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
"""Windows charges commit for GPU buffers the process never touches (HIP and Vulkan); a load that
leaves the system short of commit says so instead of exhausting the page file silently (#16015)."""
from __future__ import annotations
import sys
import types as _types
from pathlib import Path
import pytest
_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
if _BACKEND_DIR not in sys.path:
sys.path.insert(0, _BACKEND_DIR)
_loggers_stub = _types.ModuleType("loggers")
_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
sys.modules.setdefault("loggers", _loggers_stub)
_structlog_stub = _types.ModuleType("structlog")
_structlog_stub.get_logger = lambda *a, **k: __import__("logging").getLogger("stub")
sys.modules.setdefault("structlog", _structlog_stub)
import core.inference.diffusion_memory as diffusion_memory
from core.inference import llama_cpp as llama_cpp_module
from core.inference.llama_cpp import LlamaCppBackend
_MIB = 1024 * 1024
def test_report_numbers_warn_with_page_file_advice():
# The report: ~15.9 GiB untouched commit on llama-server, 1.2 GB of commit left.
message = LlamaCppBackend._gpu_commit_headroom_message(16_265, 1_150)
assert message is not None
assert "15.9 GiB" in message
assert "1.1 GiB" in message
assert "at least 16 GiB" in message
assert "Virtual memory" in message
@pytest.mark.parametrize(
("untouched_mib", "headroom_mib"),
[
(16_265, 4_096), # enough commit left
(16_265, None), # headroom unreadable
(2_047, 0), # the child holds no meaningful untouched commit
],
)
def test_no_warning_while_there_is_room_or_nothing_to_blame(untouched_mib, headroom_mib):
assert LlamaCppBackend._gpu_commit_headroom_message(untouched_mib, headroom_mib) is None
# Reporter's DirectIO load: ~16 GiB private commit, almost none of it resident.
_REPORT_COUNTERS = (16_500 * _MIB, 235 * _MIB)
def _windows_child(
monkeypatch,
backends,
headroom_mib = 1_150,
):
monkeypatch.setattr(llama_cpp_module.sys, "platform", "win32")
monkeypatch.setattr(
LlamaCppBackend, "_installed_ggml_backends", staticmethod(lambda binary = None: backends)
)
monkeypatch.setattr(diffusion_memory, "_windows_commit_headroom_mib", lambda: headroom_mib)
monkeypatch.setattr(llama_cpp_module, "_windows_private_memory", lambda pid: _REPORT_COUNTERS)
@pytest.mark.parametrize("gpu", ["hip", "vulkan", "cuda"])
def test_gpu_child_short_of_commit_gets_the_notice(monkeypatch, gpu):
_windows_child(monkeypatch, frozenset({"base", "cpu", gpu}))
message = LlamaCppBackend._windows_gpu_commit_notice(1234, "llama-server.exe")
assert message is not None and "15.9 GiB" in message
def test_cpu_only_builds_are_not_measured(monkeypatch):
_windows_child(monkeypatch, frozenset({"base", "cpu"}))
assert LlamaCppBackend._windows_gpu_commit_notice(1234, "llama-server.exe") is None
def test_off_windows_or_without_a_child_nothing_is_measured(monkeypatch):
_windows_child(monkeypatch, frozenset({"hip"}))
assert LlamaCppBackend._windows_gpu_commit_notice(None, "llama-server.exe") is None
monkeypatch.setattr(llama_cpp_module.sys, "platform", "linux")
assert LlamaCppBackend._windows_gpu_commit_notice(1234, "llama-server.exe") is None
@pytest.mark.parametrize("counters", [None, RuntimeError("OpenProcess denied")])
def test_a_failed_probe_never_fails_the_load(monkeypatch, counters):
_windows_child(monkeypatch, frozenset({"hip"}))
def _read(pid):
if isinstance(counters, Exception):
raise counters
return counters
monkeypatch.setattr(llama_cpp_module, "_windows_private_memory", _read)
assert LlamaCppBackend._windows_gpu_commit_notice(1234, "llama-server.exe") is None