Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
"""Regression coverage for #563: BF16-capable MPS text trainers."""
from __future__ import annotations
from pathlib import Path
from types import SimpleNamespace
import pytest
@pytest.fixture()
def fake_mps(monkeypatch):
import torch
probes = []
class _MPS:
available = True
def is_available(self):
return self.available
backend = _MPS()
monkeypatch.setattr(torch.backends, "mps", backend)
def apply(*, available=True, accepts_bf16=True):
backend.available = available
probes.clear()
def empty(size, *, dtype, device):
probes.append((size, dtype, device))
if not accepts_bf16:
raise TypeError("MPS runtime rejected bfloat16")
return object()
monkeypatch.setattr(torch, "empty", empty)
return probes
return apply
class TestMPSBF16Capability:
def test_available_runtime_that_accepts_bf16_is_supported(self, fake_mps):
import torch
from soup_cli.utils.gpu import mps_supports_bf16
probes = fake_mps(available=True, accepts_bf16=True)
assert mps_supports_bf16() is True
assert probes == [(1, torch.bfloat16, "mps")]
def test_unavailable_runtime_does_not_probe_or_claim_bf16(self, fake_mps):
from soup_cli.utils.gpu import mps_supports_bf16
probes = fake_mps(available=False, accepts_bf16=True)
assert mps_supports_bf16() is False
assert probes == []
def test_runtime_rejection_falls_back_to_fp32(self, fake_mps):
from soup_cli.utils.gpu import mps_supports_bf16
fake_mps(available=True, accepts_bf16=False)
assert mps_supports_bf16() is False
class TestPrecisionPolicy:
def test_verified_text_trainer_can_opt_into_mps_bf16(self, fake_mps):
from soup_cli.utils.gpu import bf16_fp16_flags
fake_mps(available=True, accepts_bf16=True)
assert bf16_fp16_flags("mps", allow_mps_bf16=True) == (True, False)
def test_mps_stays_fp32_without_per_trainer_opt_in(self, fake_mps):
from soup_cli.utils.gpu import bf16_fp16_flags
fake_mps(available=True, accepts_bf16=True)
assert bf16_fp16_flags("mps") == (False, False)
def test_cpu_stays_fp32_even_when_mps_is_available(self, fake_mps):
from soup_cli.utils.gpu import bf16_fp16_flags
fake_mps(available=True, accepts_bf16=True)
assert bf16_fp16_flags("cpu", allow_mps_bf16=True) == (False, False)
@pytest.mark.parametrize("auto_mixed_precision", [False, True])
def test_sft_requests_the_verified_mps_policy(
self, monkeypatch, auto_mixed_precision
):
from soup_cli.trainer import sft
calls = []
def resolve(device, *, allow_mps_bf16=False):
calls.append((device, allow_mps_bf16))
return (True, False)
monkeypatch.setattr(sft, "bf16_fp16_flags", resolve)
wrapper = object.__new__(sft.SFTTrainerWrapper)
wrapper.device = "mps"
assert wrapper._resolve_mixed_precision(
SimpleNamespace(auto_mixed_precision=auto_mixed_precision), "model"
) == (True, False)
assert calls == [("mps", True)]
def test_only_hardware_validated_trainers_opt_in(self):
import soup_cli.trainer as trainer_package
root = Path(trainer_package.__file__).parent
# CSM's real MPS training proof lives in test_native_csm.py.
expected = {"sft.py", "dpo.py", "grpo.py", "reward_model.py", "prm.py", "csm.py"}
opted_in = {
path.name
for path in root.glob("*.py")
if "allow_mps_bf16=True" in path.read_text(encoding="utf-8")
}
assert opted_in == expected
def test_layer_streaming_uses_the_same_mps_capability(monkeypatch):
from soup_cli.utils import gpu
from soup_cli.utils.layer_stream import resolve_stream_dtype
monkeypatch.setattr(gpu, "mps_supports_bf16", lambda: True)
assert resolve_stream_dtype("mps") == "bfloat16"
monkeypatch.setattr(gpu, "mps_supports_bf16", lambda: False)
assert resolve_stream_dtype("mps") == "float32"