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. See /studio/LICENSE.AGPL-3.0
from __future__ import annotations
import types
import pytest
torch = pytest.importorskip("torch")
hooks = pytest.importorskip("diffusers.hooks")
import core.inference.diffusion_memory as dm
class _Blocks(torch.nn.Module):
def __init__(self, n: int = 4) -> None:
super().__init__()
self.blocks = torch.nn.ModuleList([torch.nn.Linear(4, 4) for _ in range(n)])
def forward(self, x):
for block in self.blocks:
x = block(x)
return x
def _streamed(n: int = 4) -> torch.nn.Module:
module = _Blocks(n)
hooks.apply_group_offloading(
module,
onload_device = torch.device("cpu"),
offload_device = torch.device("cpu"),
offload_type = "block_level",
num_blocks_per_group = 1,
)
return module
def _pin(module: torch.nn.Module, count: int) -> None:
for group in dm._offload_groups(module)[:count]:
group._unsloth_resident = True
def test_an_unhooked_transformer_under_group_offload_is_resident():
pipe = types.SimpleNamespace(transformer = _Blocks())
assert dm.denoiser_offload(pipe, dm.OFFLOAD_GROUP) == dm.DENOISER_RESIDENT
def test_every_group_streamed_is_streamed():
pipe = types.SimpleNamespace(transformer = _streamed())
assert dm.denoiser_offload(pipe, dm.OFFLOAD_GROUP) == dm.DENOISER_STREAMED
def test_some_groups_pinned_is_partial():
module = _streamed()
_pin(module, 2)
pipe = types.SimpleNamespace(transformer = module)
assert dm.denoiser_residency(pipe) == (2, 4)
assert dm.denoiser_offload(pipe, dm.OFFLOAD_GROUP) == dm.DENOISER_PARTIAL
def test_every_group_pinned_is_resident():
module = _streamed()
_pin(module, 4)
pipe = types.SimpleNamespace(transformer = module)
assert dm.denoiser_offload(pipe, dm.OFFLOAD_GROUP) == dm.DENOISER_RESIDENT
def test_a_second_streamed_denoiser_counts():
first = _streamed()
_pin(first, 4)
pipe = types.SimpleNamespace(transformer = first, transformer_2 = _streamed())
assert dm.denoiser_residency(pipe) == (4, 8)
assert dm.denoiser_offload(pipe, dm.OFFLOAD_GROUP) == dm.DENOISER_PARTIAL
def test_an_unhooked_resident_denoiser_beside_a_streamed_one_is_partial():
pipe = types.SimpleNamespace(transformer = _streamed(), unconditional_transformer = _Blocks())
assert dm.denoiser_residency(pipe) == (1, 5)
assert dm.denoiser_offload(pipe, dm.OFFLOAD_GROUP) == dm.DENOISER_PARTIAL
@pytest.mark.parametrize(
"policy", [dm.OFFLOAD_NONE, dm.OFFLOAD_MODEL, dm.OFFLOAD_STREAMING, dm.OFFLOAD_SEQUENTIAL, None]
)
def test_other_policies_report_nothing(policy):
pipe = types.SimpleNamespace(transformer = _streamed())
assert dm.denoiser_offload(pipe, policy) is None
def test_an_unreadable_denoiser_reports_nothing_instead_of_raising():
pipe = types.SimpleNamespace(transformer = object())
assert dm.denoiser_offload(pipe, dm.OFFLOAD_GROUP) is None
def test_the_gallery_record_keeps_the_denoiser_placement():
from models.inference import DiffusionStatusResponse, GalleryImage
assert "denoiser_offload" in GalleryImage.model_fields
assert "denoiser_offload" in DiffusionStatusResponse.model_fields