",
                rel
            );
        }
    }
    println!(
        "  worst layer: {} (rel norm diff {:.2e})",
        worst_layer.0, worst_layer.1
    );

    // Full logit-vector comparison.
    let cpu_argmax = argmax(&last_logits);
    let dot: f64 = last_logits
        .iter()
        .zip(&gold_logits)
        .map(|(&a, &b)| a as f64 * b as f64)
        .sum();
    let na: f64 = last_logits.iter().map(|&a| (a as f64).powi(2)).sum::<f64>().sqrt();
    let nb: f64 = gold_logits.iter().map(|&a| (a as f64).powi(2)).sum::<f64>().sqrt();
    let cos = dot / (na * nb);
    let maxdiff = last_logits
        .iter()
        .zip(&gold_logits)
        .map(|(&a, &b)| (a - b).abs())
        .fold(0f32, f32::max);

    let cpu_top = topk(&last_logits, 8);
    let gold_top = topk(&gold_logits, 8);
    let overlap = cpu_top.iter().filter(|t| gold_top.contains(t)).count();

    println!("\n== prefill next-token ==");
    println!("  cpu argmax  = {cpu_argmax}");
    println!("  gold argmax = {gold_argmax}");
    println!("  logit cosine = {cos:.6}   max|Δlogit| = {maxdiff:.4}");
    println!("  top-8 overlap = {overlap}/8");
    println!("  cpu  top8 = {cpu_top:?}");
    println!("  gold top8 = {gold_top:?}");
    println!(
        "  prefill: {} tokens in {:.1}s ({:.2}s/tok)",
        prompt_ids.len(),
        prefill_dt,
        prefill_dt / prompt_ids.len() as f32
    );

    // ---- Greedy generation, compare to gold continuation. ----
    println!("\n== greedy continuation ==");
    let mut produced = Vec::new();
    let mut tok = cpu_argmax;
    let mut pos = prompt_ids.len();
    m.capture_trace = false;
    let t2 = std::time::Instant::now();
    for _ in 0..gold_cont.len() {
        produced.push(tok);
        let logits = m.forward(tok, pos);
        tok = argmax(&logits);
        pos += 1;
    }
    let gen_dt = t2.elapsed().as_secs_f32();
    let match_len = produced
        .iter()
        .zip(&gold_cont)
        .take_while(|(a, b)| a == b)
        .count();
    println!("  cpu  = {produced:?}");
    println!("  gold = {gold_cont:?}");
    println!(
        "  matching prefix = {match_len}/{} ({:.2}s/tok)",
        gold_cont.len(),
        gen_dt / gold_cont.len() as f32
    );

    let argmax_ok = cpu_argmax == gold_argmax;
    let cont_ok = match_len == gold_cont.len();
    println!("\n== VERDICT ==");
    println!("  argmax match : {}", if argmax_ok { "PASS" } else { "FAIL" });
    println!("  cosine>0.999 : {}", if cos > 0.999 { "PASS" } else { "FAIL" });
    println!("  continuation : {}", if cont_ok { "PASS" } else { "FAIL" });
    if argmax_ok && cont_ok && cos > 0.999 {
        println!("\nALL CHECKS PASS — pure-CPU Qwen3.5 decode matches transformers gold.");
        Ok(())
    } else {
        anyhow::bail!("parity check failed")
    }
}