import torch, pickle, math
from pathlib import Path
import lt_ep_train as L
from lt_ep_train import EQBlock
L.DD=Path('data/tinystories_bpe'); L.vocab=pickle.load(open(L.DD/'meta.pkl','rb'))['vocab_size']
dev='cuda'; eps=0.1; B=8; T=256; N=3000
torch.manual_seed(1234); idx,y=L.get_batch('val',B,T); idx=idx.to(dev) if hasattr(idx,'to') else idx
ck=torch.load('runs/redx_traj/s3200.pt',map_location=dev)
def relax_floor(alpha):
    blk=EQBlock(512,16,256,256,s=1.0,c=1.0,attn_mode='thick'); blk.qknorm=True
    with torch.no_grad():
        for p,w in zip(blk.allp,ck['allp']): p.copy_(w.to(dev))
        blk.WO.mul_(alpha)                      # scale attention OUTPUT contribution (alpha=0 -> no attention)
        xin=blk.embed(idx).detach(); z=xin.clone(); ress=[]
        for t in range(N):
            z2=z+eps*blk.force(z,xin).detach()
            r=(z2-z).norm().item()/(z.norm().item()+1e-9); ress.append(r); z=z2
            if not math.isfinite(r) or r>1e3: return ('DIVERGED',t,r,0)
    tail=ress[-800:]
    return (ress[149] if len(ress)>149 else None, ress[-1], min(tail), max(tail))
print("=== knockout: scale attention output (WO*alpha) on redx s3200, eval_relax 3000 steps ===")
print("alpha=1 is the cycling operator; if cycle dies (res->0, monotone) as alpha falls -> attention asymmetry drives it")
for a in [1.0, 0.7, 0.4, 0.2, 0.0]:
    r150,rlast,tmin,tmax=relax_floor(a)
    osc = (tmax-tmin)
    print(f"  alpha={a}: res(150)={r150:.3e}  res(3000)={rlast:.3e}  tail[min={tmin:.2e},max={tmax:.2e}] osc={osc:.2e}  {'CYCLE' if osc>1e-3 and rlast>1e-3 else 'converged' if rlast<1e-3 else 'floored'}")
print("=== DONE ===")