"""Parity + timing for holo_a_track_fast (exact halved-jvp restructure) vs holo_a_track.
Same ckpt (s2000), same batch, same T2max: a_best must match to fp noise, t_best exactly;
timing over 3 reps each (GPU contended — relative ratio is the signal)."""
import time, torch
import lt_ep_train as L
from holo_ep import holo_a_track, holo_a_track_fast

torch.manual_seed(0)
blk = L.EQBlock(512, 16, 256, 256, c=1.0, attn_mode='thick'); blk.qknorm = True
ck = torch.load('runs/redx_traj/s2000.pt', map_location=L.dev)
with torch.no_grad():
    for p, w in zip(blk.allp, ck['allp']):
        p.copy_(w.to(L.dev))
torch.manual_seed(42)
idx, y = L.get_batch('train', 24, 256)
xin = blk.embed(idx).detach()
zs = L.relax(blk, xin.clone(), xin, 150, 0.1)

r, T2, eps = 0.02, 40, 0.1
a0, t0 = holo_a_track(blk, zs, xin, y, r, T2, eps)          # warmup + reference
a1, t1 = holo_a_track_fast(blk, zs, xin, y, r, T2, eps)
rel = ((a1 - a0).norm() / (a0.norm() + 1e-12)).item()
cos = torch.nn.functional.cosine_similarity(a0.flatten(), a1.flatten(), dim=0).item()
print(f"parity: rel_diff={rel:.2e} cos={cos:.8f} t_best {t0} vs {t1}", flush=True)

for name, fn in (('orig', holo_a_track), ('fast', holo_a_track_fast)):
    ts = []
    for _ in range(3):
        torch.cuda.synchronize(); t = time.time()
        fn(blk, zs, xin, y, r, T2, eps)
        torch.cuda.synchronize(); ts.append(time.time() - t)
    print(f"{name}: {min(ts):.3f}s (best of 3)", flush=True)
print("HOLO_FAST_PARITY_DONE", flush=True)