Repository Public
1"""Mesure de la mémoire MLX réellement consommée selon le nombre de tokens en contexte.23Pour chaque palier (n requêtes simultanées × contexte de c tokens), remet le pic à zéro, lance les requêtes4(contextes tous différents : aucun cache de préfixe), puis relève le pic MLX, la pression mémoire et le swap.5Sert à fixer `token_budget` (config/models.json) et `ENGINE_MEMORY_LIMIT` (deploy/engine.env).67 python bench/memtest.py --base http://127.0.0.1:8821 --plan 1x2000,1x8000,4x8000,1x30000,8x40008"""910from __future__ import annotations1112import argparse13import asyncio14import json15import time16from pathlib import Path1718import httpx1920from bench import Client, long_text, memory_state212223async def main_async(args: argparse.Namespace) -> None:24 c = Client(args.base, "default_model", None)25 http = httpx.AsyncClient(timeout=10)26 rows = []27 base_stats = (await http.get(f"{args.base}/internal/stats")).json()28 idle = base_stats["mlx"]["active_bytes"]29 print(f"mémoire MLX au repos : {idle / 1e9:.2f} Go")30 for step in args.plan.split(","):31 n, ctx = (int(x) for x in step.split("x"))32 await http.post(f"{args.base}/internal/reset_peak")33 t0 = time.perf_counter()34 samples = await asyncio.gather(35 *[36 c.chat(37 [38 {39 "role": "user",40 "content": long_text(ctx, f"M{n}-{ctx}-{i}-{time.time()}-") + "\n\nRésume en une phrase.",41 }42 ],43 max_tokens=args.max_tokens,44 )45 for i in range(n)46 ]47 )48 wall = time.perf_counter() - t049 st = (await http.get(f"{args.base}/internal/stats")).json()50 tokens = sum(s.prompt_tokens + s.completion_tokens for s in samples)51 peak = st["mlx"]["peak_bytes"]52 row = {53 "requests": n,54 "context_tokens_each": ctx,55 "tokens_in_flight": tokens,56 "errors": sum(not s.ok for s in samples),57 "mlx_peak_bytes": peak,58 "peak_over_idle_bytes": peak - idle,59 "bytes_per_token": round((peak - idle) / tokens) if tokens else None,60 "mlx_active_after_bytes": st["mlx"]["active_bytes"],61 "prompt_cache_bytes": st["prompt_cache"]["bytes"],62 "wall_s": round(wall, 1),63 "prefill_tokens_per_s": round(64 sum(s.prompt_tokens for s in samples) / max(0.001, max((s.ttft_ms or 0) for s in samples) / 1000)65 ),66 "memory": memory_state(),67 "first_error": next((s.error for s in samples if s.error), None),68 }69 rows.append(row)70 print(json.dumps(row, ensure_ascii=False), flush=True)71 out = Path(args.out).expanduser()72 out.mkdir(parents=True, exist_ok=True)73 path = out / f"{time.strftime('%Y%m%d-%H%M%S')}-memory-{args.label}.json"74 path.write_text(75 json.dumps(76 {77 "meta": {78 "date": time.strftime("%Y-%m-%d %H:%M:%S"),79 "label": args.label,80 "plan": args.plan,81 "engine": base_stats,82 },83 "idle_bytes": idle,84 "steps": rows,85 },86 ensure_ascii=False,87 indent=1,88 )89 )90 print(f"rapport : {path}")919293def main() -> None:94 ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)95 ap.add_argument("--base", default="http://127.0.0.1:8821")96 ap.add_argument("--plan", default="1x2000,1x8000,4x8000,1x30000,8x4000")97 ap.add_argument("--max-tokens", type=int, default=64)98 ap.add_argument("--label", default="memtest")99 ap.add_argument("--out", default="~/maclustr-llm-data/bench")100 asyncio.run(main_async(ap.parse_args()))101102103if __name__ == "__main__":104 main()105