"""End-to-end router latency (tokenize + route + Core ML predict + decode) per bucket, from Python.

    uv run python bench/router_latency.py
"""
import statistics
import sys
import time
import warnings
from pathlib import Path

warnings.filterwarnings("ignore")
sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "tests"))

from cases import CASES  # noqa: E402
from gliner_decide_coreml import DecideRouter  # noqa: E402

router = DecideRouter()
seen = set()
print(f"{'case':<28}{'tokens':>7}{'bucket':>8}{'p50':>10}{'p90':>10}")
for case in CASES:
    if case["bucket"] in seen:
        continue
    seen.add(case["bucket"])
    for _ in range(5):
        router.classify(case["text"], case["tasks"])
    times = []
    for _ in range(30):
        start = time.perf_counter()
        _, route = router.classify_with_route(case["text"], case["tasks"])
        times.append((time.perf_counter() - start) * 1000)
    times.sort()
    print(f"{case['id']:<28}{route.tokens:>7}{route.bucket:>8}{statistics.median(times):>8.1f}ms"
          f"{times[int(len(times) * 0.9)]:>8.1f}ms")