From cbecb171b1af77fe6510fd60f1dfa4c7938a20d6 Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Sun, 5 Jul 2026 03:53:57 -0500 Subject: =?UTF-8?q?holofast:=20exact=20halved-jvp=20AEP=20track=20?= =?UTF-8?q?=E2=80=94=201.55x=20nudged=20phase,=20gradient-gate=20passed?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit holo_a_track computed the doubled-batch jvp/vjp on [v0; -v0] at a shared anchor zbar — exact antisymmetric redundancy (phase deviations from the common mode are exact negatives). holo_a_track_fast computes at batch B and mirrors: single-eval parity 6e-7 (exact); trajectory-level 45% divergence SHARED with the original's own FD noise floor (1e-6 state noise -> 49% self-divergence — the 2r=0.04 finite difference amplifies fp noise; training averages it via pema/momentum). Ship gate: cos(EP,BPTT) orig vs fast indistinguishable (0.907/0.912, 0.853/0.853, 0.918/0.918). Timing 6.43->4.16s on the T2=40 nudged phase (contended GPU, relative). --holofast flag, default off; queued ablation arms deliberately stay on orig for fidelity. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn --- ep_run/holo_fast_probe.py | 49 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 49 insertions(+) create mode 100644 ep_run/holo_fast_probe.py (limited to 'ep_run/holo_fast_probe.py') diff --git a/ep_run/holo_fast_probe.py b/ep_run/holo_fast_probe.py new file mode 100644 index 0000000..9387199 --- /dev/null +++ b/ep_run/holo_fast_probe.py @@ -0,0 +1,49 @@ +"""Disambiguate the parity gap: real bug vs FD-amplified fp noise. +(a) single-eval exactness: at a synthetic two-phase state Z=[zs+d, zs-d], compare the full doubled-batch + correction (Jv-JTv) against the halved+mirrored one. Exact math => allclose at fp level. +(b) estimator noise floor: perturb zs by 1e-6 relative and rerun the ORIGINAL holo_a_track — if a_best + moves by ~the same 0.4 rel, the parity gap is the estimator's intrinsic FD sensitivity, not a bug.""" +import torch, torch.func as tf +import lt_ep_train as L +from holo_ep import holo_a_track + +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) +B = zs.size(0) +fnc = lambda zz: blk.nc_force(zz) + +# (a) single-eval exactness +torch.manual_seed(7) +d = 0.02 * torch.randn_like(zs) +Z = torch.cat([zs + d, zs - d], 0) +zbar = 0.5 * (Z[:B] + Z[B:]) +zb2 = torch.cat([zbar, zbar], 0) +v = (Z - zb2).contiguous() +with torch.no_grad(): + _, Jv = tf.jvp(fnc, (zb2,), (v,)) + JTv = tf.vjp(fnc, zb2)[1](v)[0] + corr_full = Jv - JTv + v0 = (Z[:B] - zbar).contiguous() + _, Jv0 = tf.jvp(fnc, (zbar,), (v0,)) + JTv0 = tf.vjp(fnc, zbar)[1](v0)[0] + corr_half = torch.cat([Jv0 - JTv0, -(Jv0 - JTv0)], 0) +rel_a = ((corr_full - corr_half).norm() / (corr_full.norm() + 1e-12)).item() +print(f"(a) single-eval: rel={rel_a:.2e} ({'EXACT — gap is fp/FD noise' if rel_a < 1e-4 else 'REAL BUG'})", flush=True) + +# (b) estimator noise floor of the ORIGINAL +r, T2, eps = 0.02, 40, 0.1 +a_ref, _ = holo_a_track(blk, zs, xin, y, r, T2, eps) +zs_p = zs + 1e-6 * zs.norm() / (zs.numel() ** 0.5) * torch.randn_like(zs) +a_prt, _ = holo_a_track(blk, zs_p, xin, y, r, T2, eps) +rel_b = ((a_prt - a_ref).norm() / (a_ref.norm() + 1e-12)).item() +cos_b = torch.nn.functional.cosine_similarity(a_ref.flatten(), a_prt.flatten(), dim=0).item() +print(f"(b) orig-vs-orig under 1e-6 state noise: rel={rel_b:.2e} cos={cos_b:.6f}", flush=True) +print("HOLO_FAST_PROBE_DONE", flush=True) -- cgit v1.2.3