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++
"""serve/server.py: the run settings' "gpu_roles" (Main / Helper / Off) become the engine's card list, Main LAST, Off left out."""
from __future__ import annotations
import sys
import unittest
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from serve.server import card_roles_lines, child_env, gpu_list, ordered_gpus # noqa: E402
class Roles(unittest.TestCase):
def test_roles_alone_give_the_order_main_last_and_off_left_out(self):
cfg = {"gpu_roles": {"0": "helper", "1": "main", "2": "off"}, "args": []}
self.assertEqual(gpu_list(cfg), [0, 1])
self.assertEqual(gpu_list({"gpu_roles": {"0": "main", "1": "helper"}, "args": []}), [1, 0])
env = child_env({"gpu_roles": {"0": "main", "1": "helper", "2": "off"}, "args": []})
self.assertEqual(env["CUDA_VISIBLE_DEVICES"], "1,0")
def test_the_mirror_is_used_and_never_re_sorted_by_speed(self):
cfg = {"gpu_roles": {"0": "main", "1": "helper"}, "gpu": [1, 0], "gpu_order": "as_given", "args": []}
self.assertEqual(ordered_gpus(cfg, {0: 1.0, 1: 9.0}), [1, 0]) # the faster card is Helper on purpose
cfg2 = {"gpu_roles": {"0": "main", "1": "helper"}, "gpu": [1, 0], "args": []} # even without gpu_order
self.assertEqual(ordered_gpus(cfg2, {0: 1.0, 1: 9.0}), [1, 0])
def test_one_card_on(self):
cfg = {"gpu_roles": {"0": "off", "1": "main"}, "gpu": 1, "gpu_order": "as_given", "args": []}
self.assertEqual(gpu_list(cfg), [1])
self.assertEqual(child_env(cfg)["CUDA_VISIBLE_DEVICES"], "1")
def test_a_hand_edited_gpu_wins(self):
cfg = {"gpu_roles": {"0": "helper", "1": "main"}, "gpu": [1, 0], "gpu_order": "as_given", "args": []}
self.assertEqual(gpu_list(cfg), [1, 0])
self.assertTrue(any("edited by hand" in l for l in card_roles_lines(cfg)))
self.assertEqual(ordered_gpus({**cfg, "gpu_order": None}, {0: 1.0, 1: 9.0}), [0, 1]) # the legacy rule again
def test_bad_roles_are_ignored(self):
for bad in ({"0": "boss"}, {"0": "main", "1": "main"}, "x", {}):
self.assertEqual(gpu_list({"gpu_roles": bad, "gpu": 1, "args": []}), [1])
def test_legacy_settings_are_untouched(self):
self.assertEqual(gpu_list({"gpu": "0,2", "args": []}), [0, 2])
self.assertEqual(gpu_list({"args": []}), [])
self.assertEqual(card_roles_lines({"gpu": [1, 0], "args": []}), [])
def test_the_log_line_names_every_role(self):
lines = card_roles_lines({"gpu_roles": {"0": "helper", "1": "main", "2": "off"}, "gpu": [0, 1], "args": []})
self.assertEqual(lines, ["[strata] card roles: Main GPU 1, Helper GPU 0, Off GPU 2 (numbers as nvidia-smi gives them)"])
def test_every_card_off(self):
cfg = {"gpu_roles": {"0": "off", "1": "off"}, "gpu": [], "gpu_none": True, "args": []}
self.assertEqual(gpu_list(cfg), [])
env = child_env(cfg)
self.assertEqual((env["CUDA_VISIBLE_DEVICES"], env["HIP_VISIBLE_DEVICES"]), ("", ""))
with self.assertRaises(SystemExit) as e:
card_roles_lines(cfg) # a CUDA engine cannot start without a card: said plainly
self.assertIn("Turn a card on", str(e.exception))
cpu = card_roles_lines({**cfg, "backend": "cpu"})
self.assertIn("Performance is degraded on CPU only.", cpu[0])
if __name__ == "__main__":
unittest.main()