#!/usr/bin/env python3
"""Focused 256K validation for the Qwen3.8 IQ4_XS-pure artifact.

Reuses the synthetic dual-GPU runner. It reads no chats or user data, stops
production inference for the isolated test and restores the Fast profile in
the runner's finally block.
"""

from __future__ import annotations

import importlib.util
import pathlib
import sys


BASE = pathlib.Path("/opt/mike-ai/stack/run_qwen38_dualgpu_exhaustive.py")
SPEC = importlib.util.spec_from_file_location("qwen38_dualgpu", BASE)
if SPEC is None or SPEC.loader is None:
    raise RuntimeError(f"Could not load {BASE}")

runner = importlib.util.module_from_spec(SPEC)
sys.modules[SPEC.name] = runner
SPEC.loader.exec_module(runner)

runner.OUT = pathlib.Path("/data/benchmarks/qwen38-pure-256k-20260822")
runner.CASES = [
    runner.Case(
        "pure-262k-80-20-mtp2",
        "iq4-pure",
        262144,
        (80, 20),
        mtp=True,
        mtp_max=2,
        quality=False,
        long_fill=0.84,
    ),
]

raise SystemExit(runner.main())