| 1 | """Unit tests for the pure helpers inside ``backends/hf.py``. |
| 2 | |
| 3 | The helpers (``_resolve_dtype``, ``_detect_device``) gate the HF |
| 4 | backend's dtype/device choices. They're exercised end-to-end by the |
| 5 | integration suite, but doing direct unit coverage on them keeps the |
| 6 | slow lane focused on the actual backend behavior — and makes the |
| 7 | fast lane catch dtype/device regressions in seconds. |
| 8 | """ |
| 9 | |
| 10 | from __future__ import annotations |
| 11 | |
| 12 | import importlib.util |
| 13 | |
| 14 | import pytest |
| 15 | |
| 16 | # These tests need torch to construct the dtype objects we're asserting |
| 17 | # on; skip cleanly when the [hf] extra isn't installed. |
| 18 | if importlib.util.find_spec("torch") is None: |
| 19 | pytest.skip( |
| 20 | "torch not installed — install the [hf] extra to run HF helper tests", |
| 21 | allow_module_level=True, |
| 22 | ) |
| 23 | |
| 24 | from dlm_sway.backends.hf import _detect_device, _resolve_dtype |
| 25 | |
| 26 | |
| 27 | class TestResolveDtype: |
| 28 | def test_explicit_fp16(self) -> None: |
| 29 | import torch |
| 30 | |
| 31 | assert _resolve_dtype("fp16", "cpu") is torch.float16 |
| 32 | |
| 33 | def test_explicit_bf16(self) -> None: |
| 34 | import torch |
| 35 | |
| 36 | assert _resolve_dtype("bf16", "cpu") is torch.bfloat16 |
| 37 | |
| 38 | def test_explicit_fp32(self) -> None: |
| 39 | import torch |
| 40 | |
| 41 | assert _resolve_dtype("fp32", "cpu") is torch.float32 |
| 42 | |
| 43 | def test_auto_on_cpu_picks_fp32_for_numerical_stability(self) -> None: |
| 44 | import torch |
| 45 | |
| 46 | assert _resolve_dtype("auto", "cpu") is torch.float32 |
| 47 | |
| 48 | def test_auto_on_mps_picks_fp16(self) -> None: |
| 49 | import torch |
| 50 | |
| 51 | assert _resolve_dtype("auto", "mps") is torch.float16 |
| 52 | |
| 53 | |
| 54 | class TestDetectDevice: |
| 55 | def test_returns_one_of_supported_devices(self) -> None: |
| 56 | device = _detect_device() |
| 57 | assert device in ("cuda", "mps", "cpu") |