Download scripts/qa_cs/cs_retrieve.py from Raphael2099/llm-data-eval-20261030: direct link, hf CLI and curl.
- Browser
- Download file 8.42 kB
-
https://huggingface.co/datasets/Raphael2099/llm-data-eval-20261030/resolve/main/scripts/qa_cs/cs_retrieve.py
- Command line
-
hf download hf://datasets/Raphael2099/llm-data-eval-20261030/scripts/qa_cs/cs_retrieve.py
-
curl -L -o cs_retrieve.py https://huggingface.co/datasets/Raphael2099/llm-data-eval-20261030/resolve/main/scripts/qa_cs/cs_retrieve.py
8.42 kB
| #!/usr/bin/env python3 | |
| # -*- coding: utf-8 -*- | |
| """cs_retrieve.py —— 传世考释问答 v1 · 两路检索(进程内 fls,只读线上 shards.json:ft1 4B@2560 + rr3;不走 HTTP、不碰 18080、不写 service/)。 | |
| 路 1 诸家说:lib C,source_type ∈ {paper, related, collected},channel main,k=16 候选(doc_cap 1)→ 去 eval/dev_overlap; | |
| 装配时与抽样锚(C 主快照 8-gram 引用块)合并,共 ≤5 条(锚优先 ≤2)。 | |
| 路 2 传世对读:lib A2,source_type ∈ {classic, parallel},k=8(doc_cap 1)→ 去本段自身与同书他版(近似重文)、去 eval/dev_overlap。 | |
| 不跑 lib A 辞例与 private/theory 路(传世问答不需要;省一半检索时间)。 | |
| 并发:--workers N 子进程(同一张卡,CUDA_VISIBLE_DEVICES 由外层给定)× --threads T;各写 OUT.shardK(断点续跑); | |
| --deadline "YYYY-mm-dd HH:MM":到点各子进程停止取新记录、落盘退出;合并已完成部分(按输入序)。 | |
| 用法: CUDA_VISIBLE_DEVICES=2 python cs_retrieve.py --in S.jsonl --out L.jsonl --workers 6 --threads 4 --deadline "2026-10-24 13:58" | |
| """ | |
| import argparse, json, os, subprocess, sys, time, datetime | |
| W = "/root/autodl-tmp/llm_rag_20261023" | |
| SERVICE = "/root/autodl-tmp/llm_rag_20260919/service" | |
| def log(m): | |
| sys.stdout.write("%s [cs_retrieve] %s\n" % (time.strftime("%H:%M:%S"), m)); sys.stdout.flush() | |
| def read_ids(p): | |
| s = set() | |
| if os.path.exists(p): | |
| for l in open(p, encoding="utf-8"): | |
| try: | |
| s.add(json.loads(l)["id"]) | |
| except Exception: | |
| pass | |
| return s | |
| def slim(b): | |
| m = b.get("meta") or {} | |
| return {"id": b.get("id"), "cite": b.get("cite") or "", "source_type": b.get("source_type") or "", | |
| "text": b.get("text") or "", | |
| "meta": {k: m.get(k) for k in ("lib", "channel", "source_lib", "doc_key", "quotable", "eval_overlap", "dev_overlap", | |
| "book", "author", "year", "pian", "corpus") if k in m}, | |
| "eval_overlap": bool(b.get("eval_overlap") or m.get("eval_overlap")), | |
| "dev_overlap": bool(b.get("dev_overlap") or m.get("dev_overlap")), | |
| "rerank": (b.get("score") or {}).get("rerank")} | |
| def worker(a): | |
| sys.dont_write_bytecode = True | |
| os.environ["PYTHONDONTWRITEBYTECODE"] = "1" | |
| for k in ("FLS_ROOT", "FLS_SHARDS", "FLS_SCRIPTS", "FLS_PORT"): | |
| os.environ.pop(k, None) | |
| for k in ("FLS_OMP_THREADS", "OMP_NUM_THREADS", "MKL_NUM_THREADS", "OPENBLAS_NUM_THREADS", "NUMEXPR_NUM_THREADS"): | |
| os.environ[k] = "8" | |
| if a.cpu_smoke: | |
| os.environ["CUDA_VISIBLE_DEVICES"] = "" | |
| else: | |
| assert os.environ.get("CUDA_VISIBLE_DEVICES") == a.gpu, "CUDA_VISIBLE_DEVICES 须为 %s" % a.gpu | |
| os.environ["FLS_BATCH_DIR"] = os.path.join(W, "tmp", "fls_batch_unused") | |
| sys.path.insert(0, SERVICE) | |
| import fls_core # noqa: F401 | |
| import threading | |
| from concurrent.futures import ThreadPoolExecutor | |
| dl = datetime.datetime.strptime(a.deadline, "%Y-%m-%d %H:%M").timestamp() if a.deadline else 0 | |
| rows = [json.loads(l) for l in open(a.inp, encoding="utf-8") if l.strip()] | |
| if a.limit: | |
| rows = rows[:a.limit] | |
| rows = [r for i, r in enumerate(rows) if i % a.nshards == a.shard] | |
| done = read_ids(a.out) | |
| todo = [r for r in rows if r["cid"] not in done] | |
| log("shard %d/%d: %d rows, %d done, %d todo" % (a.shard, a.nshards, len(rows), len(done), len(todo))) | |
| if not todo: | |
| return 0 | |
| t0 = time.time() | |
| s = fls_core.FederatedSearcher(use_dense=not a.cpu_smoke, use_rerank=not a.cpu_smoke, quiet=True) | |
| timing, deg = s.warmup() | |
| if a.cpu_smoke: | |
| deg = [] | |
| h = s.health(); m = h.get("models") or {} | |
| log("searcher ready %.1fs degraded=%s embed=%s rerank=%s fp=%s" % (time.time() - t0, deg, m.get("embed"), m.get("rerank"), s.state_fingerprint())) | |
| assert not deg, "warmup degraded %s" % deg | |
| lock = threading.Lock(); st = {"n": 0, "deg": 0, "t1": time.time(), "stop": False} | |
| f = open(a.out, "a", encoding="utf-8") | |
| def srch(**kw): | |
| try: | |
| r = s.search(**kw) | |
| except Exception as e: | |
| return [], ["search_error:%s" % type(e).__name__] | |
| if not r.get("ok"): | |
| return [], ["search_error"] | |
| return list(r.get("blocks") or []), list(r.get("degraded") or []) | |
| def one(r): | |
| if st["stop"] or (dl and time.time() > dl): | |
| st["stop"] = True; return 0 | |
| q = r["text"][:300] | |
| cb, d1 = srch(q=q, k=16, doc_cap=1, quota={"C": 16}, | |
| filters={"lib": ["C"], "source_type": ["paper", "related", "collected"], "channel": ["main"]}) | |
| ab, d2 = srch(q=q, k=8, doc_cap=1, quota={"A2": 8}, filters={"lib": ["A2"], "source_type": ["classic", "parallel"]}) | |
| cb = [slim(b) for b in cb if (b.get("meta") or {}).get("lib") == "C"] | |
| ab = [slim(b) for b in ab if (b.get("meta") or {}).get("lib") == "A2" and b.get("id") != r["cid"]] | |
| out = {"id": r["cid"], "c_blocks": cb, "a2_blocks": ab, "degraded": sorted(set(d1) | set(d2))} | |
| with lock: | |
| f.write(json.dumps(out, ensure_ascii=False, separators=(",", ":")) + "\n") | |
| st["n"] += 1; st["deg"] += bool(out["degraded"]) | |
| if st["n"] % 100 == 0: | |
| f.flush(); os.fsync(f.fileno()) | |
| log("shard %d: %d/%d %.2f rec/s degraded=%d" % (a.shard, st["n"], len(todo), st["n"] / (time.time() - st["t1"]), st["deg"])) | |
| return 0 | |
| with ThreadPoolExecutor(max_workers=a.threads) as ex: | |
| list(ex.map(one, todo)) | |
| f.flush(); os.fsync(f.fileno()); f.close() | |
| meta = {"shard": a.shard, "n": st["n"], "todo": len(todo), "stopped_at_deadline": st["stop"], "sec": round(time.time() - st["t1"], 1), | |
| "degraded_records": st["deg"], "fingerprint": s.state_fingerprint(), "code_sha256": s.code_sha256(), | |
| "blocklist_sha256": h.get("blocklist_sha256"), "models": m, "tuning": s.tuning_defaults(), | |
| "lanes": "C{paper,related,collected}/main k16 doc_cap1; A2{classic,parallel} k8 doc_cap1", "threads": a.threads} | |
| json.dump(meta, open(a.out + ".meta.json", "w"), ensure_ascii=False, indent=1, default=str) | |
| try: | |
| s.close() | |
| except Exception: | |
| pass | |
| return 0 | |
| def main(): | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("--in", dest="inp", required=True) | |
| ap.add_argument("--out", required=True) | |
| ap.add_argument("--workers", type=int, default=1) | |
| ap.add_argument("--threads", type=int, default=4) | |
| ap.add_argument("--cpu-smoke", action="store_true") | |
| ap.add_argument("--gpu", default="2") | |
| ap.add_argument("--shard", type=int, default=-1) | |
| ap.add_argument("--nshards", type=int, default=1) | |
| ap.add_argument("--limit", type=int, default=0) | |
| ap.add_argument("--deadline", default="") | |
| a = ap.parse_args() | |
| if a.shard >= 0: | |
| return worker(a) | |
| procs = [] | |
| for k in range(max(1, a.workers)): | |
| cmd = [sys.executable, os.path.abspath(__file__), "--in", a.inp, "--out", "%s.shard%d" % (a.out, k), "--shard", str(k), | |
| "--nshards", str(max(1, a.workers)), "--threads", str(a.threads), "--gpu", a.gpu, "--deadline", a.deadline] + \ | |
| (["--cpu-smoke"] if a.cpu_smoke else []) + (["--limit", str(a.limit)] if a.limit else []) | |
| procs.append(subprocess.Popen(cmd)); time.sleep(2) | |
| rcs = [p.wait() for p in procs] | |
| log("workers rc=%s" % rcs) | |
| got = {} | |
| for k in range(max(1, a.workers)): | |
| p = "%s.shard%d" % (a.out, k) | |
| if os.path.exists(p): | |
| for l in open(p, encoding="utf-8"): | |
| if l.strip(): | |
| try: | |
| o = json.loads(l) | |
| except Exception: | |
| continue | |
| got[o["id"]] = l.rstrip("\n") | |
| ids = [json.loads(l)["cid"] for l in open(a.inp, encoding="utf-8") if l.strip()] | |
| if a.limit: | |
| ids = ids[:a.limit] | |
| tmp = a.out + ".tmp" | |
| with open(tmp, "w", encoding="utf-8") as f: | |
| for i in ids: | |
| if i in got: | |
| f.write(got[i] + "\n") | |
| os.replace(tmp, a.out) | |
| n = sum(1 for i in ids if i in got) | |
| log("merged %d / %d -> %s" % (n, len(ids), a.out)) | |
| return 0 if n > 0 and not any(rcs) else (5 if n == 0 else 0) | |
| if __name__ == "__main__": | |
| sys.exit(main()) | |