unsloth/tests/studio/test_hardware_dispatch_matrix.py

# SPDX-License-Identifier: AGPL-3.0-only
"""Studio hardware dispatch matrix: spoofs platform/torch/mlx per PROFILES to exercise CUDA/ROCm/XPU/MLX/CPU paths without real hardware."""

from __future__ import annotations

import importlib
import importlib.machinery
import importlib.util
import sys
import types
from dataclasses import dataclass, field
from pathlib import Path
from typing import Optional

import pytest


REPO_ROOT = Path(__file__).resolve().parents[2]
STUDIO_BACKEND = REPO_ROOT / "studio" / "backend"


@dataclass
class HardwareProfile:
    name: str
    system: str  # platform.system() value
    machine: str  # platform.machine() value
    cuda_available: bool  # torch.cuda.is_available() value
    hip_version: Optional[str]  # torch.version.hip; None for NVIDIA, "6.1" etc. for ROCm
    xpu_available: bool  # torch.xpu.is_available() value
    has_mlx: bool  # whether to inject a fake mlx into sys.modules
    mps_available: bool  # torch.backends.mps.is_available() value

    expect_is_mlx: bool  # unsloth._IS_MLX
    expect_device_type: str  # Studio DeviceType (uppercased name: "CUDA"/"XPU"/"MLX"/"CPU")
    expect_is_rocm: bool  # Studio IS_ROCM
    expect_apple_silicon: bool  # Studio is_apple_silicon()
    extra_notes: str = ""


PROFILES = [
    HardwareProfile(
        name = "nvidia_cuda",
        system = "Linux",
        machine = "x86_64",
        cuda_available = True,
        hip_version = None,
        xpu_available = False,
        has_mlx = False,
        mps_available = False,
        expect_is_mlx = False,
        expect_device_type = "CUDA",
        expect_is_rocm = False,
        expect_apple_silicon = False,
    ),
    HardwareProfile(
        name = "amd_rocm",
        system = "Linux",
        machine = "x86_64",
        cuda_available = True,
        hip_version = "6.1",
        xpu_available = False,
        has_mlx = False,
        mps_available = False,
        expect_is_mlx = False,
        expect_device_type = "CUDA",
        expect_is_rocm = True,
        expect_apple_silicon = False,
        extra_notes = "PyTorch ROCm reuses torch.cuda.* over HIP; "
        "Studio still uses DeviceType.CUDA but flips IS_ROCM=True.",
    ),
    HardwareProfile(
        name = "intel_xpu",
        system = "Linux",
        machine = "x86_64",
        cuda_available = False,
        hip_version = None,
        xpu_available = True,
        has_mlx = False,
        mps_available = False,
        expect_is_mlx = False,
        expect_device_type = "XPU",
        expect_is_rocm = False,
        expect_apple_silicon = False,
    ),
    HardwareProfile(
        name = "apple_silicon_mlx",
        system = "Darwin",
        machine = "arm64",
        cuda_available = False,
        hip_version = None,
        xpu_available = False,
        has_mlx = True,
        mps_available = True,
        expect_is_mlx = True,
        expect_device_type = "MLX",
        expect_is_rocm = False,
        expect_apple_silicon = True,
    ),
    HardwareProfile(
        name = "apple_silicon_no_mlx",
        system = "Darwin",
        machine = "arm64",
        cuda_available = False,
        hip_version = None,
        xpu_available = False,
        has_mlx = False,
        mps_available = True,
        expect_is_mlx = False,
        expect_device_type = "CPU",
        expect_is_rocm = False,
        expect_apple_silicon = True,
        extra_notes = "Mac without mlx falls through to CPU (chat-only).",
    ),
    HardwareProfile(
        name = "linux_arm64_with_mlx",
        system = "Linux",
        machine = "arm64",
        cuda_available = False,
        hip_version = None,
        xpu_available = False,
        has_mlx = True,
        mps_available = False,
        expect_is_mlx = False,
        expect_device_type = "CPU",
        expect_is_rocm = False,
        expect_apple_silicon = False,
        extra_notes = "Canary: Linux ARM64 with mlx package installed must NOT "
        "trigger MLX dispatch; the system check is what guards it.",
    ),
    HardwareProfile(
        name = "cpu_only",
        system = "Linux",
        machine = "x86_64",
        cuda_available = False,
        hip_version = None,
        xpu_available = False,
        has_mlx = False,
        mps_available = False,
        expect_is_mlx = False,
        expect_device_type = "CPU",
        expect_is_rocm = False,
        expect_apple_silicon = False,
    ),
]

PROFILE_IDS = [p.name for p in PROFILES]


@pytest.fixture
def spoof_hardware(monkeypatch):
    """Return a function that applies a HardwareProfile to the live process; monkeypatch cleans up on exit."""

    def _apply(profile: HardwareProfile) -> None:
        import platform
        import torch

        # platform spoof (used by both the unsloth gate and Studio's helpers)
        monkeypatch.setattr(platform, "system", lambda: profile.system)
        monkeypatch.setattr(platform, "machine", lambda: profile.machine)

        monkeypatch.setattr(torch.cuda, "is_available", lambda: profile.cuda_available)
        # Stub get_device_properties: detect_hardware reads .name, which crashes on a CPU CI runner.
        if profile.cuda_available:
            stub_props = types.SimpleNamespace(
                name = "Stub GPU" if not profile.hip_version else "Stub AMD GPU",
            )
            monkeypatch.setattr(
                torch.cuda,
                "get_device_properties",
                lambda i = 0: stub_props,
                raising = False,
            )

        # torch.version.hip: None on NVIDIA, "6.1" etc. on ROCm
        torch_version = torch.version
        monkeypatch.setattr(torch_version, "hip", profile.hip_version, raising = False)

        # Stub torch.xpu.* always; real get_device_name needs the XPU torch build.
        if hasattr(torch, "xpu"):
            monkeypatch.setattr(torch.xpu, "is_available", lambda: profile.xpu_available)
            monkeypatch.setattr(
                torch.xpu,
                "get_device_name",
                lambda i = 0: "Intel XPU (stub)",
                raising = False,
            )
        elif profile.xpu_available:
            xpu_stub = types.SimpleNamespace(
                is_available = lambda: True,
                get_device_name = lambda i = 0: "Intel XPU (stub)",
            )
            monkeypatch.setattr(torch, "xpu", xpu_stub, raising = False)

        # torch.backends.mps.is_available
        if hasattr(torch.backends, "mps"):
            monkeypatch.setattr(torch.backends.mps, "is_available", lambda: profile.mps_available)

        # mlx + mlx.core in sys.modules
        if profile.has_mlx:
            fake_mlx = types.ModuleType("mlx")
            fake_mlx.__spec__ = importlib.machinery.ModuleSpec("mlx", loader = None)
            fake_mlx.__path__ = []
            fake_mlx_core = types.ModuleType("mlx.core")
            fake_mlx.core = fake_mlx_core
            monkeypatch.setitem(sys.modules, "mlx", fake_mlx)
            monkeypatch.setitem(sys.modules, "mlx.core", fake_mlx_core)
            # detect_hardware now gates MLX on the full stack via
            # utils.mlx_repair.mlx_stack_available() (it imports mlx_lm/mlx_vlm and
            # checks dist versions), which faking only mlx.core cannot satisfy. An
            # mlx profile means a complete, healthy stack, so model that here;
            # mlx_stack_available's own internals are covered by test_mlx_repair.py.
            if str(STUDIO_BACKEND) not in sys.path:
                sys.path.insert(0, str(STUDIO_BACKEND))
            import utils.mlx_repair as _mlx_repair  # type: ignore

            monkeypatch.setattr(_mlx_repair, "mlx_stack_available", lambda: True)
        else:
            # Drop cached mlx and patch find_spec so the unsloth gate sees mlx as absent.
            monkeypatch.delitem(sys.modules, "mlx", raising = False)
            monkeypatch.delitem(sys.modules, "mlx.core", raising = False)
            real_find_spec = importlib.util.find_spec

            def _no_mlx(name, *args, **kwargs):
                if name == "mlx" or name.startswith("mlx."):
                    return None
                return real_find_spec(name, *args, **kwargs)

            monkeypatch.setattr(importlib.util, "find_spec", _no_mlx)

            # Studio's _has_mlx() does `import mlx.core`, not find_spec; block it
            # with a meta_path finder that raises ImportError for mlx.*.
            class _BlockMLXFinder:
                def find_spec(
                    self_inner,
                    name,
                    path = None,
                    target = None,
                ):
                    if name == "mlx" or name.startswith("mlx."):
                        raise ImportError(
                            f"mlx import blocked by spoof_hardware " f"(profile={profile.name})"
                        )
                    return None

            blocker = _BlockMLXFinder()
            # New list so monkeypatch fully restores on teardown.
            monkeypatch.setattr(
                sys,
                "meta_path",
                [blocker, *sys.meta_path],
            )

    return _apply


def _evaluate_unsloth_is_mlx_gate() -> bool:
    """Re-evaluate the exact expression from unsloth/__init__.py:20-24."""
    import importlib.util
    import platform

    return (
        platform.system() == "Darwin"
        and platform.machine() == "arm64"
        and importlib.util.find_spec("mlx") is not None
    )


def _import_studio_hardware_module():
    """Lazy-load Studio's hardware module under the bare-imports layout."""
    if str(STUDIO_BACKEND) not in sys.path:
        sys.path.insert(0, str(STUDIO_BACKEND))
    # Fresh import so detect_hardware re-runs under the current spoofs.
    sys.modules.pop("utils.hardware.hardware", None)
    sys.modules.pop("utils.hardware", None)
    from utils.hardware import hardware as hw  # type: ignore

    return hw


@pytest.mark.parametrize("profile", PROFILES, ids = PROFILE_IDS)
def test_unsloth_is_mlx_gate_matches_profile(profile, spoof_hardware):
    """The _IS_MLX expression in unsloth/__init__.py flips correctly per profile."""
    spoof_hardware(profile)
    actual = _evaluate_unsloth_is_mlx_gate()
    assert actual is profile.expect_is_mlx, (
        f"profile {profile.name}: expected _IS_MLX={profile.expect_is_mlx}, "
        f"got {actual}. {profile.extra_notes}"
    )


@pytest.mark.parametrize("profile", PROFILES, ids = PROFILE_IDS)
def test_studio_detect_hardware_matches_profile(profile, spoof_hardware):
    """Studio's detect_hardware() routes to the right DeviceType per profile."""
    spoof_hardware(profile)
    hw = _import_studio_hardware_module()
    detected = hw.detect_hardware()
    expected = getattr(hw.DeviceType, profile.expect_device_type)
    assert detected == expected, (
        f"profile {profile.name}: expected {profile.expect_device_type}, "
        f"got {detected!r}. {profile.extra_notes}"
    )
    assert hw.IS_ROCM is profile.expect_is_rocm, (
        f"profile {profile.name}: expected IS_ROCM={profile.expect_is_rocm}, " f"got {hw.IS_ROCM}"
    )


@pytest.mark.parametrize("profile", PROFILES, ids = PROFILE_IDS)
def test_studio_is_apple_silicon_matches_profile(profile, spoof_hardware):
    """Studio's is_apple_silicon() helper agrees with platform spoof."""
    spoof_hardware(profile)
    hw = _import_studio_hardware_module()
    assert hw.is_apple_silicon() is profile.expect_apple_silicon, (
        f"profile {profile.name}: expected is_apple_silicon={profile.expect_apple_silicon}, "
        f"got {hw.is_apple_silicon()}"
    )


# Negative-space tests: catch regressions where the dispatch order changes.


def test_cuda_takes_priority_over_mlx_when_both_available(spoof_hardware):
    """CUDA wins over MLX when both available: canary against GPU users being routed to MLX after refactors."""
    profile = HardwareProfile(
        name = "cuda_plus_mlx",
        system = "Darwin",
        machine = "arm64",
        cuda_available = True,
        hip_version = None,
        xpu_available = False,
        has_mlx = True,
        mps_available = True,
        expect_is_mlx = True,
        expect_device_type = "CUDA",
        expect_is_rocm = False,
        expect_apple_silicon = True,
    )
    spoof_hardware(profile)
    hw = _import_studio_hardware_module()
    assert hw.detect_hardware() == hw.DeviceType.CUDA


def test_xpu_takes_priority_over_mlx_when_both_available(spoof_hardware):
    """XPU is selected over MLX in the dispatch order."""
    profile = HardwareProfile(
        name = "xpu_plus_mlx",
        system = "Darwin",
        machine = "arm64",
        cuda_available = False,
        hip_version = None,
        xpu_available = True,
        has_mlx = True,
        mps_available = True,
        expect_is_mlx = True,
        expect_device_type = "XPU",
        expect_is_rocm = False,
        expect_apple_silicon = True,
    )
    spoof_hardware(profile)
    hw = _import_studio_hardware_module()
    assert hw.detect_hardware() == hw.DeviceType.XPU