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.
"""apply_tensor_parallel on a model whose Linear layers train in fp8 (torchao Float8Linear from
float8_training, or fp8 LoRA frozen bases from enable_fp8_lora), gloo world of 2 on the CPU. Tensor parallel
runs every projection through its own local-shard GEMM, so the fp8 path would never run (the model would
silently train in 16 bit) and per-shard scales would differ from one process: it is refused before any weight
is sharded, and the fp8 layers are left as they were."""
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 tempfile # noqa: E402
import pytest # noqa: E402
import torch # noqa: E402
pytest.importorskip("torchao.float8")
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from test_tp_native_gloo import _free_port, _import_checkout_unsloth # noqa: E402
def _model(kind):
from transformers import Qwen3Config, Qwen3ForCausalLM
torch.manual_seed(0)
model = Qwen3ForCausalLM(
Qwen3Config(
vocab_size = 256,
hidden_size = 64,
intermediate_size = 128,
num_hidden_layers = 2,
num_attention_heads = 4,
num_key_value_heads = 2,
head_dim = 16,
tie_word_embeddings = False,
)
).to(torch.bfloat16)
if kind in ("float8_training", "float8_subclass"):
import dataclasses
from torchao.float8 import Float8LinearConfig
from torchao.float8.float8_linear import Float8Linear
from unsloth.models.float8_training import convert_model_to_float8_training
config = dataclasses.replace(Float8LinearConfig.from_recipe_name("rowwise"), emulate = True)
assert convert_model_to_float8_training(model, config, size_filter = False)
if kind == "float8_subclass":
# A renamed subclass still runs torchao's fp8 forward.
renamed = type("MyFp8Linear", (Float8Linear,), {})
for module in model.modules():
if type(module) is Float8Linear:
module.__class__ = renamed
return model
from peft import LoraConfig, get_peft_model
from unsloth.models.float8_training import enable_fp8_lora
model = get_peft_model(
model,
LoraConfig(
r = 8, lora_alpha = 16, target_modules = ["q_proj", "v_proj", "gate_proj", "down_proj"]
),
)
assert enable_fp8_lora(model, "fp8_lora_tensorwise", allow_cpu = True)
return model
def _worker(rank, world, port, tmp, kind):
import datetime
os.environ.update(
MASTER_ADDR = "127.0.0.1",
MASTER_PORT = str(port),
RANK = str(rank),
WORLD_SIZE = str(world),
LOCAL_RANK = str(rank),
UNSLOTH_FP8_LORA_FUSED_OPS = "0",
)
torch.set_num_threads(2)
error = _import_checkout_unsloth()
if error is not None:
torch.save({"skip": error}, os.path.join(tmp, f"result_{rank}.pt"))
return
import torch.distributed as dist
from torch.distributed.device_mesh import init_device_mesh
from unsloth.models import tp_native as tpn
dist.init_process_group(
"gloo", rank = rank, world_size = world, timeout = datetime.timedelta(seconds = 300)
)
try:
model = _model(kind)
before = {n: (tuple(p.shape), p.detach().clone()) for n, p in model.named_parameters()}
refused = None
try:
tpn.apply_tensor_parallel(
model, init_device_mesh("cpu", (world,), mesh_dim_names = ("tp",))["tp"]
)
except ValueError as e:
refused = str(e)
after = dict(model.named_parameters())
res = {
"refused": refused,
"dtensors": sum(tpn._is_dtensor(p) for p in model.parameters()),
"changed": [
n
for n, (shape, value) in before.items()
if n not in after
or tuple(after[n].shape) != shape
or tpn._is_dtensor(after[n])
or not torch.equal(after[n].detach(), value)
],
"float8_linears": sum(
any(c.__name__ == "Float8Linear" for c in type(m).__mro__) for m in model.modules()
),
"fp8_lora_bases": sum("_unsloth_fp8_lora" in m.__dict__ for m in model.modules()),
"is_tp": tpn.is_tensor_parallel(model),
}
torch.save(res, os.path.join(tmp, f"result_{rank}.pt"))
finally:
dist.destroy_process_group()
@pytest.mark.parametrize("kind", ["float8_training", "float8_subclass", "fp8_lora"])
def test_tp2_low_precision_model_is_refused_before_sharding(kind):
with tempfile.TemporaryDirectory() as tmp:
ctx = multiprocessing.get_context("spawn")
port = _free_port()
procs = [ctx.Process(target = _worker, args = (r, 2, port, tmp, kind)) for r in range(2)]
for p in procs:
p.start()
for p in procs:
p.join(600)
assert [p.exitcode for p in procs] == [0, 0]
results = [
torch.load(os.path.join(tmp, f"result_{r}.pt"), weights_only = False) for r in range(2)
]
if "skip" in results[0]:
pytest.skip(f"import unsloth failed on this host: {results[0]['skip']}")
for res in results:
assert res["refused"] is not None and "fp8" in res["refused"], res
assert res["dtensors"] == 0 and res["changed"] == [] and not res["is_tp"], res
assert res["float8_linears"] + res["fp8_lora_bases"] > 0, res