scripts/eval_span.py

#!/usr/bin/env python3
"""Evaluate one span-extraction arm on the frozen val split.

Identical code path for a zero-shot base model and for a fine-tuned checkpoint.
Emits a JSON with the full threshold sweep so the bakeoff never has to guess a
threshold, plus the selected operating point chosen by the Golden Rule
(highest exact_match among thresholds meeting over-deletion < 1.0%).

Usage:
  venv/bin/python eval_span.py --model <hf-id|ckpt-dir> --tag <name> --out <json>
                               [--cpu-latency] [--export-onnx]
"""

import argparse
import json
import os
import sys
import time

import numpy as np
import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

sys.path.insert(0, "/opt/vox/sandbox/scripts")
from span_common import (  # noqa: E402
    LABELS, MAX_LEN, decode_bio, load_jsonl, split_words, sweep_thresholds,
)

THRESHOLDS = [0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9,
             0.95, 0.97, 0.99, 0.995]


@torch.no_grad()
def predict(rows, tok, model, device, max_len=MAX_LEN, batch_size=32):
    """Returns {threshold: [char spans per row]} in one forward pass per row."""
    probs = []
    order = sorted(range(len(rows)), key=lambda i: len(rows[i]["raw_text"]))
    out = [None] * len(rows)
    for b in range(0, len(order), batch_size):
        chunk = order[b:b + batch_size]
        encs, wids_all = [], []
        for i in chunk:
            words, _, _ = split_words(rows[i]["raw_text"])
            e = tok(words, is_split_into_words=True, truncation=True, max_length=max_len)
            encs.append(e)
            wids_all.append(e.word_ids())
        pad = tok.pad(encs, return_tensors="pt")
        pad = {k: v.to(device) for k, v in pad.items()}
        logits = model(**pad).logits.float()
        p = torch.softmax(logits, dim=-1).cpu().numpy()
        for j, i in enumerate(chunk):
            wids = wids_all[j]
            if wids is None:
                out[i] = ([], [])
                continue
            p_inspan = np.zeros(len(wids))
            p_begin = np.zeros(len(wids))
            for t, w in enumerate(wids):
                if w is None or t >= p.shape[1]:
                    continue
                pr = p[j][t]
                p_inspan[w] = max(0.0, 1.0 - pr[0])
                tot = pr[1] + pr[2]
                p_begin[w] = (pr[1] / tot) if tot > 1e-9 else 0.5
            out[i] = (p_inspan, p_begin)
    for t in THRESHOLDS:
        preds = [decode_bio(rows[i]["raw_text"], out[i][0], out[i][1], t) for i in range(len(rows))]
        probs.append((t, preds))
    return dict(probs)


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--model", required=True)
    ap.add_argument("--tag", required=True)
    ap.add_argument("--out", required=True)
    ap.add_argument("--val", default="/opt/vox/sandbox/corpus/pilot_v2_val.jsonl")
    ap.add_argument("--max-len", type=int, default=MAX_LEN)
    ap.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
    ap.add_argument("--cpu-latency", action="store_true")
    ap.add_argument("--export-onnx", action="store_true")
    ap.add_argument("--onnx-int8", action="store_true")
    ap.add_argument("--latency-threads", type=int, default=16)
    args = ap.parse_args()

    rows = load_jsonl(args.val)
    tok = AutoTokenizer.from_pretrained(args.model)
    model = AutoModelForTokenClassification.from_pretrained(args.model, dtype=torch.float32).to(args.device).eval()

    t0 = time.time()
    preds_by_thr = predict(rows, tok, model, args.device, args.max_len)
    gpu_s = time.time() - t0

    res = sweep_thresholds(rows, preds_by_thr, THRESHOLDS)
    n_par = sum(p.numel() for p in model.parameters()) / 1e6
    out = {
        "tag": args.tag, "model": args.model, "params_m": round(n_par, 2),
        "n_val": len(rows), "eval_device": args.device, "eval_wall_s": round(gpu_s, 2),
        "best": res["best"], "sweep": res["sweep"],
    }

    # ONNX is a Layer 5 shipping concern, not an accuracy gate (user-directed
    # 2026-10-03). It must never be able to kill an accuracy measurement.
    if args.export_onnx:
        try:
            out["onnx"] = export_and_check(model, tok, rows, args)
        except Exception as e:
            out["onnx"] = {"error": f"{type(e).__name__}: {str(e)[:200]}"}
            print(f"[{args.tag}] onnx export FAILED (non-fatal): {type(e).__name__}: {str(e)[:120]}", flush=True)

    if args.cpu_latency:
        try:
            out["cpu_latency"] = measure_cpu(model, tok, rows, args)
        except Exception as e:
            out["cpu_latency"] = {"error": f"{type(e).__name__}: {str(e)[:200]}"}
            print(f"[{args.tag}] cpu latency FAILED (non-fatal): {type(e).__name__}", flush=True)

    os.makedirs(os.path.dirname(args.out), exist_ok=True)
    with open(args.out, "w") as f:
        json.dump(out, f, indent=2)
    b = out["best"]
    print(f"[{args.tag}] P={b['span_word']['precision']} R={b['span_word']['recall']} "
          f"EM={b['exact_match']} overdel={b['over_deletion_rate']}% "
          f"harmful={b['harmful_span_rate']}% thr={b['threshold']} passes={b['passes_gate']}", flush=True)


def export_and_check(model, tok, rows, args):
    """ONNX export -> optional int8 dynamic quant -> char-level parity vs PyTorch.

    Parity is the gate, not F1: a silent divergence here becomes a word
    corruption bug in the Rust runtime.
    """
    import onnx
    import onnxruntime as ort
    from onnxruntime.quantization import QuantType, quantize_dynamic

    res = {}
    onnx_dir = os.path.join(os.path.dirname(args.out), f"onnx_{args.tag}")
    os.makedirs(onnx_dir, exist_ok=True)
    fp32_path = os.path.join(onnx_dir, "model.onnx")

    # Export must run on CPU: the live model may be on CUDA while the tracer's
    # dummy inputs are not, which silently produced a device-mismatch RuntimeError
    # for all six arms until this was found.
    was_device = next(model.parameters()).device
    model = model.cpu()
    dummy = tok(["We need to upgrade the database."], return_tensors="pt", padding=True)
    torch.onnx.export(
        model, (dummy["input_ids"], dummy["attention_mask"]), fp32_path,
        input_names=["input_ids", "attention_mask"], output_names=["logits"],
        dynamic_axes={k: {0: "b", 1: "s"} for k in ("input_ids", "attention_mask")} | {"logits": {0: "b", 1: "s"}},
        opset_version=17, do_constant_folding=True, dynamo=False,
    )
    onnx.checker.check_model(fp32_path)
    res["fp32_bytes"] = os.path.getsize(fp32_path)
    res["opset"] = 17

    cand = fp32_path
    if args.onnx_int8:
        int8_path = os.path.join(onnx_dir, "model_int8.onnx")
        quantize_dynamic(fp32_path, int8_path, weight_type=QuantType.QInt8)
        onnx.checker.check_model(int8_path)
        res["int8_bytes"] = os.path.getsize(int8_path)
        res["int8_compression"] = round(res["fp32_bytes"] / res["int8_bytes"], 2)
        cand = int8_path
    res["checked_path"] = cand

    # char-level parity: does the exported model decode to the same spans?
    sess = ort.InferenceSession(cand, providers=["CPUExecutionProvider"])
    sub = rows[:200]
    mism = 0
    for r in sub:
        words, _, _ = split_words(r["raw_text"])
        e = tok(words, is_split_into_words=True, truncation=True, max_length=args.max_len, return_tensors="np")
        ort_out = sess.run(None, {"input_ids": e["input_ids"], "attention_mask": e["attention_mask"]})[0]
        probs = torch.softmax(torch.tensor(ort_out), dim=-1).numpy()[0]
        wids = tok(words, is_split_into_words=True, truncation=True, max_length=args.max_len).word_ids()
        pi = np.zeros(len(wids)); pb = np.zeros(len(wids))
        for t, w in enumerate(wids):
            if w is None or t >= probs.shape[0]:
                continue
            pi[w] = max(0.0, 1.0 - probs[t][0])
            tot = probs[t][1] + probs[t][2]
            pb[w] = (probs[t][1] / tot) if tot > 1e-9 else 0.5
        ort_spans = decode_bio(r["raw_text"], pi, pb, 0.5)
        with torch.no_grad():
            pt = model(input_ids=e["input_ids"].to(args.device),
                       attention_mask=e["attention_mask"].to(args.device)).logits.float().cpu().numpy()[0]
        pp = torch.softmax(torch.tensor(pt), dim=-1).numpy()
        pi2 = np.zeros(len(wids)); pb2 = np.zeros(len(wids))
        for t, w in enumerate(wids):
            if w is None or t >= pp.shape[0]:
                continue
            pi2[w] = max(0.0, 1.0 - pp[t][0])
            tot = pp[t][1] + pp[t][2]
            pb2[w] = (pp[t][1] / tot) if tot > 1e-9 else 0.5
        if decode_bio(r["raw_text"], pi2, pb2, 0.5) != ort_spans:
            mism += 1
    res["parity_rows"] = len(sub)
    res["parity_mismatch"] = mism
    res["parity_rate"] = round(100.0 * (len(sub) - mism) / len(sub), 2)
    res["parity_pass"] = res["parity_rate"] >= 99.0
    model = model.to(was_device)
    return res


def measure_cpu(model, tok, rows, args):
    """CPU p50/p95 at production thread count. Invariant 9: client HW is >=2x slower."""
    import shutil
    m = AutoModelForTokenClassification.from_pretrained(args.model, dtype=torch.float32)
    t = AutoTokenizer.from_pretrained(args.model)
    torch.set_num_threads(args.latency_threads)
    try:
        torch.set_num_interop_threads(1)
    except RuntimeError:
        pass
    m.eval()
    texts = [r["raw_text"] for r in rows[:200]]
    with torch.no_grad():
        for s in texts[:20]:
            w, _, _ = split_words(s)
            e = t(w, is_split_into_words=True, truncation=True, max_length=args.max_len,
                  return_tensors="pt", padding=True)
            m(input_ids=e["input_ids"], attention_mask=e["attention_mask"])
        lat = []
        for s in texts:
            w, _, _ = split_words(s)
            t0 = time.perf_counter()
            e = t(w, is_split_into_words=True, truncation=True, max_length=args.max_len,
                  return_tensors="pt", padding=True)
            m(input_ids=e["input_ids"], attention_mask=e["attention_mask"])
            lat.append((time.perf_counter() - t0) * 1000.0)
    lat = np.array(lat)
    return {
        "threads": args.latency_threads, "n": len(lat),
        "p50_ms": round(float(np.percentile(lat, 50)), 2),
        "p95_ms": round(float(np.percentile(lat, 95)), 2),
        "mean_ms": round(float(lat.mean()), 2),
        "client_2x_p95_ms": round(float(np.percentile(lat, 95)) * 2, 2),
        "note": "2x multiplier per AGENTS.md invariant 9",
    }


if __name__ == "__main__":
    main()