Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
# SPDX-License-Identifier: AGPL-3.0-only
import importlib.util
import os
import platform
import subprocess
import sys
import time
from pathlib import Path
from types import SimpleNamespace
import pytest
_REPO_ROOT = Path(__file__).resolve().parents[2]
def _load_legacy_cli():
spec = importlib.util.spec_from_file_location(
"unsloth_legacy_cli", _REPO_ROOT / "unsloth-cli.py"
)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
def _run_collective_probe():
sys.path.insert(0, str(_REPO_ROOT / "studio" / "backend"))
import mlx.core as mx
from core.training.worker import _configure_mlx_training_schedule, _finalize_mlx_training
from unsloth_zoo.mlx.trainer import MLXTrainer
root = Path(os.environ["UNSLOTH_MLX_DDP_PROBE_DIR"])
trainer = object.__new__(MLXTrainer)
rank = trainer.distributed_rank
args = SimpleNamespace(max_steps = 0, warmup_steps = 0, eval_steps = 0)
schedule_trainer = SimpleNamespace(args = args, distributed_world_size = 2)
assert _configure_mlx_training_schedule(schedule_trainer, 0, 16, 2, 2, 3) == (6, 0)
import unsloth_cli.commands.train as train_cmd
import utils.transformers_version as transformers_version
active, contender = root / "sidecar-active", root / "sidecar-contender"
def activate_sidecar(*_args):
try:
active.touch(exist_ok = False)
if rank == 0:
deadline = time.monotonic() + 5
while not contender.exists() and time.monotonic() < deadline:
time.sleep(0.01)
assert contender.exists()
time.sleep(0.1)
finally:
active.unlink(missing_ok = True)
transformers_version.activate_transformers_for_subprocess = activate_sidecar
if rank:
deadline = time.monotonic() + 5
while not active.exists() and time.monotonic() < deadline:
time.sleep(0.01)
assert active.exists()
contender.touch()
train_cmd._activate_mlx_transformers("test/model", None)
trainer.stop_requested = False
trainer.save_model = lambda _path: (root / "worker-save").touch(exist_ok = False)
complete = {}
_finalize_mlx_training(
trainer,
lambda: (False, True),
str(root),
mx.synchronize,
lambda event_type, **payload: complete.update(type = event_type, **payload),
lambda: True,
)
assert complete["type"] == "complete"
assert complete["output_dir"] == (str(root) if rank == 0 else None)
legacy_cli = _load_legacy_cli()
legacy_cli._save_or_push_model = lambda *_args: (root / "legacy-save").touch(exist_ok = False)
legacy_cli._save_or_push_model_with_mlx_ddp(object(), object(), object(), True, trainer)
trainer.save_model = lambda _path: (_ for _ in ()).throw(KeyboardInterrupt("probe"))
with pytest.raises((KeyboardInterrupt, RuntimeError)):
_finalize_mlx_training(
trainer,
lambda: (False, True),
str(root),
mx.synchronize,
lambda *_args, **_kwargs: None,
lambda: True,
)
(root / f"passed-{rank}").touch()
@pytest.mark.skipif(
platform.system() != "Darwin" or platform.machine() != "arm64",
reason = "real MLX collectives require Apple Silicon",
)
def test_mlx_launch_cli_ddp_contracts(tmp_path):
launcher = Path(sys.executable).with_name("mlx.launch")
env = os.environ.copy()
env["UNSLOTH_MLX_DDP_PROBE_DIR"] = str(tmp_path)
env["UNSLOTH_STUDIO_HOME"] = str(tmp_path / "studio")
env["PYTHONPATH"] = os.pathsep.join((str(_REPO_ROOT), str(_REPO_ROOT / "studio/backend")))
result = subprocess.run(
[launcher, "-n", "2", "--", sys.executable, str(Path(__file__).resolve())],
cwd = _REPO_ROOT,
env = env,
capture_output = True,
timeout = 60,
)
assert result.returncode == 0, result.stdout + result.stderr
assert len(list(tmp_path.glob("passed-*"))) == 2, result.stdout + result.stderr
if __name__ == "__main__":
_run_collective_probe()