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 long-chat gate's WebKit leg: verdicts, known ceilings and rechecks (CI wiring: test_webkit_leg_ci.py)."""
import sys
from pathlib import Path
import pytest
sys.path.insert(0, str(Path(__file__).resolve().parent))
import _area_matrix as am # noqa: E402
import _webkit_leg as wk # noqa: E402
REPO = Path(__file__).resolve().parents[2]
def _cell(
open_ms,
close_ms = 50.0,
frame = 40.0,
):
return {
"status": "OK",
"reps": 3,
"open_ms": open_ms,
"open_frame": frame,
"close_ms": close_ms,
"close_frame": frame,
}
def _result(
short,
long_,
engine = "webkit_mac",
mode = "default",
surface = "browser_pane",
):
return {
(engine, mode, wk.SHORT): {"cells": {surface: short}},
(engine, mode, wk.LONG): {"cells": {surface: long_}},
}
@pytest.mark.parametrize(
"short, long_, fails",
[
(150, 500, True), # Round 11's macOS WebKit pane: 144-165 -> 460-562 ms
(150, 180, False),
(60, 190, False), # over 2.5x but under 200 ms
(300, 700, False), # under 2.5x
],
)
def test_open_time_ratio(short, long_, fails):
rows = [
v for v in wk.judge(_result(_cell(short), _cell(long_)), known = {}) if v["metric"] == "open"
]
assert [v["fail"] for v in rows] == [fails]
if fails:
known = {"webkit_mac:default:browser_pane:open": 4.0}
assert not [
v for v in wk.judge(_result(_cell(short), _cell(long_)), known = known) if v["fail"]
]
def test_long_frame_fails_and_can_be_pinned():
res = _result(_cell(100, frame = 30), _cell(120, frame = 300))
assert [v["metric"] for v in wk.judge(res, known = {}) if v["fail"]] == [
"open_frame",
"close_frame",
]
known = {
"webkit_mac:default:browser_pane:open_frame": 350,
"webkit_mac:default:browser_pane:close_frame": 350,
}
assert not [v for v in wk.judge(res, known = known) if v["fail"]]
def test_a_surface_that_stops_running_in_the_long_chat_fails():
res = _result(_cell(100), {"status": "NOT_RUN: never reached sel"})
assert [v["metric"] for v in wk.judge(res, known = {}) if v["fail"]] == ["run"]
def test_a_long_page_that_does_not_load_fails():
res = {("webkit_mac", "default", wk.LONG): {"error": "did not load"}}
assert [v["metric"] for v in wk.judge(res, known = {}) if v["fail"]] == ["load"]
def test_a_short_page_that_does_not_load_fails():
res = {
("webkit_mac", "default", wk.SHORT): {"error": "did not load"},
("webkit_mac", "default", wk.LONG): {"cells": {"browser_pane": _cell(10_000)}},
}
assert [v["metric"] for v in wk.judge(res, known = {}) if v["fail"]] == ["load"]
def test_a_surface_without_a_short_baseline_fails():
res = _result({"status": "NOT_RUN: never reached sel"}, _cell(10_000))
assert [v["metric"] for v in wk.judge(res, known = {}) if v["fail"]] == ["run"]
def test_a_surface_that_runs_at_neither_length_fails():
res = _result({"status": "NOT_RUN: a"}, {"status": "NOT_RUN: b"})
assert [v["metric"] for v in wk.judge(res, known = {}) if v["fail"]] == ["run"]
def test_a_crashed_engine_worker_fails_every_planned_page(monkeypatch):
def boom(*args, **kwargs):
raise RuntimeError("webkit did not launch")
class _Server:
def shutdown(self):
pass
monkeypatch.setattr(wk, "_engine_worker", boom)
monkeypatch.setattr(wk.am, "serve_app", lambda dist: (_Server(), "http://x"))
res = wk.measure(Path("."), modes = ["default"], engines = ["webkit_mac"], log = lambda *a, **k: None)
assert set(res) == {("webkit_mac", "default", wk.SHORT), ("webkit_mac", "default", wk.LONG)}
assert all("webkit did not launch" in page["error"] for page in res.values())
assert [v["metric"] for v in wk.judge(res, known = {}) if v["fail"]] == ["load", "load"]
def test_short_floor_keeps_instant_surfaces_from_inflating_the_ratio():
v = [v for v in wk.judge(_result(_cell(5), _cell(210)), known = {}) if v["metric"] == "open"][0]
assert v["ratio"] == pytest.approx(210 / wk.SHORT_FLOOR_MS)
def test_a_recheck_only_lowers_readings():
first = _result(_cell(150), _cell(600, frame = 300))
again = _result(_cell(140), _cell(400, frame = 320))
merged = wk.lower(first, again)
cell = merged[("webkit_mac", "default", wk.LONG)]["cells"]["browser_pane"]
assert cell["open_ms"] == 400 and cell["open_frame"] == 300
def test_every_surface_is_a_matrix_area_with_a_timing_predicate():
for name in wk.SURFACES:
spec = am.AREAS[name]
if spec["open"][0] != "scroll":
wk.pred_js(spec["ready"], 0, "x")
if spec.get("close"):
wk.pred_js(spec["gone"], 1, "x")
def test_known_ceilings_name_their_fix():
text = (Path(wk.__file__)).read_text(encoding = "utf-8")
block = text[text.index("KNOWN_DEFAULT = {") : text.index("KNOWN = {")]
for key in wk.KNOWN_DEFAULT:
engine, mode, surface, metric = key.split(":")
assert engine in wk.ENGINES and mode in wk.MODES and surface in wk.SURFACES
assert metric in ("open", "close", "open_frame", "close_frame")
if wk.KNOWN_DEFAULT:
assert "Fix work:" in block
def test_a_frame_that_is_already_long_in_the_short_chat_is_a_mount_cost_not_a_long_chat_one():
# WebKit on a CI runner draws a 300-600 ms frame mounting Settings at 1 turn.
res = _result(_cell(600, frame = 400), _cell(620, frame = 450))
assert not [v for v in wk.judge(res, known = {}) if v["fail"]]