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++
"""Graphics card roles in setup (strata_control/gpu_roles.py): the best card is Main, the others start as Helper, a card with 8 GB or less
starts Off; one question, "Use these roles?"; --gpu-roles names them.
python tools/test_setup_gpu_fit.py
"""
from __future__ import annotations
import sys
import unittest
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))
sys.path.insert(0, str(ROOT / "tools"))
import setup # noqa: E402
from test_setup_golden import card, install # noqa: E402
def run(ram, found, model, answers=None, extra=()):
argv = ["--family", "qwen", "--no-start", "--model", model, "--context", "32768", *extra]
return install(ram, found, argv, answers=answers or {})
class SetupFitRule(unittest.TestCase):
def test_one_card_has_no_question(self):
code, out, cfg, asked = run(62.0, [card(0, "NVIDIA GeForce RTX 3070 Ti", 8.0, "86")], "IQ3_XXS")
self.assertEqual(code, 0, out)
self.assertEqual(cfg["gpu"], 0) if "gpu" in cfg else None
self.assertFalse(any("Which GPUs" in q for q in asked), asked)
def test_two_unequal_cards_the_big_one_is_main_and_a_card_with_8_gb_or_less_starts_off(self):
found = [card(0, "NVIDIA GeForce RTX 3060 Ti", 8.0, "86"), card(1, "NVIDIA GeForce RTX 3070 Ti", 16.0, "86")]
code, out, cfg, asked = run(62.0, found, "IQ3_XXS")
self.assertEqual(code, 0, out)
self.assertEqual(cfg["gpu"], 1) # the 16 GB card, alone
self.assertEqual(cfg["gpu_roles"], {"0": "off", "1": "main"})
self.assertNotIn("layer_split", cfg)
self.assertTrue(any("Use these roles?" in q and "[y]" in q for q in asked), asked)
# one answer away: the 8 GB card as a Helper
code, out, cfg, _ = run(62.0, found, "IQ3_XXS", answers={"Use these roles?": "n", "Roles, e.g.": "1=main,0=helper"})
self.assertEqual(code, 0, out)
self.assertEqual(cfg["gpu"], [0, 1]) # Main LAST: the engine's stage order
self.assertEqual(cfg["gpu_order"], "as_given")
self.assertEqual(cfg["gpu_roles"], {"0": "helper", "1": "main"})
self.assertEqual(cfg["layer_split"], "auto")
def test_two_equal_cards_start_with_the_highest_number_as_main(self):
found = [card(0, "NVIDIA GeForce RTX 3060", 12.0, "86"), card(1, "NVIDIA GeForce RTX 3060", 12.0, "86")]
code, out, cfg, asked = run(26.0, found, "IQ3_XXS")
self.assertEqual(code, 0, out)
self.assertEqual(cfg["gpu"], [0, 1])
self.assertEqual(cfg["gpu_roles"], {"0": "helper", "1": "main"})
self.assertEqual(cfg["gpu_order"], "as_given")
self.assertEqual(cfg["layer_split"], "auto")
def test_gpu_roles_flag(self):
found = [card(0, "NVIDIA GeForce RTX 3060", 12.0, "86"), card(1, "NVIDIA GeForce RTX 3070 Ti", 16.0, "86")]
code, out, cfg, _ = run(62.0, found, "IQ3_XXS", extra=("--gpu-roles", "0=main,1=helper"))
self.assertEqual(code, 0, out)
self.assertEqual(cfg["gpu"], [1, 0]) # the slower card is Main on purpose: it goes last
self.assertEqual(cfg["gpu_roles"], {"0": "main", "1": "helper"})
code, out, cfg, _ = run(62.0, found, "IQ3_XXS", extra=("--gpu-roles", "1=main,0=off"))
self.assertEqual((cfg["gpu"], cfg["gpu_roles"]), (1, {"0": "off", "1": "main"}))
self.assertNotIn("layer_split", cfg)
code, out, cfg, _ = run(62.0, found, "IQ3_XXS", extra=("--gpu-roles", "0=off,1=off"))
self.assertNotEqual(code, 0)
def test_naming_the_cards_still_wins(self):
found = [card(0, "NVIDIA GeForce RTX 3070 Ti", 16.0, "86"), card(1, "NVIDIA GeForce RTX 3060 Ti", 8.0, "86")]
code, out, cfg, _ = run(62.0, found, "IQ3_XXS", extra=("--gpus", "0,1"))
self.assertEqual(code, 0, out)
self.assertEqual(sorted(cfg["gpu"]), [0, 1])
code, out, cfg, _ = run(62.0, found, "IQ3_XXS", extra=("--gpu", "1"))
self.assertEqual(cfg["gpu"], 1)
def test_a_bigger_amd_card_beside_the_nvidia_one_is_said_in_one_line_and_never_mixed(self):
"""RTX 5060 Ti (16 GB) + AMD R9700 (32 GB): the question offers the two makes apart, says one engine drives one make and that two
models at once are not supported, and names the bigger AMD card; the NVIDIA answer runs on the NVIDIA card alone."""
from unittest import mock
found = [card(0, "NVIDIA GeForce RTX 5060 Ti", 16.0, "120")]
amd = [{"index": 0, "name": "AMD Radeon AI PRO R9700", "vram_gb": 32.0, "arch": "gfx1201", "vendor": "amd"}]
patches = [mock.patch.object(setup, "amd_gpus", lambda *a, **k: [dict(g) for g in amd]),
mock.patch.object(setup, "amd_problem", lambda g: None)]
argv = ["--family", "qwen", "--no-start", "--model", "IQ3_XXS", "--context", "32768"]
code, out, cfg, asked = install(62.0, found, argv, answers={"Which cards?": "1"}, extra=patches)
self.assertEqual(code, 0, out)
self.assertIn("This PC has NVIDIA and AMD cards Strata can use", out)
self.assertIn("never split across NVIDIA and AMD cards", out)
self.assertIn("two models at the same", out)
self.assertIn("not supported yet", out)
self.assertIn("This computer also has an AMD AI PRO R9700 (32 GB). The AMD build of Strata can use it", out)
self.assertNotIn("Which GPUs?", " ".join(asked)) # one NVIDIA card: no split question
self.assertEqual(cfg.get("gpu", 0), 0)
if __name__ == "__main__":
unittest.main()