eval_job.py

"""In-job entrypoint for the CADGenBench eval on HF Jobs.

Invoked by the leaderboard Space's worker (see
``AI4Engineering/submit.py``) via::

    hf jobs run --image hf.co/spaces/HuggingAI4Engineering/cadgenbench-eval-gpu \\
        --flavor a10g-large \\
        --env CADGENBENCH_DATA_REPO=HuggingAI4Engineering/cadgenbench-data \\
        --env CADGENBENCH_DATA_GT_REPO=HuggingAI4Engineering/cadgenbench-data-gt \\
        --env HF_SUBMISSIONS_REPO=HuggingAI4Engineering/cadgenbench-submissions \\
        --env EVAL_WORKER_COUNT=8 \\
        --secrets HF_TOKEN \\
        python /opt/eval_job.py <submission_id> <zip_url>

Two run modes:

**Whole-submission (default, no ``--fixtures``)** -- the original path.
Synchronous, no fallbacks. Any failure raises and the container exits
non-zero; the Space's poller catches the ERROR stage and flips the
submission row to ``failed``.

1. Download ``submissions/<id>.zip`` from the submissions dataset
   via ``hf_hub_download`` (auth via ``HF_TOKEN``).
2. Unpack into ``/tmp/run/``.
3. ``cadgenbench evaluate /tmp/run --workers <n>`` (subprocess).
4. ``cadgenbench report single /tmp/run -o /tmp/<id>.html``
   (subprocess).
5. Build ``report.json`` bundling ``run_summary.json`` + every
   per-fixture ``result.json`` (mirror of submit.py's
   ``_build_report_json``).
6. Upload ``reports/<id>.html`` + ``reports/<id>.json`` back to the
   submissions dataset via ``HfApi.upload_file``.
7. Exit 0.

The Space-side worker then downloads ``reports/<id>.json``, reads
``run_summary`` out of it, and flips the row to ``completed``.

**Shard (``--fixtures f1,f2,... --shard-id shard_000``)** -- used by
the Space's sharded submit path (UC3) to fan a large submission across
several jobs. Steps 1-2 are identical, then the run dir is pruned to
just this shard's fixtures, ``cadgenbench evaluate`` runs over that
subset, and the resulting per-fixture dirs (``result.json`` + renders)
are staged *verbatim*. If ``CADGENBENCH_SHARD_BUCKET`` is set, the shard
syncs them into that HF Storage Bucket via the bucket API; otherwise it
uploads under ``reports/<id>/shards/<shard_id>/`` in the submissions
dataset. No
report HTML, ``report.json``, or gallery render is produced per shard:
the Space reads every shard's fixture dirs, merges them into one run dir,
and builds the single ``run_summary`` + report + gallery from the merged
whole (mirroring the orchestrator's ``_merge_eval``). Exit 0 on success;
any failure exits non-zero and the Space marks that shard ERROR and
retries it.
"""
from __future__ import annotations

import argparse
import json
import os
import shutil
import subprocess
import sys
import zipfile
from pathlib import Path
from typing import Any

from huggingface_hub import CommitOperationAdd, HfApi, hf_hub_download


RUN_DIR = Path("/tmp/run")
REPORT_HTML_DIR = Path("/tmp")

EVAL_TIMEOUT_SECONDS = 30 * 60
REPORT_TIMEOUT_SECONDS = 5 * 60

REPORTS_DIR_IN_REPO = "reports"
RENDERS_DIR_IN_REPO = "renders"
SHARD_BUCKET_ENV = "CADGENBENCH_SHARD_BUCKET"
SHARD_BUCKET_PREFIX_ENV = "CADGENBENCH_SHARD_BUCKET_PREFIX"

# Public HF Storage Bucket the eval job uploads gallery/report renders to (the
# job is the sole render uploader; the Space never handles render bytes). The
# hosted report + gallery reference these by anonymous bucket URL. Submission
# renders only; GT renders stay in the private GT dataset.
RENDER_BUCKET = os.environ.get(
    "CADGENBENCH_RENDER_BUCKET", "HuggingAI4Engineering/cadgenbench-eval-staging",
).strip()
HF_ENDPOINT = os.environ.get("HF_ENDPOINT", "https://huggingface.co").rstrip("/")

# Space-relative proxy roots the hosted report references for the *private*
# assets it can't link from the public bucket. The report is served by the
# Space at ``/reports/<id>.html``, so these absolute-path URLs resolve against
# the Space origin and the token-holding proxy streams the bytes. Kept in sync
# with the routes registered in the leaderboard Space's ``app.py``.
GT_PROXY_BASE_URL = "/gt"
INPUT_PROXY_BASE_URL = "/task-input"


def _render_base_url(submission_id: str) -> str:
    """Public ``.../resolve/renders/<id>`` base; report appends ``/<fixture>/<file>``."""
    return f"{HF_ENDPOINT}/buckets/{RENDER_BUCKET}/resolve/{RENDERS_DIR_IN_REPO}/{submission_id}"


def _submission_zip_url(submission_id: str, submissions_repo: str) -> str:
    """Hub resolve URL of ``submissions/<id>.zip`` (the report's download link).

    Same canonical blob URL the submit handler records as
    ``submission_blob_url`` and the gallery links, so the report's download
    button points at the identical artifact.
    """
    return (
        f"{HF_ENDPOINT}/datasets/{submissions_repo}"
        f"/resolve/main/submissions/{submission_id}.zip"
    )


def _upload_renders_to_bucket(
    run_dir: Path, submission_id: str, token: str,
) -> list[str]:
    """Upload every fixture's renders to ``renders/<id>/<fixture>/`` in the bucket.

    One ``batch_bucket_files`` call for the whole submission (cheaper than a
    per-file fan-out). Returns the bucket object paths that were uploaded (so
    the caller can warm the CDN for them).
    """
    add: list[tuple[str, str]] = []
    for fixture_dir in sorted(d for d in run_dir.iterdir() if d.is_dir()):
        dest_prefix = (
            f"{RENDERS_DIR_IN_REPO}/{submission_id}/{fixture_dir.name}"
        )
        renders_dir = fixture_dir / "renders"
        if renders_dir.is_dir():
            for render_path in sorted(renders_dir.iterdir()):
                if render_path.suffix.lower() not in {".png", ".webp"}:
                    continue
                add.append((str(render_path), f"{dest_prefix}/{render_path.name}"))
        # The interface overlay is a per-fixture report artifact that lives at
        # the fixture-dir root (not under renders/). It must ride to the bucket
        # alongside the turntables so the hosted report can reference it by URL
        # instead of base64-inlining it; without this the report would have a
        # broken overlay link. Uploaded under the same per-fixture prefix.
        overlay = fixture_dir / "interface_overlay.png"
        if overlay.is_file():
            add.append((str(overlay), f"{dest_prefix}/{overlay.name}"))
    if not add:
        return []
    HfApi(token=token).batch_bucket_files(RENDER_BUCKET, add=add, token=token)
    print(
        f"[eval_job] uploaded {len(add)} render(s) -> "
        f"hf://buckets/{RENDER_BUCKET}/{RENDERS_DIR_IN_REPO}/{submission_id}",
        flush=True,
    )
    return [dest for _, dest in add]


def _warm_render_cdn(object_paths: list[str]) -> None:
    """Prime the CDN by fetching each freshly-uploaded render once.

    A bucket serves a render via a 302 to a signed Xet CDN URL, and the very
    first fetch of a brand-new object pays the chunk-reconstruction cost, which
    is the lag a viewer sees opening a just-published report. Fetching each
    object here (in parallel, anonymously, best-effort) warms the edge cache so
    the first human hits a warm object instead. Failures are swallowed: warming
    is an optimisation, never a publish blocker.
    """
    import urllib.request
    from concurrent.futures import ThreadPoolExecutor

    def _warm(path: str) -> None:
        url = f"{HF_ENDPOINT}/buckets/{RENDER_BUCKET}/resolve/{path}"
        try:
            with urllib.request.urlopen(url, timeout=30) as resp:
                resp.read()
        except Exception:
            pass

    if not object_paths:
        return
    with ThreadPoolExecutor(max_workers=16) as pool:
        list(pool.map(_warm, object_paths))
    print(f"[eval_job] warmed CDN for {len(object_paths)} render(s)", flush=True)

# Sub-prefix under ``reports/<id>/`` where each shard uploads its raw
# per-fixture dirs in shard mode. The Space merges these and deletes the
# whole ``shards/`` tree after a successful merge.
SHARDS_DIR_NAME = "shards"

def main() -> int:
    parser = argparse.ArgumentParser(
        description="Run the CADGenBench eval pipeline on an HF Job.",
    )
    parser.add_argument(
        "submission_id",
        help="Filesystem-safe slug minted by the Space's submit handler.",
    )
    parser.add_argument(
        "zip_url",
        help=(
            "Canonical Hub blob URL of submissions/<id>.zip "
            "(submission_blob_url from the row)."
        ),
    )
    parser.add_argument(
        "--fixtures",
        default=None,
        help=(
            "Comma-separated fixture subset for shard mode. When set, the "
            "run dir is pruned to just these fixtures, evaluated, and the "
            "per-fixture dirs are uploaded under "
            "reports/<id>/shards/<shard-id>/ for the Space to merge. "
            "Omit for the original whole-submission path."
        ),
    )
    parser.add_argument(
        "--shard-id",
        default=None,
        help=(
            "Shard label (e.g. shard_000) naming this shard's upload prefix. "
            "Required when --fixtures is set."
        ),
    )
    args = parser.parse_args()

    submission_id: str = args.submission_id
    zip_url: str = args.zip_url

    token = _require_env("HF_TOKEN")
    submissions_repo = _require_env("HF_SUBMISSIONS_REPO")
    worker_count = int(os.environ.get("EVAL_WORKER_COUNT", "8"))

    shard_fixtures = _parse_fixtures_arg(args.fixtures)
    if shard_fixtures is not None:
        if not args.shard_id:
            raise RuntimeError("--shard-id is required when --fixtures is set.")
        print(
            f"[eval_job] submission_id={submission_id} shard={args.shard_id} "
            f"fixtures={len(shard_fixtures)} workers={worker_count} "
            f"repo={submissions_repo}",
            flush=True,
        )
        _prepare_run_dir(submission_id, zip_url, submissions_repo, token)
        _prune_run_dir(RUN_DIR, shard_fixtures)
        _run_eval(RUN_DIR, worker_count)
        # The shard job is the sole uploader of its fixtures' renders to the
        # permanent bucket prefix; the Space merge only assembles the report.
        _warm_render_cdn(_upload_renders_to_bucket(RUN_DIR, submission_id, token))
        _upload_shard_artifacts(
            submission_id, args.shard_id, RUN_DIR, submissions_repo, token,
        )
        print(
            f"[eval_job] done: {submission_id} shard={args.shard_id}",
            flush=True,
        )
        return 0

    print(
        f"[eval_job] submission_id={submission_id} "
        f"workers={worker_count} repo={submissions_repo}",
        flush=True,
    )

    _prepare_run_dir(submission_id, zip_url, submissions_repo, token)
    _run_eval(RUN_DIR, worker_count)
    # Upload renders to the public bucket and warm the CDN, then build the
    # report referencing them by URL (so the heavy WebP/PNG bytes never land in
    # the HTML and the first viewer hits an already-warm edge cache).
    _warm_render_cdn(_upload_renders_to_bucket(RUN_DIR, submission_id, token))
    html_path = REPORT_HTML_DIR / f"{submission_id}.html"
    _run_report(
        RUN_DIR, html_path,
        render_base_url=_render_base_url(submission_id),
        download_url=_submission_zip_url(submission_id, submissions_repo),
    )
    report_json = _build_report_json(RUN_DIR)
    _publish_reports_and_gallery(
        submission_id, html_path, report_json, submissions_repo, token,
    )
    print(f"[eval_job] done: {submission_id}", flush=True)
    return 0


def _parse_fixtures_arg(raw: str | None) -> list[str] | None:
    """Parse the ``--fixtures`` CSV into a deduped list, or ``None``.

    ``None`` (flag absent) selects the whole-submission path. A present
    but empty/whitespace value is a usage error: a shard with no
    fixtures is never something the Space should dispatch.
    """
    if raw is None:
        return None
    names: list[str] = []
    seen: set[str] = set()
    for part in raw.split(","):
        name = part.strip()
        if not name or name in seen:
            continue
        seen.add(name)
        names.append(name)
    if not names:
        raise RuntimeError("--fixtures was set but resolved to no fixture names.")
    return names


def _require_env(name: str) -> str:
    """Return env var *name* or raise with a clear message."""
    value = os.environ.get(name)
    if not value:
        raise RuntimeError(
            f"Required environment variable {name!r} is unset or empty."
        )
    return value


def _prepare_run_dir(
    submission_id: str,
    zip_url: str,
    submissions_repo: str,
    token: str,
) -> None:
    """Download the submission zip and unpack into ``RUN_DIR``.

    Derives the in-repo path from *zip_url* and pulls via
    ``hf_hub_download`` so token auth is handled and the file lands
    in the Hub cache. *zip_url* is expected to look like
    ``https://huggingface.co/datasets/<repo>/resolve/main/submissions/<id>.zip``;
    we accept any URL shape that ends in ``submissions/<id>.zip`` and
    re-derive the in-repo filename from the *submission_id*.
    """
    if RUN_DIR.exists():
        shutil.rmtree(RUN_DIR)
    RUN_DIR.mkdir(parents=True)

    in_repo_path = f"submissions/{submission_id}.zip"
    print(
        f"[eval_job] downloading {submissions_repo}:{in_repo_path}",
        flush=True,
    )
    local_zip = hf_hub_download(
        repo_id=submissions_repo,
        filename=in_repo_path,
        repo_type="dataset",
        token=token,
    )

    # Defensive: matches the validated shape from submit.py's
    # _extract_zip, but the Space already gate-checked the zip
    # contents pre-upload so we extract directly without re-
    # validating zip-slip / symlinks here.
    with zipfile.ZipFile(local_zip) as zf:
        zf.extractall(RUN_DIR)
    print(f"[eval_job] unpacked into {RUN_DIR}", flush=True)


def _prune_run_dir(run_dir: Path, fixtures: list[str]) -> None:
    """Drop every fixture dir under *run_dir* not in *fixtures*.

    Shard mode unpacks the whole zip (the candidate STEPs for every
    fixture) but should only evaluate this shard's slice, so we delete
    the other fixture dirs before ``cadgenbench evaluate`` walks the
    tree. Non-fixture files at the root (e.g. ``meta.json``) are left
    untouched. Raises if a requested fixture is absent from the zip,
    which would mean the Space sharded a name the submission didn't
    contain (a contract violation worth a loud, retried failure).
    """
    wanted = set(fixtures)
    present = {p.name for p in run_dir.iterdir() if p.is_dir()}
    missing = wanted - present
    if missing:
        raise RuntimeError(
            f"Shard fixtures missing from submission zip: "
            f"{', '.join(sorted(missing))}"
        )
    removed = 0
    for child in run_dir.iterdir():
        if child.is_dir() and child.name not in wanted:
            shutil.rmtree(child)
            removed += 1
    print(
        f"[eval_job] pruned run dir to {len(wanted)} shard fixture(s) "
        f"(removed {removed})",
        flush=True,
    )


def _upload_shard_artifacts(
    submission_id: str,
    shard_id: str,
    run_dir: Path,
    submissions_repo: str,
    token: str,
) -> None:
    """Upload this shard's evaluated per-fixture dirs for the Space to merge.

    Persists the pruned ``run_dir`` (each ``<fixture>/`` with its
    ``result.json`` + ``renders/`` + any overlay PNGs) verbatim. In
    bucket mode, this syncs the dir into the HF Storage Bucket via the
    bucket API (no volume mount); in legacy mode, it is one dataset-repo
    commit under ``reports/<id>/shards/<shard_id>/``. The Space reads
    every shard's tree, copies the fixture dirs into a single merged run
    dir, and builds the aggregate ``run_summary`` + report + gallery from
    the whole. The per-shard ``run_summary.json`` written by
    ``cadgenbench evaluate`` rides along harmlessly; the merge recomputes
    it over the union and ignores the partials.
    """
    bucket_id = os.environ.get(SHARD_BUCKET_ENV, "").strip()
    if bucket_id:
        if bucket_id.startswith("hf://buckets/"):
            bucket_id = bucket_id[len("hf://buckets/"):]
        bucket_id = bucket_id.rstrip("/")
        prefix = os.environ.get(SHARD_BUCKET_PREFIX_ENV, "submissions").strip("/")
        dest = (
            f"hf://buckets/{bucket_id}/{prefix}/{submission_id}/"
            f"{SHARDS_DIR_NAME}/{shard_id}"
        )
        api = HfApi(token=token)
        api.sync_bucket(source=str(run_dir), dest=dest, token=token)
        print(
            f"[eval_job] synced shard {shard_id} -> {dest}",
            flush=True,
        )
        return

    api = HfApi(token=token)
    path_in_repo = f"{REPORTS_DIR_IN_REPO}/{submission_id}/{SHARDS_DIR_NAME}/{shard_id}"
    api.upload_folder(
        folder_path=str(run_dir),
        path_in_repo=path_in_repo,
        repo_id=submissions_repo,
        repo_type="dataset",
        commit_message=f"add eval shard {shard_id} for {submission_id}",
    )
    print(
        f"[eval_job] uploaded shard {shard_id} -> {path_in_repo}",
        flush=True,
    )


def _run_eval(run_dir: Path, workers: int) -> None:
    """Invoke ``cadgenbench evaluate`` over *run_dir*; raise on non-zero."""
    cmd = [
        sys.executable, "-m", "cadgenbench.cli", "evaluate", str(run_dir),
        "--workers", str(workers),
    ]
    print(f"[eval_job] {' '.join(cmd)}", flush=True)
    proc = subprocess.run(
        cmd,
        timeout=EVAL_TIMEOUT_SECONDS,
        env=os.environ.copy(),
        check=False,
    )
    if proc.returncode != 0:
        raise RuntimeError(
            f"cadgenbench evaluate exited {proc.returncode}"
        )


def _run_report(
    run_dir: Path,
    html_out: Path,
    *,
    render_base_url: str | None = None,
    download_url: str | None = None,
) -> None:
    """Invoke ``cadgenbench report single`` for *run_dir*; raise on non-zero.

    Passes ``--render-base-url`` so candidate renders are referenced from the
    public bucket rather than base64-inlined into the hosted HTML.
    """
    cmd = [
        sys.executable, "-m", "cadgenbench.cli", "report", "single",
        str(run_dir), "-o", str(html_out),
    ]
    if render_base_url:
        cmd += [
            "--render-base-url", render_base_url,
            # GT + inputs are private, so they link through the Space proxy
            # rather than the public bucket. Passed alongside the render base
            # so the whole hosted report is lazy-loaded links, not base64.
            "--gt-base-url", GT_PROXY_BASE_URL,
            "--input-base-url", INPUT_PROXY_BASE_URL,
        ]
    if download_url:
        cmd += ["--download-url", download_url]
    print(f"[eval_job] {' '.join(cmd)}", flush=True)
    proc = subprocess.run(
        cmd,
        timeout=REPORT_TIMEOUT_SECONDS,
        env=os.environ.copy(),
        check=False,
    )
    if proc.returncode != 0 or not html_out.is_file():
        raise RuntimeError(
            f"cadgenbench report single exited {proc.returncode} "
            f"(html exists={html_out.is_file()})"
        )


def _build_report_json(run_dir: Path) -> dict[str, Any]:
    """Bundle ``run_summary.json`` + every per-fixture ``result.json``.

    Identical shape to submit.py's ``_build_report_json``: the
    Space-side worker reads ``report.json`` after the Job completes
    and pulls ``run_summary`` out of it to flip the row.
    """
    summary_path = run_dir / "run_summary.json"
    if not summary_path.is_file():
        raise RuntimeError(
            f"run_summary.json not produced under {run_dir} (eval issue?)"
        )
    summary = json.loads(summary_path.read_text(encoding="utf-8"))
    per_fixture: dict[str, dict[str, Any]] = {}
    for fixture_dir in sorted(d for d in run_dir.iterdir() if d.is_dir()):
        rp = fixture_dir / "result.json"
        if rp.is_file():
            per_fixture[fixture_dir.name] = json.loads(
                rp.read_text(encoding="utf-8")
            )
    return {"run_summary": summary, "per_fixture_results": per_fixture}


def _publish_reports_and_gallery(
    submission_id: str,
    html_path: Path,
    report_json: dict[str, Any],
    submissions_repo: str,
    token: str,
) -> None:
    """Publish the report HTML + JSON to the submissions dataset in one commit.

    Renders are **not** committed here: :func:`_upload_renders_to_bucket` has
    already pushed them to the public render bucket under ``renders/<id>/``, and
    the report HTML references them by bucket URL. Keeping the binary renders
    out of the dataset repo avoids bloating its git history and the commit-queue
    429s the per-file fan-out used to cause.
    """
    operations: list[CommitOperationAdd] = [
        CommitOperationAdd(
            path_in_repo=f"{REPORTS_DIR_IN_REPO}/{submission_id}.html",
            path_or_fileobj=str(html_path),
        ),
        CommitOperationAdd(
            path_in_repo=f"{REPORTS_DIR_IN_REPO}/{submission_id}.json",
            path_or_fileobj=json.dumps(
                report_json, ensure_ascii=False, indent=2,
            ).encode("utf-8"),
        ),
    ]
    api = HfApi(token=token)
    api.create_commit(
        repo_id=submissions_repo,
        repo_type="dataset",
        operations=operations,
        commit_message=f"publish report for {submission_id}",
    )
    print(
        f"[eval_job] published reports/{submission_id}.{{html,json}}",
        flush=True,
    )


if __name__ == "__main__":
    sys.exit(main())