Files
buryan-infra/ops/work/llm_dispatcher.py
T

169 lines
5.4 KiB
Python

#!/usr/bin/env python3
"""Единый диспетчер LLM-сбора: один поток, общая очередь, модели грузятся по нужде.
Зачем: раньше каждый вендор — отдельный сервис, и они дрались за Ollama
(7B стоял в очереди за 3B, а RAG падал по таймауту). Мини-ПК работает 24/7,
значит достаточно одной очереди: берём по одному документу, отдаём модели,
пишем результат. Ollama сама загружает/выгружает нужную модель.
Конфиг: ops/work/research/jobs.yaml — список источников с весом и моделью.
"""
import json
import os
import sys
import time
import yaml
sys.path.insert(0, "/opt/work-hub/research")
OUT = "/opt/work-hub/research"
JOBS = "/opt/work-hub/research/jobs.yaml"
LOG = "/var/log/llm-dispatch.log"
LOCK = "/var/lib/llm-dispatcher.lock"
os.environ.setdefault("OLLAMA_URL", "http://192.168.1.229:11434")
def log(msg):
line = "[%s] %s" % (time.strftime("%Y-%m-%d %H:%M:%S"), msg)
try:
with open(LOG, "a", encoding="utf-8") as f:
f.write(line + "\n")
except Exception:
pass
print(line, flush=True)
def one_instance():
"""Не даём запуститься двум диспетчерам."""
try:
import fcntl
fh = open(LOCK, "w")
fcntl.flock(fh, fcntl.LOCK_EX | fcntl.LOCK_NB)
return fh
except Exception:
return None
def progress_path(job):
tag = ("-" + job["tag"]) if job.get("tag") else ""
return os.path.join(OUT, "progress-%s%s-0.json" % (job["vendor"], tag))
def results_path(job):
tag = ("-" + job["tag"]) if job.get("tag") else ""
return os.path.join(OUT, "results-%s%s-0.jsonl" % (job["vendor"], tag))
def load_progress(job):
try:
return json.load(open(progress_path(job), encoding="utf-8"))
except Exception:
return {}
def save_progress(job, done):
json.dump(done, open(progress_path(job), "w", encoding="utf-8"))
def pending(job):
q = os.path.join(OUT, job["queue"])
try:
items = [json.loads(l) for l in open(q, encoding="utf-8") if l.strip()]
except Exception:
return []
done = load_progress(job)
return [i for i in items if done.get(i["url"]) not in ("ok", "empty", "skip")]
def process(job, item, L):
"""Обрабатывает один документ выбранной моделью."""
L.VENDOR = job["vendor"]
L.MODEL = job.get("model", "qwen2.5:3b")
L.PROXY = job.get("proxy", "")
L.MAX_PREDICT = int(job.get("max_predict", 450))
window = int(job.get("window", 2600))
url = item["url"]
done = load_progress(job)
txt = L.fetch(url)
if len(txt) < 200:
done[url] = "empty"
save_progress(job, done)
return "empty", 0
low = txt.lower()
if job["vendor"] == "yealink":
hits = sum(1 for t in L.MODEL_TOKENS if t in low)
gate = hits > 0 and any(w in low for w in L.SPEC_WORDS)
else:
hits = sum(1 for t in L.VENDOR_TOKENS.get(job["vendor"], []) if t in low)
gate = hits > 0
if not gate:
done[url] = "skip"
save_progress(job, done)
return "skip", 0
chunk = L.pick_excerpt(txt, window)
try:
res = L.ask(chunk)
except Exception as e:
log(" ошибка LLM (%s): %s" % (job["vendor"], str(e)[:100]))
return "error", 0
with open(results_path(job), "a", encoding="utf-8") as f:
f.write(json.dumps({"url": item.get("source_url") or url,
"vendor": job["vendor"], "model": L.MODEL,
"product": item.get("product"), "slug": item.get("slug"),
"chunks": 1, "data": [res] if res else []},
ensure_ascii=False) + "\n")
done[url] = "ok"
save_progress(job, done)
nfacts = len((res or {}).get("facts") or [])
ndeps = len((res or {}).get("dependencies") or [])
return "ok", nfacts + ndeps
def main():
fh = one_instance()
if fh is None:
log("диспетчер уже запущен — выходим")
return
import llm_research as L
log("=== диспетчер старт ===")
empty_rounds = 0
while True:
cfg = yaml.safe_load(open(JOBS, encoding="utf-8")) or {}
jobs = cfg.get("jobs") or []
did_work = False
for job in jobs:
if job.get("enabled") is False:
continue
weight = int(job.get("weight", 1))
for _ in range(weight):
items = pending(job)
if not items:
continue
item = items[0]
try:
status, n = process(job, item, L)
except Exception as e:
log("сбой обработки %s: %s" % (job["vendor"], str(e)[:120]))
continue
did_work = True
log("%-9s %-5s +%-2d %s" % (job["vendor"], status, n, item["url"][-70:]))
time.sleep(1)
if not did_work:
empty_rounds += 1
if empty_rounds == 1:
log("очереди пусты — ждём")
time.sleep(300)
else:
empty_rounds = 0
if __name__ == "__main__":
main()