#!/usr/bin/env python3 """LLM Lab — отдельный трек ускорения локальной LLM (read-only). Замеряет prefill и генерацию Ollama на разных размерах промпта и моделях. Ничего не меняет в сервисах: только HTTP-вызовы и отчёт. Запуск: llm_lab.py --quick # быстрый смоук (1 модель, 3 размера) llm_lab.py --full # все модели/размеры (долго) llm_lab.py --models qwen2.5:3b --sizes 500,4000 llm_lab.py --json # только JSONL Результаты: /var/lib/llm-lab/YYYY-MM-DD.jsonl и summary.md """ import argparse import json import os import sys import time import urllib.request from datetime import date LLM = "http://192.168.1.229:11434" OUT = "/var/lib/llm-lab" DEFAULT_MODELS = ["qwen2.5:3b-32k", "qwen2.5:3b", "qwen2.5:7b"] DEFAULT_SIZES = [300, 2000, 8000, 14000] QUICK_SIZES = [300, 2000] def make_prompt(tokens): # ~4 символа на токен для английского; для русского будет больше токенов n = max(40, int(tokens * 4)) base = ("The quick brown fox jumps over the lazy dog. ") return (base * (n // len(base) + 1))[:n] def post(path, payload, timeout): req = urllib.request.Request(LLM + path, data=json.dumps(payload).encode(), headers={"Content-Type": "application/json"}, method="POST") t0 = time.time() with urllib.request.urlopen(req, timeout=timeout) as r: body = json.loads(r.read().decode()) return body, time.time() - t0 def test_native(model, tokens, timeout): prompt = make_prompt(tokens) + "\nОтветь одним словом: ок" payload = {"model": model, "messages": [{"role": "user", "content": prompt}], "stream": False, "options": {"num_ctx": 32768, "num_predict": 8}} body, wall = post("/api/chat", payload, timeout) pe = body.get("prompt_eval_count") or 0 pd = (body.get("prompt_eval_duration") or 0) / 1e9 ec = body.get("eval_count") or 0 ed = (body.get("eval_duration") or 0) / 1e9 return { "model": model, "target_tokens": tokens, "prompt_tokens": pe, "gen_tokens": ec, "prefill_s": round(pd, 2), "gen_s": round(ed, 2), "wall_s": round(wall, 2), "prefill_tps": round(pe / pd, 1) if pd > 0 else None, "gen_tps": round(ec / ed, 1) if ed > 0 else None, "reply": (body.get("message") or {}).get("content", "")[:80], } def test_openai(model, tokens, timeout): prompt = make_prompt(tokens) + "\nОтветь одним словом: ок" payload = {"model": model, "messages": [{"role": "user", "content": prompt}], "max_tokens": 8} body, wall = post("/v1/chat/completions", payload, timeout) usage = body.get("usage") or {} return { "model": model, "target_tokens": tokens, "endpoint": "openai", "prompt_tokens": usage.get("prompt_tokens"), "wall_s": round(wall, 2), "reply": ((body.get("choices") or [{}])[0].get("message") or {}).get("content", "")[:80], } def warm(model, timeout=300): try: post("/api/chat", {"model": model, "messages": [{"role": "user", "content": "ок"}], "stream": False, "options": {"num_predict": 1}}, timeout) except Exception: pass def main(): ap = argparse.ArgumentParser() ap.add_argument("--quick", action="store_true") ap.add_argument("--full", action="store_true") ap.add_argument("--models", default="") ap.add_argument("--sizes", default="") ap.add_argument("--timeout", type=int, default=900) ap.add_argument("--json", action="store_true") args = ap.parse_args() models = [m.strip() for m in args.models.split(",") if m.strip()] or DEFAULT_MODELS sizes = [int(s) for s in args.sizes.split(",") if s.strip()] or \ (QUICK_SIZES if args.quick else DEFAULT_SIZES) if args.quick: models = models[:1] os.makedirs(OUT, exist_ok=True) out = os.path.join(OUT, date.today().isoformat() + ".jsonl") rows = [] print("LLM Lab: модели=%s размеры=%s" % (models, sizes)) for m in models: print(" прогрев", m) warm(m) for n in sizes: try: r = test_native(m, n, args.timeout) r["endpoint"] = "native" except Exception as e: r = {"model": m, "target_tokens": n, "error": str(e)[:120]} rows.append(r) with open(out, "a", encoding="utf-8") as f: f.write(json.dumps(r, ensure_ascii=False) + "\n") if not args.json: if r.get("error"): print(" %-20s %5d → ошибка: %s" % (m, n, r["error"])) else: print(" %-20s %5d → prefill %.1fс (%.0f т/с), gen %.1fс, всего %.1fс" % (m, n, r["prefill_s"], r["prefill_tps"] or 0, r["gen_s"], r["wall_s"])) # один замер через OpenAI-совместимый endpoint для сравнения try: r = test_openai(models[0], sizes[0], args.timeout) rows.append(r) with open(out, "a", encoding="utf-8") as f: f.write(json.dumps(r, ensure_ascii=False) + "\n") if not args.json: print(" openai endpoint: %.1fс" % r["wall_s"]) except Exception as e: print(" openai endpoint: ошибка", str(e)[:100]) # оценка: сколько займёт промпт OpenClaw (~15k токенов) native = [r for r in rows if r.get("endpoint") == "native" and r.get("prefill_s")] if native: rate = sum(r["prefill_tps"] for r in native if r["prefill_tps"]) / max( 1, len([r for r in native if r["prefill_tps"]])) est = 15000 / rate if rate else None print("\nОценка: при %.0f т/с prefill 15k токенов ≈ %.0f с" % (rate, est or 0)) if not args.json: path = os.path.join(OUT, "summary.md") with open(path, "w", encoding="utf-8") as f: f.write("# LLM Lab — %s\n\n" % date.today().isoformat()) f.write("| модель | токенов | prefill, с | prefill, т/с | gen, с | всего, с |\n") f.write("|---|---|---|---|---|---|\n") for r in rows: if r.get("endpoint") == "native" and not r.get("error"): f.write("| %s | %d | %.1f | %.0f | %.1f | %.1f |\n" % ( r["model"], r["target_tokens"], r["prefill_s"], r["prefill_tps"] or 0, r["gen_s"], r["wall_s"])) print("отчёт:", path) return 0 if __name__ == "__main__": sys.exit(main())