| |
| """Teacher-forced top-k logprobs for every window of eval-set.json from any vLLM OpenAI server. |
| |
| uv run --with numpy python logprobs.py --base http://127.0.0.1:8002 --model <id> \ |
| --set eval-set.json --out arm.npz [--topk 64] [--sets kld,ppl-wiki,ppl-code] [--workers 4] |
| |
| Server needs `--max-logprobs 64` and speculation OFF (prompt_logprobs is NaN under MTP on the |
| B65 stack). Per set, writes: |
| <set>.target float32 [W, L-1] logprob of the true next token (positions 1..L-1) |
| <set>.ids int32 [W, L-1, k] top-k token ids |
| <set>.lp float16 [W, L-1, k] their logprobs |
| """ |
| import argparse, json, math, time, urllib.request |
| from concurrent.futures import ThreadPoolExecutor |
| import numpy as np |
|
|
| ap = argparse.ArgumentParser() |
| ap.add_argument("--base", required=True); ap.add_argument("--model", required=True) |
| ap.add_argument("--set", required=True); ap.add_argument("--out", required=True) |
| ap.add_argument("--topk", type=int, default=64); ap.add_argument("--workers", type=int, default=4) |
| ap.add_argument("--sets", default="kld,ppl-wiki,ppl-code") |
| ap.add_argument("--api-key", default="local") |
| a = ap.parse_args() |
| es = json.load(open(a.set)) |
|
|
|
|
| def score(ids): |
| body = {"model": a.model, "prompt": ids, "max_tokens": 1, "temperature": 0, |
| "prompt_logprobs": a.topk, "add_special_tokens": False} |
| for attempt in range(5): |
| try: |
| req = urllib.request.Request(a.base.rstrip("/") + "/v1/completions", json.dumps(body).encode(), |
| {"Content-Type": "application/json", "Authorization": f"Bearer {a.api_key}"}) |
| pl = json.load(urllib.request.urlopen(req, timeout=1800))["choices"][0]["prompt_logprobs"] |
| break |
| except Exception as ex: |
| if attempt == 4: |
| raise |
| print(f" retry {attempt + 1}: {ex}", flush=True); time.sleep(10) |
| L = len(ids) |
| tgt = np.empty(L - 1, np.float32); tid = np.zeros((L - 1, a.topk), np.int32) |
| tlp = np.full((L - 1, a.topk), -np.inf, np.float16) |
| for p in range(1, L): |
| d = pl[p] |
| tgt[p - 1] = d[str(ids[p])]["logprob"] |
| top = sorted(((int(t), v["logprob"]) for t, v in d.items() if v.get("rank", 0) <= a.topk), |
| key=lambda x: -x[1])[:a.topk] |
| for j, (t, lp) in enumerate(top): |
| tid[p - 1, j] = t; tlp[p - 1, j] = lp |
| if not np.isfinite(tgt).all(): |
| raise SystemExit("non-finite prompt logprobs: is speculation on?") |
| return tgt, tid, tlp |
|
|
|
|
| out = {} |
| for s in a.sets.split(","): |
| W = es[s]; t0 = time.time() |
| with ThreadPoolExecutor(a.workers) as ex: |
| res = list(ex.map(score, W)) |
| out[f"{s}.target"] = np.stack([r[0] for r in res]); out[f"{s}.ids"] = np.stack([r[1] for r in res]) |
| out[f"{s}.lp"] = np.stack([r[2] for r in res]) |
| nll = -out[f"{s}.target"].mean() |
| print(f"{s}: {len(W)} windows, ppl {math.exp(nll):.4f}, {time.time() - t0:.0f}s", flush=True) |
| np.savez_compressed(a.out, **out, meta=json.dumps({"model": a.model, "topk": a.topk, |
| "eval_set_sha256": es["meta"]["sha256_ids"]})) |
| print("wrote", a.out) |
|
|
