fix(artifact-score): checkpoint resume

src/gpu/modal_score_artifact.py

@@ -0,0 +1,107 @@
+"""Modal app: score a SHIPPED IMF artifact end-to-end through the
+Python runtime — the protocol's 1400-byte windowing, greedy ONNX
+decode, haraqat projection, sadeedbench DER — so a published number
+can be re-derived from the exact bytes users download.
+
+    modal run --detach src/gpu/modal_score_artifact.py::score \\
+        --zip imf/ara-diac-small-2/ara-diac-small-2.0-fp32.zip \\
+        --expect-sha d9aa95d0... --out artifact-2.0-preds.jsonl
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import modal
+
+REPO_ROOT = Path(__file__).resolve().parent.parent.parent
+
+IMAGE = (
+    modal.Image.debian_slim(python_version="3.11")
+    .pip_install(
+        "onnxruntime==1.23.2",
+        "numpy>=1.26",
+        "pandas>=2.0",
+        "pyarrow>=14.0",
+        "pyarabic>=0.6",
+        "prettytable>=3.9",
+        "pyyaml>=6.0",
+    )
+    .add_local_dir(str(REPO_ROOT), "/root/interscript-ml", copy=True)
+    .add_local_dir(str(REPO_ROOT.parent / "rababa"), "/opt/rababa", copy=True)
+    .workdir("/root/interscript-ml")
+)
+
+MODELS_VOLUME = modal.Volume.from_name("secryst-models")
+
+app = modal.App("interscript-ml-artifact-score", image=IMAGE)
+
+
+@app.function(cpu=8, memory=32 * 1024, timeout=5 * 3600,
+              volumes={"/outputs": MODELS_VOLUME})
+def score(zip_path: str, expect_sha: str = "", out: str = "") -> dict:
+    import hashlib
+    import json
+    import sys
+
+    sys.path.insert(0, "/root/interscript-ml/runtime/src")
+    sys.path.insert(0, "/root/interscript-ml/src")
+
+    volume_zip = Path("/outputs") / zip_path
+    digest = hashlib.sha256(volume_zip.read_bytes()).hexdigest()
+    if expect_sha and digest != expect_sha:
+        raise RuntimeError(f"volume zip sha {digest} != expected {expect_sha}")
+
+    import pandas as pd
+
+    from harness.sadeed import project_haraqat, split_windows, strip_diacritics
+    from interscript_ml import Model
+    from sadeedbench.scoring import score_predictions
+
+    inputs = pd.read_parquet("/opt/rababa/data/sadeed-diac-25/train.parquet")[
+        "input"
+    ].tolist()
+    model = Model.load(volume_zip)
+
+    # resumable: worker evictions restart the input; completed rows are
+    # durably on the volume and skipped on relaunch
+    out_path = Path("/outputs") / out if out else None
+    done: dict[int, str] = {}
+    if out_path and out_path.exists():
+        for line in out_path.read_text(encoding="utf-8").splitlines():
+            if line.strip():
+                row = json.loads(line)
+                done[row["idx"]] = row["student"]
+        print(f"resume: {len(done)} rows already complete", flush=True)
+
+    preds: list[str] = []
+    for i, src in enumerate(inputs):
+        if i in done:
+            preds.append(done[i])
+            continue
+        stripped = strip_diacritics(src)
+        pieces = []
+        for w in split_windows(stripped, 1400):
+            out_text = model.translate(w, max_len=max(256, 2 * len(w)))
+            pieces.append(project_haraqat(out_text, w))
+        preds.append("".join(pieces))
+        if (i + 1) % 100 == 0:
+            print(f"{i+1}/{len(inputs)}", flush=True)
+            if out_path:
+                with out_path.open("w", encoding="utf-8") as fh:
+                    for j, q in enumerate(preds):
+                        if j < len(preds):
+                            fh.write(json.dumps({"idx": j, "student": q}, ensure_ascii=False) + "\n")
+                MODELS_VOLUME.commit()
+
+    gts = pd.read_parquet("/opt/rababa/data/sadeed-diac-25/train.parquet")[
+        "output"
+    ].tolist()
+    result = score_predictions(preds, gts)
+    result["sha256"] = digest
+    return result
+
+
+@app.local_entrypoint()
+def main(zip_path: str, expect_sha: str = "", out: str = "") -> None:
+    print(score.remote(zip_path, expect_sha, out))