| # Box calibration: the same 16 HumanEval items, same config (SDAR-1.7B + 128M judge, no pool), old 8x box vs this box. | |
| import glob, json, sys | |
| R = "/workspace/dmge" | |
| def load(pat): | |
| out = {} | |
| for f in glob.glob(pat): | |
| if f.endswith("running.jsonl"): continue | |
| for l in open(f): | |
| if l.strip(): | |
| r = json.loads(l); out[int(r["idx"])] = r | |
| return out | |
| old = load(f"{R}/logs/phase2/ref/S0_128M_nopool_by_set/humaneval/*.jsonl") | |
| new = load(f"{R}/logs/phase3/setup/check_judge/*.jsonl") | |
| ids = sorted(new) | |
| same = sum(old[i]["gen_ids"] == new[i]["gen_ids"] for i in ids) | |
| print(f"paired items {len(ids)}, identical outputs {same}; old gpu {old[ids[0]]['gpu']} utc {old[ids[0]]['utc']}") | |
| def comp(r): | |
| c = {f"moe_{k}": v for k, v in r["ms_by_kind"].items()} | |
| c.update({f"dr_{k}": v for k, v in r["drafter_ms_by_kind"].items()}) | |
| c["other"] = 1000 * r["seconds"] - sum(c.values()) | |
| return c | |
| keys = sorted(comp(new[ids[0]])) | |
| print(f"{'component':14s} {'old ms':>9s} {'new ms':>9s} {'new/old':>8s} counts old/new") | |
| for k in keys: | |
| o = sum(comp(old[i])[k] for i in ids); n = sum(comp(new[i])[k] for i in ids) | |
| if o or n: print(f"{k:14s} {o:9.0f} {n:9.0f} {n/o if o else float('nan'):8.3f}") | |
| to = sum(old[i]["seconds"] for i in ids); tn = sum(new[i]["seconds"] for i in ids); g = sum(new[i]["n_gen_to_eos"] for i in ids) | |
| print(f"total s old {to:.1f} new {tn:.1f} ratio {tn/to:.3f}; tok/s old {g/to:.2f} new {g/tn:.2f}") | |
| print("per-item ratio new/old:", " ".join(f"{new[i]['seconds']/old[i]['seconds']:.2f}" for i in ids)) | |