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
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
"""native_save_optimizer (gloo CPU world of 2): only rank 0 writes the full optimizer state, and the
other rank must not leave before that write finishes, or its next collective (the next step's
all-gather) waits on rank 0 for as long as a large torch.save takes and can hit the collective
timeout. Rank 0's write is slowed down; rank 1 must find the file complete when the call returns."""
import datetime
import os
import sys
import time
import pytest
import torch
if __import__("multiprocessing").parent_process() is not None or getattr(
__import__("multiprocessing").current_process(), "_inheriting", False
):
os.environ["CUDA_VISIBLE_DEVICES"] = ""
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from test_fsdp2_native_units import _load_native # noqa: E402
from test_fsdp2_native_guards_gloo import _tiny_llama # noqa: E402
DELAY_S = 3.0
def _worker(rank, world, init_file, path, out):
os.environ.update(RANK = str(rank), WORLD_SIZE = str(world), LOCAL_RANK = str(rank))
import torch.distributed as dist
dist.init_process_group(
"gloo",
init_method = f"file://{init_file}",
rank = rank,
world_size = world,
timeout = datetime.timedelta(seconds = 60),
)
try:
native = _load_native()
model = native.shard_model(_tiny_llama(), native.FSDP2Config(param_dtype = torch.float32))
opt = torch.optim.AdamW(model.parameters(), lr = 1e-3, foreach = False)
model(torch.tensor([[1, 2, 3]])).logits.float().mean().backward()
opt.step()
if rank == 0:
save = torch.save
def slow_save(*args, **kwargs):
time.sleep(DELAY_S) # a large state still being written
return save(*args, **kwargs)
torch.save = slow_save
native.native_save_optimizer(model, opt, path, should_save = rank == 0)
out.put((rank, {"file_complete_on_return": os.path.exists(path)}))
except Exception as exc:
import traceback
out.put((rank, {"crash": f"{type(exc).__name__}: {exc}\n{traceback.format_exc()}"}))
finally:
try:
dist.destroy_process_group()
except Exception:
pass
def test_non_writers_wait_for_the_optimizer_write(tmp_path):
pytest.importorskip("transformers")
import torch.multiprocessing as mp
ctx = mp.get_context("spawn")
out = ctx.Queue()
args = (2, str(tmp_path / "store"), str(tmp_path / "optimizer.pt"), out)
procs = mp.start_processes(_worker, args = args, nprocs = 2, join = False, start_method = "spawn")
deadline = time.time() + 180
while not procs.join(timeout = 5):
if time.time() > deadline:
for p in procs.processes:
p.kill()
pytest.fail("the ranks did not finish within 180 s")
results = {}
while not out.empty():
rank, r = out.get(timeout = 5)
results[rank] = r
assert sorted(results) == [0, 1], results
for rank, r in results.items():
assert "crash" not in r, r["crash"]
assert r["file_complete_on_return"], (rank, results)