#!/usr/bin/env python3
"""Final acceptance of the fastest stable Abliterated 80K candidate."""

from __future__ import annotations

import base64
import importlib.util
import pathlib
import sys
import time


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.IMAGE = "mike-ai/llama.cpp:3f545bec-test"
runner.PROJECTOR = "/models/qwen3.8-27b-abliterated/mmproj-Qwen3.8-27B-ABLITERATED-F16.gguf"
runner.OUT = pathlib.Path("/data/benchmarks/qwen38-abliterated-final-20260822")
runner.TASK_FILE = pathlib.Path("/opt/mike-ai/stack/dev/QWEN38-FINAL-ACCEPTANCE-v1.json")
runner.MODELS["abliterated"] = "/models/qwen3.8-27b-abliterated/Qwen3.8-27B-ABLITERATED-Q4_K_M.gguf"
runner.CASES = [
    runner.Case("17-abliterated-80k-90x10-mtp2-p010-final", "abliterated", 80000, (90, 10), projector="gpu", mtp_max=2, quality=True, long_fill=0.75, vision=True),
]

_base_command = runner.command


def command(case):
    args = _base_command(case)
    image_index = args.index(runner.IMAGE)
    args[image_index:image_index] = ["-e", "MTMD_BACKEND_DEVICE=CUDA1"]
    return args + ["--mmproj-device", "CUDA1", "--spec-draft-p-min", "0.10"]


def production(action: str) -> None:
    names = [
        "mike-ai-open-webui", "mike-ai-router", "mike-ai-profile-controller",
        "mike-ai-llama-fast", "mike-ai-llama-medium", "mike-ai-llama-large",
        "mike-ai-llama-ultra", "mike-ai-llama-uncensored",
        "mike-ai-llama-experimental",
    ]
    if action == "stop":
        runner.run(["docker", "stop", *names], check=False, timeout=240)
        return
    runner.run([
        "docker", "compose", "-f", "/opt/mike-ai/stack/compose.yaml",
        "--env-file", "/etc/mike-ai/stack.env", "--profile", "inference",
        "up", "-d", "router", "open-webui", "profile-controller", "llama-medium",
    ], check=False, timeout=900)


def vision_probe(case, image: bytes) -> dict:
    payload = {
        "model": case.id, "temperature": 0.1, "max_tokens": 1200,
        "messages": [{"role": "user", "content": [
            {"type": "text", "text": "Describe the image precisely. What animal or object is visible, and what text can you read? Do not guess."},
            {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64," + base64.b64encode(image).decode()}},
        ]}],
    }
    started = time.monotonic()
    try:
        response = runner.api("/v1/chat/completions", payload, timeout=1200)
        choice = (response.get("choices") or [{}])[0]
        message = choice.get("message") or {}
        return {"ok": True, "elapsed_seconds": round(time.monotonic() - started, 3),
                "finish_reason": choice.get("finish_reason"), "content": message.get("content", ""),
                "reasoning_content": message.get("reasoning_content", ""), "timings": response.get("timings", {})}
    except Exception as exc:
        return {"ok": False, "elapsed_seconds": round(time.monotonic() - started, 3),
                "error": f"{type(exc).__name__}: {exc}"}


runner.command = command
runner.production = production
runner.vision_probe = vision_probe
raise SystemExit(runner.main())