| |
| """GPU fix loop on data/fix_todo.jsonl with the clean labeler: up to --rounds rounds, each round regenerates |
| the first still-failing band of every row (prefix = game + earlier bands, --n candidates, first one passing |
| the rule check is spliced in) and re-checks. Rows with 3+ bad bands first try a whole new set. |
| Writes labels/fix/fix_wK.jsonl {key, labels, src, bad_after} + .done.""" |
| import argparse |
| import json |
| import sys |
| import time |
| import types |
| from pathlib import Path |
|
|
| sys.path.insert(0, str(Path(__file__).resolve().parent)) |
| import grounding as G |
| import train as T |
| from repair_labels import bad_bands |
|
|
| I = "/iopsstor/scratch/cscs/ibadanin/projects/llm_chess/" |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--ckpt", default=I + "results/labeler/ckpt_full_gemma3_1b_clean") |
| ap.add_argument("--worker", type=int, required=True) |
| ap.add_argument("--n-workers", type=int, required=True) |
| ap.add_argument("--rounds", type=int, default=5) |
| ap.add_argument("--n", type=int, default=8) |
| ap.add_argument("--sub", type=int, default=0, help="rerun one worker's rows split over --n-sub GPUs") |
| ap.add_argument("--n-sub", type=int, default=1) |
| a = ap.parse_args() |
| out = Path(I + f"labels/fix/fix_w{a.worker}" + (f"s{a.sub}" if a.n_sub > 1 else "") + ".jsonl") |
| mark = Path(str(out) + (".ok" if a.n_sub > 1 else ".done")) |
| out.parent.mkdir(parents=True, exist_ok=True) |
| rows = [json.loads(l) for i, l in enumerate(open(I + "data/fix_todo.jsonl")) if i % a.n_workers == a.worker] |
| rows = rows[a.sub::a.n_sub] |
| from transformers import AutoTokenizer |
| from vllm import LLM, SamplingParams |
|
|
| cfg = types.SimpleNamespace(input="compact", tag=0, special_tokens=0, max_in=2048) |
| tok = AutoTokenizer.from_pretrained(a.ckpt) |
| llm = LLM(model=a.ckpt, dtype="bfloat16", max_model_len=2560, gpu_memory_utilization=0.9, |
| max_num_seqs=1024, max_num_batched_tokens=32768, enable_prefix_caching=True) |
| sp_line = SamplingParams(n=a.n, temperature=0.8, max_tokens=120, stop=["\n"]) |
| sp_set = SamplingParams(n=a.n, temperature=0.8, max_tokens=320) |
| enc = lambda text: {"prompt_token_ids": tok(text, add_special_tokens=True)["input_ids"][-2300:]} |
| for r in rows: |
| r["bad_now"] = bad_bands(r["user"], r["labels"]) |
| print(f"worker {a.worker}: {len(rows)} rows, {sum(map(len, (r['bad_now'] for r in rows)))} bad bands", flush=True) |
| t0 = time.time() |
| multi = [r for r in rows if len(r["bad_now"]) >= 3] |
| for r, o in zip(multi, llm.generate([enc(T.prompt_text(r, cfg, None)) for r in multi], sp_set, use_tqdm=False)): |
| sets = [s for s in (T.parse_joint(c.text) for c in o.outputs) if all(s.get(b) for b in T.BANDS)] |
| if sets: |
| best = min(sets, key=lambda s: len(bad_bands(r["user"], s))) |
| if len(bad_bands(r["user"], best)) < len(r["bad_now"]): |
| r["labels"], r["src"], r["bad_now"] = {b: best[b] for b in T.BANDS}, "replace", bad_bands(r["user"], best) |
| for k in range(a.rounds): |
| todo = [r for r in rows if r["bad_now"]] |
| if not todo: |
| break |
| ps = [] |
| for r in todo: |
| b = r["bad_now"][0] |
| head = "".join(f"{T.BAND_NAME[x]}: {r['labels'][x].strip()}\n" for x in T.BANDS[: T.BANDS.index(b)]) |
| ps.append(enc(T.prompt_text(r, cfg, None) + head + f"{T.BAND_NAME[b]}:")) |
| stuck = 0 |
| for r, o in zip(todo, llm.generate(ps, sp_line, use_tqdm=False)): |
| b = r["bad_now"][0] |
| ok = [c.text.strip() for c in o.outputs if c.text.strip() and not bad_bands(r["user"], {b: c.text.strip()})] |
| if ok: |
| r["labels"] = dict(r["labels"], **{b: ok[0]}) |
| r["src"] = "patch" if r["src"] == "orig" else r["src"] |
| else: |
| stuck += 1 |
| r["bad_now"] = sorted(bad_bands(r["user"], r["labels"]), key=lambda x: (x == b and not ok, T.BANDS.index(x))) |
| left = sum(len(r["bad_now"]) for r in rows) |
| print(f"round {k + 1}: tried {len(todo)} stuck {stuck} bad bands left {left} ({time.time() - t0:.0f}s)", flush=True) |
| with open(str(out) + ".tmp", "w") as fo: |
| for r in rows: |
| fo.write(json.dumps({"key": r["key"], "labels": r["labels"], "src": r["src"], "bad_after": r["bad_now"]}) + "\n") |
| Path(str(out) + ".tmp").rename(out) |
| mark.write_text(json.dumps({"rows": len(rows), "bad_left": sum(len(r["bad_now"]) for r in rows), |
| "sec": round(time.time() - t0)})) |
|
|
