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
"""POST /unload finds a model under the ids Studio publishes for it, and logs only real unloads (#13766)."""
import asyncio
import pytest
import core.inference.model_slots as model_slots
import core.inference.orchestrator as orchestrator
import routes.inference as inf
from models.inference import UnloadRequest
SCAN_DIR = "/srv/models/org/X-GGUF"
class FakeLlama:
holds_no_vram = False
def __init__(
self,
identifier = None,
variant = None,
advertised = None,
):
self.model_identifier = identifier
self.hf_variant = variant
self._openai_advertised_id = advertised
self.is_loaded = self.is_active = identifier is not None
self.is_vision = False
self.context_length = 4096
def unload_model(self):
self.is_loaded = self.is_active = False
class FakeOrchestrator:
def __init__(
self,
active = None,
advertised = None,
):
self.active_model_name = active
self.models = {active: {}} if active else {}
self._openai_advertised_id = advertised
self.unloaded = []
def unload_model(self, name):
if name == self.active_model_name:
self.unloaded.append(name)
self.active_model_name = None
self.models = {}
return True
@pytest.fixture
def lifecycle(monkeypatch):
rows = []
monkeypatch.setattr(model_slots, "slots", [])
monkeypatch.setattr(model_slots, "stuck", [])
monkeypatch.setattr(inf, "_raise_if_sidecar_swap_in_progress", lambda: None)
monkeypatch.setattr(inf.api_monitor, "record_lifecycle", lambda **kwargs: rows.append(kwargs))
return rows
def _serve(
monkeypatch,
llama,
orch = None,
):
monkeypatch.setattr(inf, "_llama_cpp_backend", llama)
monkeypatch.setattr(orchestrator, "_inference_backend", orch or FakeOrchestrator())
def _unload(model_path):
return asyncio.run(inf._unload_model_impl(UnloadRequest(model_path = model_path), "s"))
@pytest.mark.parametrize(
"named",
[
"org/X-GGUF", # /v1/models and status.active_model
"ORG/x-gguf",
"org/X-GGUF:X-custom-NVFP4", # the API monitor's loaded-model label
SCAN_DIR, # status.model_identifier, what the Unload button sends
],
)
def test_scan_folder_gguf_unloads_under_every_published_id(monkeypatch, lifecycle, named):
llama = FakeLlama(SCAN_DIR, "X-custom-NVFP4", advertised = "org/X-GGUF")
_serve(monkeypatch, llama)
assert _unload(named).status == "unloaded"
assert not llama.is_loaded
assert [row["model"] for row in lifecycle] == ["org/X-GGUF:X-custom-NVFP4"]
@pytest.mark.parametrize("named", ["org/X-GGUF:Q8_0", "org/Other-GGUF", "X-custom-NVFP4"])
def test_a_name_for_another_model_leaves_it_loaded_and_logs_nothing(monkeypatch, lifecycle, named):
llama = FakeLlama(SCAN_DIR, "X-custom-NVFP4", advertised = "org/X-GGUF")
_serve(monkeypatch, llama)
# Still the documented idempotent no-op for a stale tab.
assert _unload(named).status == "unloaded"
assert llama.is_loaded
assert lifecycle == []
def test_unload_with_nothing_loaded_logs_nothing(monkeypatch, lifecycle):
_serve(monkeypatch, FakeLlama())
assert _unload("org/X-GGUF").status == "unloaded"
assert lifecycle == []
def test_a_load_in_flight_is_not_matched_by_a_stale_advertised_id(monkeypatch):
llama = FakeLlama("/srv/models/org/New-GGUF", "Q4_K_M", advertised = "org/Old-GGUF")
llama.is_loaded = False
assert not inf._names_the_resident_gguf(llama, "org/Old-GGUF")
assert inf._names_the_resident_gguf(llama, "/srv/models/org/New-GGUF")
def test_auto_switched_safetensors_unloads_under_its_advertised_id(monkeypatch, lifecycle):
orch = FakeOrchestrator("/srv/models/org/Y", advertised = "org/Y")
_serve(monkeypatch, FakeLlama(), orch)
assert _unload("org/Y").status == "unloaded"
assert orch.unloaded == ["/srv/models/org/Y"]
assert [row["model"] for row in lifecycle] == ["org/Y"]