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
import types
import pytest
import utils.hardware as hardware
from core.inference import diffusion_engine_router
from core.inference import flux3_prompt_upsampler as pup
from core.inference.sd_cpp_backend import _SdState
def _native_state(**overrides):
fields = dict(
repo_id = "unsloth/FLUX.3-klein-GGUF",
base_repo = "unsloth/FLUX.3-klein",
family = types.SimpleNamespace(name = "flux3-klein"),
device = "cuda",
files = types.SimpleNamespace(),
)
fields.update(overrides)
return _SdState(**fields)
def _activate(monkeypatch, state, visible):
monkeypatch.setattr(
diffusion_engine_router,
"get_active_diffusion_engine",
lambda: types.SimpleNamespace(_state = state),
)
monkeypatch.setattr(hardware, "get_parent_visible_gpu_ids", lambda: list(visible))
@pytest.mark.parametrize(
"visible, physical, expected",
[([0, 1], 1, 1), ([0, 1], 0, 0), ([3, 5], 5, 1), ([3, 5], 4, None)],
)
def test_upsampler_reads_the_native_sd_cpp_card(monkeypatch, visible, physical, expected):
_activate(monkeypatch, _native_state(physical_gpu_id = physical), visible)
assert pup._resident_cuda_ordinal() == expected
def test_unresolved_or_off_torch_native_placement_stays_unplaced(monkeypatch):
_activate(monkeypatch, _native_state(physical_gpu_id = None), [0, 1])
assert pup._resident_cuda_ordinal() is None
_activate(monkeypatch, _native_state(physical_gpu_id = 1, off_torch_device = "Vulkan1"), [0, 1])
assert pup._resident_cuda_ordinal() is None
def test_llama_server_is_masked_to_the_native_card(monkeypatch):
torch = pytest.importorskip("torch")
_activate(monkeypatch, _native_state(physical_gpu_id = 5), [3, 5])
monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "3,5")
monkeypatch.delenv("HIP_VISIBLE_DEVICES", raising = False)
monkeypatch.delenv("LLAMA_ARG_DEVICE", raising = False)
monkeypatch.setattr(torch.cuda, "device_count", lambda: 2)
from core.inference.llama_cpp import LlamaCppBackend
monkeypatch.setattr(LlamaCppBackend, "_torch_is_rocm", staticmethod(lambda _torch: False))
assert pup._child_device_mask(pup._resident_cuda_ordinal()) == {"CUDA_VISIBLE_DEVICES": "5"}