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.
from __future__ import annotations
import pytest
import torch
from torch import nn
import unsloth # noqa: F401
from unsloth.kernels import multi_lora as ml
def _model(proj: nn.Module) -> nn.Module:
model = nn.Module()
model.layer = nn.Module()
model.layer.proj = proj
model.layer.dense = nn.Linear(16, 12, bias = False)
return model
def _assert_refused_untouched(model: nn.Module, proj: nn.Module) -> None:
with pytest.raises(NotImplementedError, match = r"layer\.proj.*4-bit or in 16-bit"):
ml.inject_multi_lora(model, ["proj", "dense"], num_adapters = 2, r = 2)
assert model.layer.proj is proj
assert isinstance(model.layer.dense, nn.Linear)
assert model.layer.dense.weight.requires_grad
def test_bitsandbytes_linear8bit_base_is_refused():
bnb = pytest.importorskip("bitsandbytes")
proj = bnb.nn.Linear8bitLt(16, 12, bias = False, has_fp16_weights = False)
_assert_refused_untouched(_model(proj), proj)
with pytest.raises(NotImplementedError, match = "Linear8bitLt"):
ml.MultiLoraLinear(proj, 2, r = 2)
def test_a_weight_carrying_int8_scales_is_refused():
proj = nn.Linear(16, 12, bias = False)
proj.weight.SCB = torch.ones(12)
_assert_refused_untouched(_model(proj), proj)
def test_a_dense_base_is_still_wrapped():
model = _model(nn.Linear(16, 12, bias = False))
assert ml.inject_multi_lora(model, ["proj", "dense"], num_adapters = 2, r = 2) == [
"layer.proj",
"layer.dense",
]
assert isinstance(model.layer.proj, ml.MultiLoraLinear)