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
"""The Live monitor polls /api/system every 1-10 s; GPU readings follow it, never faster than 1 s."""
import types
def test_requested_gpu_age_is_clamped_to_one_to_ten_seconds():
import main
assert main._system_gpu_max_age_s(None) == main._SYSTEM_GPU_CACHE_TTL_SECONDS
assert main._system_gpu_max_age_s(1) == 1.0
assert main._system_gpu_max_age_s(0.1) == 1.0
assert main._system_gpu_max_age_s(60) == main._SYSTEM_GPU_CACHE_TTL_SECONDS
assert main._system_gpu_max_age_s(float("nan")) == main._SYSTEM_GPU_CACHE_TTL_SECONDS
def test_route_passes_the_age_only_when_asked(monkeypatch):
import main
seen = []
def fake(logger, **kwargs):
seen.append(kwargs)
return {"available": False}, {"available": False}
monkeypatch.setattr(main, "_get_cached_system_gpu_info", fake)
main.get_system_info(current_subject = "test")
main.get_system_info(current_subject = "test", gpu_max_age = 1)
assert "max_age_s" not in seen[0]
assert seen[1]["max_age_s"] == 1.0
def test_dense_quant_reprobe_is_throttled(monkeypatch):
import sys
import main
calls = []
monkeypatch.setitem(sys.modules, "torch", sys.modules.get("torch") or types.ModuleType("torch"))
monkeypatch.setitem(sys.modules, "torchao", types.ModuleType("torchao"))
monkeypatch.setattr(main, "_refresh_dense_quant_capability", lambda: calls.append(1) or True)
monkeypatch.setattr(main, "_dense_quant_capability", True)
monkeypatch.setattr(main, "_dense_quant_probed_at", None)
clock = iter([100.0, 101.0, 104.9, 105.1])
monkeypatch.setattr(main.time, "monotonic", lambda: next(clock))
for _ in range(4):
main._dense_quant_supported()
# Probed at 100, skipped at 101 and 104.9, probed again at 105.1.
assert len(calls) == 2
def test_process_tree_children_are_reused_within_the_ttl(monkeypatch):
import psutil
import main
scans = []
me = psutil.Process()
monkeypatch.setattr(main, "_descendants_cache", {})
monkeypatch.setattr(type(me), "children", lambda self, recursive = False: scans.append(1) or [])
main._cached_descendants(me)
main._cached_descendants(me)
assert len(scans) == 1
def test_llama_pin_support_is_not_rescanned_on_every_fast_read(monkeypatch):
import logging
import main
from core.inference.llama_cpp import LlamaCppBackend
scans = []
monkeypatch.setattr(main, "_gguf_pin_support_cache", None)
monkeypatch.setattr(
LlamaCppBackend, "_is_vulkan_backend", staticmethod(lambda: scans.append(1) or False)
)
monkeypatch.setattr(LlamaCppBackend, "_backend_lacks_gpu_lib", staticmethod(lambda: False))
log = logging.getLogger("t")
main._gguf_pin_support(log, refresh = True)
main._gguf_pin_support(log)
main._gguf_pin_support(log)
assert len(scans) == 1
# A full (past-the-window) read re-resolves, so a reinstall shows within 10 s.
main._gguf_pin_support(log, refresh = True)
assert len(scans) == 2
def test_a_slow_gpu_probe_spaces_out_fast_polls(monkeypatch):
import logging
import main
cached = ({"available": True, "devices": []}, {"available": True, "devices": []})
now = main.time.monotonic()
monkeypatch.setattr(main, "_system_gpu_cache", (now - 2.0, cached))
log = logging.getLogger("t")
# A 1.5 s probe (Windows ROCm) keeps a 2 s old reading for a 1 s poll.
monkeypatch.setattr(main, "_system_gpu_probe_s", 1.5)
assert main._get_cached_system_gpu_info(log, max_age_s = 1.0) is cached
# A fast probe lets the same poll through to a fresh read.
monkeypatch.setattr(main, "_system_gpu_probe_s", 0.05)
monkeypatch.setattr(main, "_system_gpu_cache", (now - 2.0, cached))
assert main._get_cached_system_gpu_info(log, max_age_s = 1.0) is not cached