tests/test_runtime_accelerator.py

"""Picking the build that suits the machine, and stepping down when it does not.

The benchmark that prompted all of this ran twenty agent workloads on a laptop
with a CUDA card, at 3.1 tokens per second, with the GPU sitting at its 7 W idle
draw. Nothing was broken: the only recorded build was the CPU one, so the CPU
did the work. These are the properties that stop that happening silently, and
the ones that stop the cure being worse -- a CUDA build installed on a machine
that cannot start it is a worker that does not run at all, which is worse than a
slow one.
"""

from __future__ import annotations

import io
import json
import zipfile
from pathlib import Path

import pytest

from distinct_agent import runtime

CUDA_URL = "https://github.com/ggml-org/llama.cpp/releases/download/bT/cuda.zip"
CUDART_URL = "https://github.com/ggml-org/llama.cpp/releases/download/bT/cudart.zip"
CPU_URL = "https://github.com/ggml-org/llama.cpp/releases/download/bT/cpu.zip"


def _zip(*names: str) -> bytes:
    buf = io.BytesIO()
    with zipfile.ZipFile(buf, "w") as handle:
        for name in names:
            handle.writestr(f"build/bin/{name}", b"x" * 16)
    return buf.getvalue()


def _sha(payload: bytes) -> str:
    import hashlib

    return hashlib.sha256(payload).hexdigest()


def _serve(bodies: dict[str, bytes]):
    """An opener that answers each pinned URL with its own archive."""

    class _Response:
        def __init__(self, payload: bytes) -> None:
            self.headers = {"Content-Length": str(len(payload))}
            self._buffer = io.BytesIO(payload)

        def read(self, size=-1):
            return self._buffer.read(size)

        def __enter__(self):
            return self

        def __exit__(self, *exc):
            return False

    def opener(request, timeout=None):
        url = request.full_url if hasattr(request, "full_url") else str(request)
        if url not in bodies:
            raise AssertionError(f"unpinned URL was fetched: {url}")
        return _Response(bodies[url])

    return opener


def _pins(tmp_path: Path, entries: dict) -> Path:
    path = tmp_path / "pins.json"
    path.write_text(json.dumps({"builds": entries}), encoding="utf-8")
    return path


@pytest.fixture()
def pinned(tmp_path: Path):
    cuda, cudart, cpu = _zip("llama-server.exe"), _zip("cudart64_12.dll"), _zip("llama-server.exe")
    path = _pins(
        tmp_path,
        {
            "test-x64": {
                "tag": "bT",
                "asset": "cpu.zip",
                "url": CPU_URL,
                "sha256": _sha(cpu),
                "bytes": len(cpu),
            },
            "test-x64-cuda": {
                "tag": "bT",
                "accelerator": "cuda",
                "requires_cuda": 12040,
                "archives": [
                    {"asset": "cuda.zip", "url": CUDA_URL, "sha256": _sha(cuda), "bytes": len(cuda)},
                    {
                        "asset": "cudart.zip",
                        "url": CUDART_URL,
                        "sha256": _sha(cudart),
                        "bytes": len(cudart),
                    },
                ],
            },
        },
    )
    return path, _serve({CUDA_URL: cuda, CUDART_URL: cudart, CPU_URL: cpu})


def test_a_cuda_build_is_offered_only_when_the_driver_can_run_it(pinned, monkeypatch) -> None:
    path, _ = pinned
    monkeypatch.setattr(runtime, "cuda_capability", lambda: 12040)
    assert runtime.preferred_keys("test-x64", path) == ["test-x64-cuda", "test-x64"]


def test_a_driver_below_the_pinned_cuda_version_gets_the_cpu_build(pinned, monkeypatch) -> None:
    path, _ = pinned
    monkeypatch.setattr(runtime, "cuda_capability", lambda: 12000)
    assert runtime.preferred_keys("test-x64", path) == ["test-x64"]


def test_a_machine_with_no_nvidia_card_is_never_offered_cuda(pinned, monkeypatch) -> None:
    path, _ = pinned
    monkeypatch.setattr(runtime, "cuda_capability", lambda: 0)
    assert runtime.preferred_keys("test-x64", path) == ["test-x64"]


def test_the_probe_is_not_run_when_no_accelerated_build_is_recorded(tmp_path, monkeypatch) -> None:
    """No NVML call, no nvidia-smi subprocess, on the machines that cannot use one."""

    asked = []
    monkeypatch.setattr(runtime, "cuda_capability", lambda: asked.append(1) or 0)
    path = _pins(tmp_path, {"test-x64": {"tag": "bT", "url": CPU_URL, "sha256": "a" * 64}})
    assert runtime.preferred_keys("test-x64", path) == ["test-x64"]
    assert asked == []


def test_an_accelerated_pin_installs_every_archive_it_names(pinned, tmp_path, monkeypatch) -> None:
    """The CUDA runtime is not optional: without it the executable does not load."""

    path, opener = pinned
    monkeypatch.setattr(runtime, "cuda_capability", lambda: 12040)
    root = tmp_path / "proj"
    root.mkdir()
    found = runtime.ensure(root=root, key="test-x64-cuda", opener=opener, pins_path=path)
    directory = root / "runtime" / "test-x64-cuda"
    assert Path(found).parent == directory
    assert {item.name for item in directory.iterdir()} == {"llama-server.exe", "cudart64_12.dll"}


def test_a_second_archive_that_fails_leaves_no_half_install(tmp_path, monkeypatch) -> None:
    cuda, cudart = _zip("llama-server.exe"), _zip("cudart64_12.dll")
    path = _pins(
        tmp_path,
        {
            "test-x64-cuda": {
                "tag": "bT",
                "accelerator": "cuda",
                "requires_cuda": 12040,
                "archives": [
                    {"asset": "cuda.zip", "url": CUDA_URL, "sha256": _sha(cuda)},
                    {"asset": "cudart.zip", "url": CUDART_URL, "sha256": "0" * 64},
                ],
            }
        },
    )
    root = tmp_path / "proj"
    root.mkdir()
    with pytest.raises(runtime.RuntimeVerificationError):
        runtime.ensure(
            root=root,
            key="test-x64-cuda",
            opener=_serve({CUDA_URL: cuda, CUDART_URL: cudart}),
            pins_path=path,
        )
    assert not (root / "runtime" / "test-x64-cuda").exists()


def test_a_build_that_will_not_start_steps_down_to_the_next_one(pinned, tmp_path, monkeypatch) -> None:
    """A driver may advertise a version it cannot honour. Starting it is the proof."""

    path, opener = pinned
    monkeypatch.setattr(runtime, "cuda_capability", lambda: 12040)
    root = tmp_path / "proj"
    root.mkdir()
    refused = []

    def verify(executable: str) -> bool:
        if "cuda" in executable:
            refused.append(executable)
            return False
        return True

    found = runtime.ensure_best(
        root=root, key="test-x64", opener=opener, pins_path=path, verify=verify
    )
    assert refused, "the accelerated build should have been tried first"
    assert Path(found).parent.name == "test-x64"


def test_when_nothing_starts_the_reason_names_every_build_that_was_tried(
    pinned, tmp_path, monkeypatch
) -> None:
    path, opener = pinned
    monkeypatch.setattr(runtime, "cuda_capability", lambda: 12040)
    root = tmp_path / "proj"
    root.mkdir()
    with pytest.raises(runtime.RuntimeUnavailable) as caught:
        runtime.ensure_best(
            root=root, key="test-x64", opener=opener, pins_path=path, verify=lambda _: False
        )
    assert "test-x64-cuda" in str(caught.value)
    assert "test-x64" in str(caught.value)


def test_an_older_flat_install_is_moved_rather_than_re_downloaded(tmp_path, monkeypatch) -> None:
    """Builds used to unpack straight into runtime/. That install still counts."""

    monkeypatch.setattr(runtime, "platform_key", lambda *a, **k: "test-x64")
    root = tmp_path / "proj"
    flat = root / "runtime"
    flat.mkdir(parents=True)
    (flat / "llama-server.exe").write_bytes(b"old")
    (flat / "ggml.dll").write_bytes(b"old")

    runtime._migrate_flat_install(root)

    assert not (flat / "llama-server.exe").exists()
    moved = flat / "test-x64"
    assert (moved / "llama-server.exe").read_bytes() == b"old"
    assert (moved / "ggml.dll").exists()
    assert runtime.installed_server(root, "test-x64") == str(moved / "llama-server.exe")


def test_the_cpu_build_is_always_the_last_resort(pinned, monkeypatch) -> None:
    """Whatever else is recorded, the list ends with the build that runs anywhere."""

    path, _ = pinned
    for capability in (0, 12000, 12040, 13000):
        monkeypatch.setattr(runtime, "cuda_capability", lambda c=capability: c)
        assert runtime.preferred_keys("test-x64", path)[-1] == "test-x64"


def test_an_accelerator_this_selector_cannot_check_is_not_installed(tmp_path, monkeypatch) -> None:
    """A HIP pin on an unexamined machine is exactly the failure being prevented."""

    payload = _zip("llama-server.exe")
    path = _pins(
        tmp_path,
        {
            "test-x64": {"tag": "bT", "url": CPU_URL, "sha256": _sha(payload)},
            "test-x64-hip": {"tag": "bT", "accelerator": "hip", "url": CUDA_URL, "sha256": _sha(payload)},
        },
    )
    assert runtime.preferred_keys("test-x64", path) == ["test-x64"]


def test_a_pin_pointing_off_github_is_refused_before_anything_is_fetched(tmp_path) -> None:
    payload = _zip("llama-server.exe")
    path = _pins(
        tmp_path,
        {
            "test-x64-cuda": {
                "tag": "bT",
                "accelerator": "cuda",
                "archives": [
                    {"asset": "cuda.zip", "url": CUDA_URL, "sha256": _sha(payload)},
                    {"asset": "x.zip", "url": "https://example.invalid/x.zip", "sha256": _sha(payload)},
                ],
            }
        },
    )
    root = tmp_path / "proj"
    root.mkdir()

    def refuse(request, timeout=None):
        raise AssertionError("nothing should have been fetched")

    with pytest.raises(runtime.RuntimeUnavailable, match="not GitHub"):
        runtime.ensure(root=root, key="test-x64-cuda", opener=refuse, pins_path=path)


def test_the_cuda_probe_survives_a_machine_with_no_bindings_and_no_smi(monkeypatch) -> None:
    """Probing is best-effort by construction: it must never be the thing that fails."""

    monkeypatch.setattr(runtime, "_cuda_from_nvml", lambda: 0)
    monkeypatch.setattr(runtime.shutil, "which", lambda name: None)
    assert runtime.cuda_capability() == 0