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
"""CPU tests of determinism.main() through the real measurement path."""
import sys
from pathlib import Path
from types import SimpleNamespace
import pytest
torch = pytest.importorskip("torch")
HARNESS = Path(__file__).parents[1] / "benchmarks" / "harness"
sys.path.insert(0, str(HARNESS))
import bitexact # noqa: E402
import mechanism_cost # noqa: E402
from tests.test_issue379_determinism_measure import VOCAB, _Out, _Toy, module # noqa: E402
_REAL_GENERATOR = torch.Generator
_REAL_RANDINT = torch.randint
class _CpuGenerator(_REAL_GENERATOR):
def __new__(cls, device="cpu"):
return _REAL_GENERATOR.__new__(cls)
class _NanLoss(_Toy):
def forward(self, input_ids, labels):
out = super().forward(input_ids, labels)
return _Out(out.logits, out.loss * float("nan"))
class _Runtime:
def close(self):
pass
def _run(monkeypatch, streamed, resident=None):
resident = resident if resident is not None else _Toy(defect=False)
streamed.config = resident.config = SimpleNamespace(vocab_size=VOCAB)
seen = {"arms": [], "restored": 0}
monkeypatch.setattr(torch, "Generator", _CpuGenerator)
monkeypatch.setattr(
torch,
"randint",
lambda *args, **kwargs: _REAL_RANDINT(
*args, **{key: value for key, value in kwargs.items() if key != "device"}
),
)
monkeypatch.setattr(
mechanism_cost, "prepare_shards", lambda weights, shards: ({}, "llama")
)
def build(weights, shards, index, arch, arm, buffers):
seen["arms"].append(arm)
return streamed, _Runtime(), lambda: seen.__setitem__(
"restored", seen["restored"] + 1
)
monkeypatch.setattr(mechanism_cost, "build_streamed_arm", build)
monkeypatch.setattr(bitexact, "load_resident_reference", lambda weights, quant: resident)
import soup_cli.utils.spectrum_scan as scan
monkeypatch.setattr(scan, "resolve_model_weights", lambda value: value)
monkeypatch.setattr(
sys,
"argv",
["determinism.py", "--weights", "w", "--shards", "s", "--seq", "8"],
)
calls = {"n": 0}
def available_once():
calls["n"] += 1
return calls["n"] == 1
monkeypatch.setattr(torch.cuda, "is_available", available_once)
return module.main(), seen
def test_record_pattern_is_reproduced_with_the_historical_control(monkeypatch):
code, seen = _run(monkeypatch, _Toy(defect=True))
assert code == 0
assert seen["arms"] == ["control"]
assert seen["restored"] == 1
def test_healthy_pair_is_not_reproduced(monkeypatch):
code, seen = _run(monkeypatch, _Toy(defect=False))
assert code == 1
assert seen["restored"] == 1
def test_nan_loss_is_an_invalid_measurement(monkeypatch):
code, seen = _run(monkeypatch, _NanLoss(defect=True))
assert code == 3
assert seen["restored"] == 1