tests/test_model_card.py

from __future__ import annotations

import json

import pytest

from shiftedx_bench.model_card import (
    END_MARKER,
    START_MARKER,
    render_benchmark_block,
    update_model_card,
    validate_post_publish_run,
)


MODEL_REPO = "Shiftedx/example-quant"
MODEL_REVISION = "a" * 40
BENCHMARK_REVISION = "b" * 40
VARIANT = "example"


def _write_run(tmp_path, *, omit_lane=None):
    counts = {"long-context": 15, "quality": 10, "tools": 6, "agentic": 2, "vision": 4}
    names = {"long-context": "context", "quality": "quality", "tools": "tools", "agentic": "agentic", "vision": "vision"}
    for lane, count in counts.items():
        if lane == omit_lane:
            count -= 1
        rows = []
        for index in range(count):
            metadata = {}
            if lane == "long-context":
                metadata = {
                    "target_prompt_tokens": [4096, 16384, 65536, 131072][min(index // 4, 3)],
                    "requested_position": 0.5,
                    "family": "single_key",
                }
            rows.append(
                {
                    "schema_version": "1.0",
                    "variant": VARIANT,
                    "lane": lane,
                    "case_id": f"{lane}-{index}",
                    "passed": index != count - 1 if lane == "quality" else True,
                    "metadata": metadata,
                    "telemetry": {
                        "wall_s": 2.0,
                        "decode_tokens_per_second": 20.0,
                        "active_memory_bytes": 16 * 1024 ** 3,
                    },
                }
            )
        (tmp_path / f"{names[lane]}.jsonl").write_text(
            "".join(json.dumps(row) + "\n" for row in rows), encoding="utf-8"
        )
        manifest = {
            "benchmark_version": "0.3.0",
            "variants": [{"label": VARIANT, "request_overrides": {}}],
            "evaluated_artifact": {"repo_id": MODEL_REPO, "revision": MODEL_REVISION},
            "benchmark_source": {
                "repo_id": "Shiftedx/shiftedx-bench",
                "revision": BENCHMARK_REVISION,
            },
        }
        (tmp_path / f"{names[lane]}.manifest.json").write_text(json.dumps(manifest), encoding="utf-8")


def test_post_publish_model_card_block_is_complete_and_idempotent(tmp_path):
    _write_run(tmp_path)
    rows, version = validate_post_publish_run(
        tmp_path,
        variant=VARIANT,
        profile="vision",
        model_repo=MODEL_REPO,
        model_revision=MODEL_REVISION,
        benchmark_revision=BENCHMARK_REVISION,
    )
    block = render_benchmark_block(
        rows,
        variant=VARIANT,
        model_repo=MODEL_REPO,
        model_revision=MODEL_REVISION,
        benchmark_revision=BENCHMARK_REVISION,
        benchmark_version=version,
        host="Apple Silicon, 64 GiB",
        runtime="MTPLX 2.7.1, reasoning on/medium, native tools",
        kv_cache="off",
        mtp_depth="3",
        max_window_status="host-limited",
    )
    readme = "# Example\n\n## Provenance and integrity\n"
    first = update_model_card(readme, block)
    second = update_model_card(first, block)
    assert first == second
    assert first.count(START_MARKER) == first.count(END_MARKER) == 1
    assert "15/15" in first
    assert "host watchdog" in first
    assert MODEL_REVISION in first and BENCHMARK_REVISION in first


def test_model_card_export_rejects_incomplete_or_wrong_revision(tmp_path):
    _write_run(tmp_path, omit_lane="vision")
    with pytest.raises(ValueError, match="Incomplete"):
        validate_post_publish_run(
            tmp_path,
            variant=VARIANT,
            profile="vision",
            model_repo=MODEL_REPO,
            model_revision=MODEL_REVISION,
            benchmark_revision=BENCHMARK_REVISION,
        )
    with pytest.raises(ValueError, match="provenance mismatch"):
        validate_post_publish_run(
            tmp_path,
            variant=VARIANT,
            profile="text",
            model_repo=MODEL_REPO,
            model_revision="c" * 40,
            benchmark_revision=BENCHMARK_REVISION,
        )


def test_model_card_renders_null_effective_context_as_not_established(tmp_path):
    _write_run(tmp_path)
    context_path = tmp_path / "context.jsonl"
    rows = [json.loads(line) for line in context_path.read_text().splitlines()]
    for row in rows:
        row["passed"] = False
    context_path.write_text(
        "".join(json.dumps(row) + "\n" for row in rows), encoding="utf-8"
    )
    validated, version = validate_post_publish_run(
        tmp_path,
        variant=VARIANT,
        profile="vision",
        model_repo=MODEL_REPO,
        model_revision=MODEL_REVISION,
        benchmark_revision=BENCHMARK_REVISION,
    )
    block = render_benchmark_block(
        validated,
        variant=VARIANT,
        model_repo=MODEL_REPO,
        model_revision=MODEL_REVISION,
        benchmark_revision=BENCHMARK_REVISION,
        benchmark_version=version,
        host="Apple Silicon, 64 GiB",
        runtime="MTPLX 2.7.1",
        kv_cache="off",
        mtp_depth="2",
        max_window_status="not-run",
    )
    assert "effective tested context: not established at the benchmark's 90% threshold" in block