Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
"""Step 1's PCIe line: the generation card and board both run, the lanes, and a warning when the card runs on fewer
lanes than it has. No GPU needed: nvidia-smi's answer is given.
python -m unittest tools.test_setup_pcie
"""
from __future__ import annotations
import sys
import unittest
from pathlib import Path
from unittest import mock
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))
import setup # noqa: E402
class Link(unittest.TestCase):
def test_parse(self):
with mock.patch.object(setup, "out", return_value="3, 4, 3, 8, 8\n"):
self.assertEqual(setup.pcie_link(0), {"gen": 3, "gpu_gen": 4, "host_gen": 3, "width": 8, "max_width": 8})
for bad in ("", "[N/A], 4, 3, 8, 8\n", "garbage"):
with mock.patch.object(setup, "out", return_value=bad):
self.assertIsNone(setup.pcie_link(0), bad)
def test_board_limit_and_slow_link(self): # llmserver: RTX 4060 Ti (4.0 x8) on an X99 board (3.0)
line, warn = setup.pcie_lines({"gen": 3, "gpu_gen": 4, "host_gen": 3, "width": 8, "max_width": 8})
self.assertIn("PCIe: 3.0 x8 (the card supports 4.0, the board 3.0)", line)
self.assertIn("~8 GB/s", line)
self.assertIsNone(warn)
def test_fast_link_says_nothing_more(self):
line, warn = setup.pcie_lines({"gen": 4, "gpu_gen": 4, "host_gen": 5, "width": 16, "max_width": 16})
self.assertEqual(line, "PCIe: 4.0 x16")
self.assertIsNone(warn)
def test_fewer_lanes_warns(self):
_, warn = setup.pcie_lines({"gen": 3, "gpu_gen": 3, "host_gen": 3, "width": 4, "max_width": 16})
self.assertIn("reads 4 of its 16 lanes right now", warn)
self.assertIn("narrow it when idle", warn) # the width is read at idle: a hint, not a verdict
if __name__ == "__main__":
unittest.main()