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.
"""Tensor parallel evaluation and cached generate on a gloo CPU world vs one process, fp32.
- The HF Trainer with eval_strategy = "steps" (every step), Trainer.evaluate and Trainer.predict on a tp model
(and dp x tp after shard_model): the gather covers the dp replicas only, so eval_loss / predictions equal
one process (the tp ranks of a replica hold identical results; a world gather would count each tp times).
- Cached greedy generate after a forward wrapper is installed over the tensor parallel forward guard (what the
GRPO hidden states wrapper in models/rl.py does to the model the GRPO trainer generates with): the gathered
replica decodes with a KV cache instead of being refused mid-run."""
import multiprocessing
import os
if multiprocessing.parent_process() is not None or getattr(
multiprocessing.current_process(), "_inheriting", False
):
os.environ["CUDA_VISIBLE_DEVICES"] = "" # spawned gloo ranks: CPU only, also on GPU hosts
import sys # noqa: E402
import pytest # noqa: E402
import torch # noqa: E402
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import tempfile # noqa: E402
from test_tp_native_gloo import _free_port, _import_checkout_unsloth, _Rows, _trainer_args # noqa: E402
from test_tp_native_oracle import LLAMA, _batch, build # noqa: E402
def _case_eval(rank, world, tmp, dp, tp):
"""3 Trainer steps evaluating every step, then evaluate() and predict() on 7 rows (a dp remainder)."""
import transformers
error = _import_checkout_unsloth()
if error is not None:
return {"skip": error}
from torch.distributed.device_mesh import init_device_mesh
from unsloth.models import fsdp2_native, tp_native
from unsloth.models._utils import patch_gradient_accumulation_fix
patch_gradient_accumulation_fix(transformers.Trainer)
class SequentialTrainer(transformers.Trainer):
def _get_train_sampler(self, *args, **kwargs):
return torch.utils.data.SequentialSampler(self.train_dataset)
model = build(LLAMA)
data = _Rows(list(_batch(model, bsz = 12 * dp, seq = 9)))
held_out = _Rows(list(_batch(model, seed = 2, bsz = 7, seq = 9)))
if tp > 1:
mesh = init_device_mesh("cpu", (dp, tp), mesh_dim_names = ("dp", "tp"))
tp_native.apply_tensor_parallel(model, mesh["tp"])
if dp > 1:
fsdp2_native.shard_model(
model, fsdp2_native.FSDP2Config(param_dtype = torch.float32, mesh = mesh["dp"])
)
args = _trainer_args(
tmp,
2 * dp if tp == 1 else 2,
eval_strategy = "steps",
eval_steps = 1,
per_device_eval_batch_size = 1,
dataloader_drop_last = False,
)
trainer = SequentialTrainer(model = model, args = args, train_dataset = data, eval_dataset = held_out)
trainer.train()
history = trainer.state.log_history
final = trainer.evaluate()
predicted = trainer.predict(held_out)
return {
"losses": [h["loss"] for h in history if "loss" in h],
"eval_losses": [h["eval_loss"] for h in history if "eval_loss" in h],
"final_eval": final["eval_loss"],
"predictions": torch.as_tensor(predicted.predictions),
"label_ids": torch.as_tensor(predicted.label_ids),
"predict_loss": predicted.metrics["test_loss"],
}
def _case_generate(rank, world, tmp, tp):
"""Greedy cached generate after the GRPO hidden states forward wrapper wraps the tp forward guard."""
error = _import_checkout_unsloth()
if error is not None:
return {"skip": error}
from torch.distributed.device_mesh import init_device_mesh
from unsloth.models import rl, tp_native, tp_replica
model = build(LLAMA)
prompt = _batch(model, seed = 3, bsz = 2, seq = 5)
if tp > 1:
mesh = init_device_mesh("cpu", (tp,), mesh_dim_names = ("tp",))
tp_native.apply_tensor_parallel(model, mesh["tp"])
wrapped = rl._install_grpo_hidden_states_forward_wrapper(model)
model.eval()
try:
with torch.no_grad():
out = model.generate(
input_ids = prompt,
attention_mask = torch.ones_like(prompt),
max_new_tokens = 6,
do_sample = False,
use_cache = True,
pad_token_id = 0,
)
error = None
except Exception as e:
out, error = None, f"{type(e).__name__}: {e}"
active = tp_replica.replica_active(model) if tp > 1 else False
return {"out": out, "error": error, "wrapped": wrapped, "replica_after": active}
CASES = {"eval": _case_eval, "generate": _case_generate}
def _worker(rank, world, port, case, tmp, kwargs):
os.environ.update(
MASTER_ADDR = "127.0.0.1",
MASTER_PORT = str(port),
RANK = str(rank),
WORLD_SIZE = str(world),
LOCAL_RANK = str(rank),
LOCAL_WORLD_SIZE = str(world),
)
import datetime
import torch.distributed as dist
torch.set_num_threads(2)
dist.init_process_group(
"gloo", rank = rank, world_size = world, timeout = datetime.timedelta(seconds = 120)
)
try:
res = CASES[case](rank, world, tmp, **kwargs)
torch.save(res, os.path.join(tmp, f"result_{rank}.pt"))
finally:
dist.destroy_process_group()
def _spawn(
case,
world = 2,
**kwargs,
):
ctx = multiprocessing.get_context("spawn")
port = _free_port()
with tempfile.TemporaryDirectory() as tmp:
procs = [
ctx.Process(target = _worker, args = (r, world, port, case, tmp, kwargs))
for r in range(world)
]
for p in procs:
p.start()
for p in procs:
p.join(600)
codes = [p.exitcode for p in procs]
assert codes == [0] * world, f"rank exit codes {codes}"
return [
torch.load(os.path.join(tmp, f"result_{r}.pt"), weights_only = False)
for r in range(world)
]
def _torch_below_2_7():
from packaging.version import Version
return Version(torch.__version__.split("+")[0]) < Version("2.7")
@pytest.mark.parametrize("dp, tp", [(1, 2), (2, 2)])
def test_trainer_evaluates_a_tensor_parallel_model_like_one_process(dp, tp):
if dp > 1 and _torch_below_2_7():
pytest.skip(
"torch < 2.7: tensor parallel x FSDP2 is refused (no fully_shard(ignored_params))"
)
ref = _spawn("eval", 1, dp = dp, tp = 1)[0]
if "skip" in ref:
pytest.skip(f"import unsloth failed on this host: {ref['skip']}")
results = _spawn("eval", dp * tp, dp = dp, tp = tp)
assert (
len(ref["losses"]) == 3 and len(ref["eval_losses"]) == 4
) # 3 during training + evaluate()
assert ref["predictions"].shape[0] == 7
for rank, res in enumerate(results):
assert res["losses"] == pytest.approx(ref["losses"], rel = 1e-5), (
rank,
res["losses"],
ref["losses"],
)
assert res["eval_losses"] == pytest.approx(ref["eval_losses"], rel = 1e-5), (
rank,
res["eval_losses"],
ref["eval_losses"],
)
assert res["final_eval"] == pytest.approx(ref["final_eval"], rel = 1e-5)
assert res["predict_loss"] == pytest.approx(ref["predict_loss"], rel = 1e-5)
assert res["predictions"].shape == ref["predictions"].shape, (
rank,
res["predictions"].shape,
)
torch.testing.assert_close(res["predictions"], ref["predictions"], rtol = 1e-4, atol = 1e-5)
assert torch.equal(res["label_ids"], ref["label_ids"])
def test_cached_generate_through_a_wrapped_tp_forward_matches_one_process():
ref = _spawn("generate", 1, tp = 1)[0]
if "skip" in ref:
pytest.skip(f"import unsloth failed on this host: {ref['skip']}")
assert ref["error"] is None, ref["error"]
results = _spawn("generate", 2, tp = 2)
for rank, res in enumerate(results):
assert res[
"wrapped"
], "the GRPO hidden states wrapper was not installed: the test proves nothing"
assert res["error"] is None, (rank, res["error"])
assert torch.equal(res["out"], ref["out"]), (rank, res["out"], ref["out"])
assert res["replica_after"] is False