"""Lightweight cos monitor for ep_hr02: probe each new ckpt, log step->cos(g_EP,exact-adjoint). Fire on: cos degrades <0.90 (gradient going bad) / survived to step>=9500 (cleared old danger zone) / death.""" import time, os, re, subprocess, shutil WD = "/home/yurenh2/ept/ep_run"; os.chdir(WD) LOG, CK, FROZEN, COSLOG, PID = "runs/ep_hr02.log", "runs/ep_hr02.pt", "runs/ep_hr02_cosprobe.pt", "runs/cos_monitor.log", 1684249 def alive(): try: os.kill(PID, 0); return True except Exception: return False def cur_step(): try: ls = [l for l in open(LOG) if l.startswith("step")] if ls: return int(re.search(r"step (\d+)", ls[-1]).group(1)) except Exception: pass return 0 open(COSLOG, "a").write(f"# cos monitor start (ep_hr02, hr=0.2)\n") last = -1; fired = None; traj = []; t0 = time.time() while fired is None and time.time() - t0 < 18 * 3600: time.sleep(120) if not alive(): fired = f"ep_hr02 EXITED at step {cur_step()}"; break step = cur_step() 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, zres = None, "?" try: r = subprocess.run(["python3", "asym_probe.py", "--ckpt", FROZEN, "--B", "8"], env=env, capture_output=True, text=True, timeout=600) out = r.stdout + r.stderr m = re.search(r"cos\(g_EP, ?g_transpose\)=([+-][0-9.]+)", out) zr = re.search(r"z\* residual.*?step_rel=([0-9.eE+-]+)", out) cosv = float(m.group(1)) if m else None zres = zr.group(1) if zr else "?" except Exception as e: zres = f"probe-err:{e}" # also grab current val from the training log val = "?" try: val = re.search(r"val CE ([\d.eE+-]+)", [l for l in open(LOG) if l.startswith("step")][-1]).group(1) except Exception: pass line = f"step {step}: cos={cosv} val={val} z_res={zres}" traj.append((step, cosv, val)); open(COSLOG, "a").write(line + "\n"); print(line, flush=True) if cosv is not None and cosv < 0.82: # below the historical per-batch floor (~0.85) => real degradation, not variance fired = f"COS DEGRADED to {cosv:.3f} at step {step} (val {val}) — below historical floor, real gradient degradation"; break if step >= 9500: fired = f"ep_hr02 SURVIVED to step {step} (val {val}) with cos staying high — cleared the old ~9200 danger zone"; break print("=== COS MONITOR FIRED ==="); print("trigger:", fired or "18h timeout") print("cos trajectory (step | cos | val):") for s, c, v in traj: print(f" {s:6d} | {c} | {v}")