diff options
Diffstat (limited to 'ep_run/resreg_warm_probe_loop.py')
| -rw-r--r-- | ep_run/resreg_warm_probe_loop.py | 49 |
1 files changed, 49 insertions, 0 deletions
diff --git a/ep_run/resreg_warm_probe_loop.py b/ep_run/resreg_warm_probe_loop.py new file mode 100644 index 0000000..21c73a2 --- /dev/null +++ b/ep_run/resreg_warm_probe_loop.py @@ -0,0 +1,49 @@ +import time, os, re, torch, pickle, numpy as np, subprocess +from pathlib import Path +from scipy.sparse.linalg import eigs, LinearOperator +import lt_ep_train as L +from lt_ep_train import EQBlock, relax +os.chdir("/home/yurenh2/ept/ep_run") +L.DD=Path('data/tinystories_bpe'); L.vocab=pickle.load(open(L.DD/'meta.pkl','rb'))['vocab_size'] +dev='cuda'; B=24; T=256; eps=0.1 +CK='runs/ep_resreg_warm.pt'; LOG='runs/ep_resreg_warm.log'; OUT='runs/resreg_warm_probe.log' +torch.manual_seed(0); idx,y=L.get_batch('val',B,T); idx=idx.to(dev) if hasattr(idx,'to') else idx + +def alive(): return subprocess.run(["pgrep","-f","ckpt runs/ep_resreg_warm.pt"],capture_output=True).returncode==0 +def curinfo(): + try: + ls=[l for l in open(LOG) if l.startswith("step")] + m=re.search(r"step\s+(\d+)",ls[-1]); b=re.search(r"best ([\d.]+)",ls[-1]); v=re.search(r"val CE ([\d.]+)",ls[-1]) + return (m.group(1), b.group(1), v.group(1)) + except Exception: return ("?","?","?") +def probe(): + ck=torch.load(CK,map_location=dev) + blk=EQBlock(512,16,256,256,s=1.0,c=1.0,attn_mode='thick'); blk.qknorm=True; blk.fnoise=0.0 + with torch.no_grad(): + for p,w in zip(blk.allp,ck['allp']): p.copy_(w.to(dev)) + xin=blk.embed(idx).detach() + z=relax(blk,xin.clone(),xin,300,eps).detach() + g=(blk.force(z,xin).detach().norm()/z.norm()).item() + F0=blk.force(z,xin).detach(); h=1e-3*z.norm().item(); N=z.numel() + def Jv(v): + vt=torch.from_numpy(v).float().to(dev).view_as(z); nv=vt.norm().item() + if nv<1e-20: return np.zeros_like(v) + return ((blk.force(z+h*(vt/nv),xin).detach()-F0)/h*nv).view(-1).cpu().numpy() + lam=eigs(LinearOperator((N,N),matvec=lambda v:v+eps*Jv(v),dtype=np.float32),k=6,which='LM',return_eigenvectors=False,maxiter=300,tol=1e-3) + mu=(lam-1)/eps; top=mu[np.argmax(mu.real)] + return g, float(top.real), float(abs(top.imag)) + +t0=time.time() +hdr="=== resreg_warm g/Reμ probe loop (every 20min) — track if g holds (bounded) or grows (toward blow) ===" +print(hdr,flush=True); open(OUT,'a').write(hdr+"\n") +while time.time()-t0 < 24*3600: + if not alive(): + line=f"[+{int(time.time()-t0)}s] ep_resreg_warm DEAD/done -> stop"; print(line,flush=True); open(OUT,'a').write(line+"\n"); break + try: + st,bb,vv=curinfo(); g,re_mu,im_mu=probe() + line=f"[+{int((time.time()-t0)/60)}min] step{st} best{bb} valCE{vv} | g_floor={g:.2e} Re_mu={re_mu:+.3f} Im={im_mu:.3f}" + print(line,flush=True); open(OUT,'a').write(line+"\n") + except Exception as e: + print("probe skip:",repr(e),flush=True) + time.sleep(1200) +print("loop end",flush=True) |
