"""Combined monitor for ep_sn (hr=0.2 + specnorm 0.9): watch loss + probe cos per ckpt. Fire on: diverge / val<2.0 (cleared wall) / cos<0.82 (specnorm not holding gradient) / exit.""" import time, os, re, subprocess, shutil WD = "/home/yurenh2/ept/ep_run"; os.chdir(WD) LOG, CK, FROZEN, COSLOG, PID = "runs/ep_sn.log", "runs/ep_sn.pt", "runs/ep_sn_cosprobe.pt", "runs/cos_monitor_sn.log", 2428946 BLOG = "runs/bptt_clean.log" def alive(p): try: os.kill(p, 0); return True except Exception: return False def latest(log): try: ls = [l for l in open(log) if l.startswith("step")] except Exception: return None if not ls: return None m = re.search(r"step (\d+)/.*val CE ([\d.eE+-]+).*res=([\d.eE+-]+)", ls[-1]) return (int(m.group(1)), float(m.group(2)), float(m.group(3))) if m else None open(COSLOG, "a").write("# ep_sn monitor (hr=0.2 + specnorm 0.9)\n") last = -1; fired = None; t0 = time.time() while fired is None and time.time() - t0 < 18 * 3600: time.sleep(120) if not alive(PID): fired = f"ep_sn EXITED; last {latest(LOG)}"; break d = latest(LOG) if not d: continue step, val, res = d if res > 0.2 or val > 15: fired = f"ep_sn DIVERGED step {step} val {val:.2f} res {res:.2e} — specnorm did NOT prevent it"; break if val < 2.0: fired = f"ep_sn reached val {val:.4f} step {step} res {res:.2e} — CLEARED the wall (hr+specnorm worked)"; break if step >= last + 450 and os.path.exists(CK) and os.path.getsize(CK) > 1e6: try: shutil.copy2(CK, FROZEN) except Exception: continue last = step env = dict(os.environ, CUDA_VISIBLE_DEVICES="0", PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True") cosv = "?" try: r = subprocess.run(["python3", "asym_probe.py", "--ckpt", FROZEN, "--B", "8"], env=env, capture_output=True, text=True, timeout=600) m = re.search(r"cos\(g_EP, ?g_transpose\)=([+-][0-9.]+)", r.stdout + r.stderr) cosv = float(m.group(1)) if m else "?" except Exception as e: cosv = f"err:{e}" line = f"step {step}: cos={cosv} val={val:.4f} res={res:.2e}" open(COSLOG, "a").write(line + "\n"); print(line, flush=True) if isinstance(cosv, float) and cosv < 0.82: fired = f"ep_sn COS DEGRADED to {cosv:.3f} step {step} (res {res:.2e}) — specnorm not holding the gradient"; break print("=== EP_SN MONITOR FIRED ==="); print("trigger:", fired or "18h timeout") print("ep_sn:", latest(LOG), "| bptt:", latest(BLOG))