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 asyncio
import os
import pytest
from core.inference import model_identity as mi
from studio_mcp import loading, model_downloads, placement
from studio_mcp.outputs import SlotModel
AUDIO = "audio-cpp/audio.cpp-gguf/Qwen3-TTS"
NT = os.name == "nt"
@pytest.mark.parametrize(
"name, local, split",
[
("unsloth/Qwen3-GGUF:Q4_K_M", False, ("unsloth/Qwen3-GGUF", "Q4_K_M")),
("unsloth/Qwen3-GGUF:", False, ("unsloth/Qwen3-GGUF", None)),
(f"{AUDIO}:Q8_0", False, (AUDIO, "Q8_0")),
("org/repo:sub/dir", False, ("org/repo:sub/dir", None)),
("ab:x", False, ("ab", "x")),
("1:x", False, ("1", "x")),
("~/m/a.gguf", True, ("~/m/a.gguf", None)),
("C:\\models\\a.gguf", True, ("C:\\models\\a.gguf", None)),
("C:/models/a.gguf", True, ("C:/models/a.gguf", None)),
("C:x", True, ("C:x", None)),
("m\\a.gguf", True, ("m\\a.gguf", None)),
("ref:0123abcd", True, ("ref:0123abcd", None)),
],
)
def test_names(monkeypatch, name, local, split):
assert mi.is_local(name) is placement.is_local(name) is local
assert model_downloads._looks_local(name) is local
assert mi.split_variant(name) == split
if local or split[1]:
assert placement._split(name, None) == model_downloads._split(name, None) == split
asked = []
async def route_json(method, path, **kwargs):
asked.append(path)
return {}
monkeypatch.setattr(loading, "route_json", route_json)
asyncio.run(loading._llm_poll(None, name)())
# Only a Hub name has download progress to read.
assert ("/api/models/download-progress" in asked) is not local
@pytest.mark.parametrize(
"left, right, model, repo",
[
("unsloth/Qwen3-0.6B-GGUF", "UNSLOTH/qwen3-0.6b-gguf", True, True),
(" unsloth/Qwen3 ", "unsloth/qwen3", True, True),
("unsloth/Qwen3-GGUF:Q4_K_M", "unsloth/qwen3-gguf", False, True),
("unsloth/Qwen3-GGUF:q4_k_m", "unsloth/Qwen3-GGUF:Q4_K_M", True, True),
("org/repo:Q4_K_M-3.97bpw", "org/repo:Q4_K_M-4.19bpw", False, True),
("unsloth/Qwen3-GGUF", "unsloth/Gemma-GGUF", False, False),
(AUDIO, AUDIO.lower(), True, True),
("/models/Foo.gguf", "/models/foo.gguf", NT, NT),
("/models/Foo/", "/models/Foo", True, True),
("./m/Foo", "./m/foo", NT, NT),
("C:\\models\\a.gguf", "C:/models/a.gguf", True, True),
("C:\\models\\a.gguf", "C:\\models\\b.gguf", False, False),
("/models/a:b.gguf", "/models/a:c.gguf", False, False),
("ref:ABC", "ref:ABC", True, True),
("ref:ABC", "ref:abc", NT, NT),
(None, "unsloth/Qwen3", False, False),
],
)
def test_same_model_and_repo(left, right, model, repo):
for a, b in ((left, right), (right, left)):
assert mi.same_model(a, b) is mi.matches_any(a, (None, "", b)) is model
assert mi.same_repo(a, b) is repo
@pytest.mark.parametrize(
"variant, filename, settled, result",
[
("Q4_K_M", "model-Q4_K_M.gguf", True, True),
("q4_k_m", "model-Q4_K_M.gguf", False, True),
("Q8_0", "model-Q4_K_M.gguf", True, False),
# Status drops the bpw suffix: mid-load a sibling build cannot be told apart.
("IQ4_XS", "model-IQ4_XS-3.97bpw.gguf", True, True),
("IQ4_XS", "model-IQ4_XS-3.97bpw.gguf", False, False),
("Q4_K_M", "model.safetensors", True, None),
],
)
def test_resident_quant_matches(variant, filename, settled, result):
assert mi.resident_quant_matches(variant, filename, settled = settled) is result
@pytest.fixture
def models_dir(tmp_path, monkeypatch):
monkeypatch.chdir(tmp_path)
(tmp_path / "models").mkdir()
(tmp_path / "models" / "Foo.gguf").write_bytes(b"")
def _predict(
kind,
model,
resident,
mode = "always",
):
slots = [SlotModel(kind = kind, model = resident)]
args = dict(device = None, alongside = None, mode = mode, status = {}, slots = slots)
return placement.predict(kind = kind, model = model, **args)
@pytest.mark.parametrize(
"left, right",
[
("models/Foo.gguf", "models/foo.gguf"),
("unsloth/Qwen3-GGUF:Q4_K_M", "UNSLOTH/qwen3-gguf"),
("/m/Foo", "/m/foo"),
("C:\\m\\a.gguf", "C:\\m\\b.gguf"),
(AUDIO, AUDIO.upper()),
("org/repo:sub/dir", "org/repo"),
],
)
def test_load_and_plan_agree(models_dir, left, right):
assert _predict("llm", left, right)[2] is not loading._same_id(left, right)
def test_an_existing_relative_path_keeps_its_case(models_dir):
assert mi.is_local("models/Foo.gguf", probe = True) and not mi.is_local("models/Foo.gguf")
assert not mi.is_local("unsloth/Qwen3", probe = True)
assert mi.same_model("models/foo.gguf", "models/Foo.gguf", probe = True) is NT
assert mi.same_model("models/Foo.gguf", "models/foo.gguf")
others, exact, _ = _predict("stt", "models/Foo.gguf", "models/foo.gguf", mode = None)
# Windows reads them as one file; elsewhere they are two.
assert NT or ([s.model for s in others], exact) == (["models/foo.gguf"], False)
assert model_downloads.keys_for(None, "llm", "~/m/a.gguf", None) == []
def test_route_and_status_keep_their_old_rules():
import routes.inference as route
# Not vouched for as a path there: a dot-relative or forward-slash drive name.
for name in ("/m", "~/m", "m\\a", "./m", "C:/m", "o/r"):
assert route._looks_like_local_path(name) is (name in ("/m", "~/m", "m\\a")), name
# Status may spell the resident path differently from the request.
assert loading._resident_matches({"loaded": True, "repo_id": "/m/foo"}, "/m/Foo")
assert not loading._resident_matches({"loaded": True, "repo_id": "/m/bar"}, "/m/Foo")