benchmark_inference.py

"""Benchmark AniFileBERT PyTorch and ONNX Runtime inference.

The benchmark measures end-to-end parser latency after model/session loading.
It includes tokenization, model forward pass, constrained BIO decoding, and
field postprocessing.
"""

import argparse
import json
import statistics
import time
from pathlib import Path
from typing import Callable, Dict, List

import torch
import onnxruntime as ort
from transformers import BertForTokenClassification

from config import Config
from evaluate_parser_cases import DEFAULT_CASE_FILE, load_cases
from inference import parse_filename
from onnx_inference import OnnxFilenameParser
from tokenizer import load_tokenizer


def percentile(values: List[float], pct: float) -> float:
    if not values:
        return 0.0
    ordered = sorted(values)
    index = (len(ordered) - 1) * pct
    lower = int(index)
    upper = min(lower + 1, len(ordered) - 1)
    if lower == upper:
        return ordered[lower]
    weight = index - lower
    return ordered[lower] * (1 - weight) + ordered[upper] * weight


def summarize(name: str, load_ms: float, latencies_ms: List[float]) -> Dict:
    total_ms = sum(latencies_ms)
    count = len(latencies_ms)
    return {
        "name": name,
        "load_ms": load_ms,
        "runs": count,
        "avg_ms": statistics.fmean(latencies_ms) if latencies_ms else 0.0,
        "p50_ms": percentile(latencies_ms, 0.50),
        "p95_ms": percentile(latencies_ms, 0.95),
        "p99_ms": percentile(latencies_ms, 0.99),
        "min_ms": min(latencies_ms) if latencies_ms else 0.0,
        "max_ms": max(latencies_ms) if latencies_ms else 0.0,
        "throughput_fps": (count / (total_ms / 1000.0)) if total_ms > 0 else 0.0,
    }


def run_benchmark(
    name: str,
    parser_fn: Callable[[str], Dict],
    filenames: List[str],
    warmup: int,
    repeat: int,
) -> Dict:
    for idx in range(warmup):
        parser_fn(filenames[idx % len(filenames)])

    latencies: List[float] = []
    for _ in range(repeat):
        for filename in filenames:
            start = time.perf_counter()
            parser_fn(filename)
            latencies.append((time.perf_counter() - start) * 1000.0)
    return {"name": name, "latencies_ms": latencies}


def load_case_filenames(case_file: str, limit: int | None) -> List[str]:
    cases = load_cases(case_file)
    filenames = [case["filename"] for case in cases if case.get("filename")]
    if limit is not None and limit > 0:
        filenames = filenames[:limit]
    if not filenames:
        raise ValueError(f"No filenames found in {case_file}")
    return filenames


def main() -> None:
    parser = argparse.ArgumentParser(description="Benchmark AniFileBERT inference speed")
    parser.add_argument("--model-dir", default=".", help="Directory containing the PyTorch checkpoint")
    parser.add_argument("--onnx", default="exports/anime_filename_parser.onnx", help="ONNX model path")
    parser.add_argument("--case-file", default=DEFAULT_CASE_FILE, help="JSON regression case file")
    parser.add_argument("--max-length", type=int, default=None, help="Override sequence length")
    parser.add_argument("--limit-cases", type=int, default=None, help="Use only the first N cases")
    parser.add_argument("--repeat", type=int, default=5, help="Repeat the case set this many times")
    parser.add_argument("--warmup", type=int, default=10, help="Warmup parses per backend")
    parser.add_argument("--backend", choices=["both", "torch", "onnx"], default="both")
    parser.add_argument("--torch-threads", type=int, default=1, help="torch intra-op thread count")
    parser.add_argument("--ort-threads", type=int, default=1, help="ONNX Runtime intra/inter-op thread count")
    parser.add_argument("--no-constrained-bio", action="store_true", help="Use greedy labels for PyTorch backend")
    parser.add_argument("--output", default=None, help="Optional JSON output path")
    args = parser.parse_args()

    filenames = load_case_filenames(args.case_file, args.limit_cases)
    model_dir = Path(args.model_dir)
    max_length = args.max_length

    if args.torch_threads and args.torch_threads > 0:
        torch.set_num_threads(args.torch_threads)
        torch.set_num_interop_threads(args.torch_threads)

    results: List[Dict] = []

    if args.backend in {"both", "torch"}:
        cfg = Config()
        load_start = time.perf_counter()
        tokenizer = load_tokenizer(str(model_dir))
        model = BertForTokenClassification.from_pretrained(model_dir)
        model.eval()
        resolved_max_length = max_length or int(getattr(model.config, "max_seq_length", 128))
        id2label = {int(k): v for k, v in getattr(model.config, "id2label", cfg.id2label).items()}
        load_ms = (time.perf_counter() - load_start) * 1000.0

        def parse_torch(filename: str) -> Dict:
            return parse_filename(
                filename,
                model,
                tokenizer,
                id2label,
                max_length=resolved_max_length,
                debug=False,
                constrain_bio=not args.no_constrained_bio,
            )

        raw = run_benchmark("pytorch", parse_torch, filenames, args.warmup, args.repeat)
        results.append(summarize(raw["name"], load_ms, raw["latencies_ms"]))

    if args.backend in {"both", "onnx"}:
        session_options = ort.SessionOptions()
        if args.ort_threads and args.ort_threads > 0:
            session_options.intra_op_num_threads = args.ort_threads
            session_options.inter_op_num_threads = args.ort_threads
        load_start = time.perf_counter()
        onnx_parser = OnnxFilenameParser(
            model_dir=model_dir,
            onnx_path=Path(args.onnx),
            max_length=max_length or 128,
            session_options=session_options,
        )
        load_ms = (time.perf_counter() - load_start) * 1000.0

        def parse_onnx(filename: str) -> Dict:
            return onnx_parser.parse(filename)

        raw = run_benchmark("onnxruntime", parse_onnx, filenames, args.warmup, args.repeat)
        results.append(summarize(raw["name"], load_ms, raw["latencies_ms"]))

    report = {
        "model_dir": str(model_dir),
        "onnx": args.onnx,
        "case_file": args.case_file,
        "case_count": len(filenames),
        "repeat": args.repeat,
        "warmup": args.warmup,
        "torch_threads": args.torch_threads,
        "ort_threads": args.ort_threads,
        "constrain_bio": not args.no_constrained_bio,
        "results": results,
    }

    print(json.dumps(report, ensure_ascii=False, indent=2))
    print("\nSummary:")
    print("| Backend | Load ms | Avg ms | P50 ms | P95 ms | P99 ms | Throughput files/s |")
    print("| --- | ---: | ---: | ---: | ---: | ---: | ---: |")
    for item in results:
        print(
            f"| {item['name']} | {item['load_ms']:.2f} | {item['avg_ms']:.3f} | "
            f"{item['p50_ms']:.3f} | {item['p95_ms']:.3f} | {item['p99_ms']:.3f} | "
            f"{item['throughput_fps']:.1f} |"
        )

    if args.output:
        output_path = Path(args.output)
        output_path.parent.mkdir(parents=True, exist_ok=True)
        output_path.write_text(json.dumps(report, ensure_ascii=False, indent=2), encoding="utf-8")


if __name__ == "__main__":
    main()