summaryrefslogtreecommitdiff
path: root/ep_run/lt_ep_train.py
AgeCommit message (Collapse)Author
2026-07-07SPECTRAL GOVERNOR (--governor): the from-scratch recipe candidate, ↵Yuren Hao
dip-map-calibrated Dip screening (4/4 seeds have dips; first-dip cluster 1200-1700, width 100-300 steps, depth |lam|~0.998 = s2000-class; s2 shows excursion->dip in sequence): reg-free until a scout (warm lead_rho, every gov_k=100, deep-400 subbatch) flags rho<0.985 after gov_min=1000, then ARPACK-certify all top-3 |lam|<1 -> leash ON (pair regs). Post-engagement: sustained rho_scout>1.02 x2 -> rescue boost (resreg x3 for 500 steps; the proven hr2 maneuver). abl_delay's fixed-2000 death explained: it engaged AFTER the dip cluster, mid-excursion. Two governor seeds launched on 107 (trained-state engagement is Pascal-safe per hr2 precedent). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-06Tier-3 gates: Anderson math-yes/impl-no (parked for v2); bf16polish UNSAFE ↵Yuren Hao
near-edge (eval-only); dp_ep.py ready Anderson: res 25-35x deeper per budget but 5.8x slower (naive history stacks + per-iter safeguard eval) — v2 = ring buffers + periodic safeguard, est +1.3x on the speed tier. bf16+20polish: 1.41x free phase, res parity, BUT z-diff 1.2e-3 — near-marginal operators contract too slowly for a 20-step polish (0.998^20≈0.96), same magnitude as the TF32 kill verdict and the estimator's 50%-sensitivity input. Predicted by our own depth/noise theory. Flags kept with warnings; neither ships for training. dp_ep.py: manual-allreduce EP DP (controller in lockstep, aligned collectives), smoke pending freed 1080s. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-06aggregate speed bench: speed tier hf+sd 1.47x; accuracy tier hf+sd+t80+avg ↵Yuren Hao
(0.91x, cos 0.94, avg free); compile demoted Full-ep_step wall times on quiet A6000 (warm s2000, B24), res parity across all 8 configs. compile only 1.12x at this shape (historical 1.46x was a different workload split); FULL cmp_sdpa saves 4% over eager at t80 — not worth the guard complexity. tforce_sdpa added (flash baked into compiled graph, flag-free so grad paths never see SDPA). bp_lm --tie probe in flight. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-06--holoavg shipped (gate 0.913->0.936 @t2sel160, batch2 +0.051) + farm wave-2 ↵Yuren Hao
tooling trend-aware stop + plateau averaging recovers the semi-convergence victims exactly as predicted. Estimator pack now: holofast+sdpa+t2sel80(+holoavg). Also: bp_lm stdinit/beta2/sched flags (anchor archaeology), dipfarm_freezer, staged tol_sweep.sh (gated on hr2 verdict), bp_sweep.sh. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-05--sdpa: fused flash attention in the no_grad relax loop — 1.45x free ↵Yuren Hao
phase, z* parity 4e-7 Scoped via blk._sdpa set only inside relax()'s loop (grad paths jvp/vjp/resreg keep the manual attention: no forward-mode-through-flash risk). Combined with --holofast: ~1.51x full-step exact-math tier. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-05holofast: exact halved-jvp AEP track — 1.55x nudged phase, gradient-gate ↵Yuren Hao
passed 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 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-04magic-s2000 study: reg_delay/noadaptc flags, 4-arm queue v2, redx trajectory ↵Yuren Hao
audit - lt_ep_train: --reg_delay N (reg-free early phase: resreg/jr/floss/adaptc off for first N steps) + --noadaptc (kill hidden jacreg==0 damping feedback that would pollute single-reg ablation arms) - queue v2: 4 arms delay-first (abl_delay = reg-free 2k -> proven pair) - eig_traj/2/3: ARPACK audit of redx_traj — the run crossed the edge EARLY and oscillated (s1000 rotating-unstable, s1400 excursion mu=+2.1 self-recovered, s2000 the ONLY stable snapshot mu=-0.02, s2100/s2200 already back out) => s2000 is a post-excursion STABILITY-DIP capture, dip width <100 steps; learning survives mild instability (val fell through unstable stretches). lead_rho cold-40 under-reads clusters — NOT a classifier; ARPACK for audits. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-03floss-ept: graded finite-horizon LE penalty (--floss) + three-arm ↵Yuren Hao
from-scratch ablation queue - ep_step: floss block after resreg — unroll q=10 steps past z_T1 on a sub-batch WITH graph, rho_hat = mean per-step delta growth, one-sided relu(rho_hat - 0.995)^2, ramp keyed on (rho_hat - target) NOT resT1 (de-cliffed resreg: same fundamental path-LE quantity, linear early signal), capped at floss fraction of task-grad norm (resreg convention). - smoke: below-target = untouched (cos 1.0000); force-fire = finite grads, capped perturbation (cos 0.9803). - runs/abl3_queue.sh (runner live): waits for free GPU slots (0/1/3, GPU2 excluded), launches abl_floss (floss-only) / abl_resreg (resreg-only, never cleanly run) / abl_pair (proven 2.09 recipe, control) with identical remaining flags + seed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-03omega/norm-family refuted as stability signal; fingerprint story retracted; ↵Yuren Hao
eigreg v2 = true map-eigenvalue (spec_penalty) - eig_control: fix plain-PI bug (shifted PI for lambda_max of indefinite Sym); add lead_rho + spec_penalty (soft one-sided cap on |lam|(I+eps*J_F), 2-D Rayleigh-Ritz, matvec-only) — aep 'spectral' ported. eig_penalty demoted to diagnostic. - eig_recheck.py (Lanczos audit): omega=+5..+13 on ALL operators incl the stablest (s2000 +12.8 while true alpha=-0.02); gap omega-alpha~10; old 'warm -10.14 vs scratch +1.11' numbers were PI-mixture artifacts. RETRACTED. - eig_v2_smoke/depth: v2 mechanics validated vs ARPACK; z_T1 readings >1 are unconverged-state contamination (150: 1.009 -> 400/800: 0.997-0.999, mu=-0.02..-0.006 matching eig_probe); fixed-point top = BAND of slow modes. - lt_ep_train: --eigreg now spec_penalty (--eig_margin 0.995 = rho target); --fingerprint reports rho/Re_mu instead of num_abscissa. - ONBOARDING §4-7 + FINDINGS 2026-07-03: retraction + verdict (fundamental quantity = finite-horizon path LE / resreg axis; de-cliff via floss-ept; spec_penalty = measure-mode scalpel for a detaching Hopf pair). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-03Initial commit: ept — backprop-free equilibrium transformer (EP)Yuren Hao
Code (ep_run/), organized docs (docs/{method,campaign,hardware,outreach,paper}), analysis scripts (scripts/), ONBOARDING.md entry point. Large data/checkpoints git-ignored (share separately). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
generated by cgit v1.2.3 (git 2.25.1) at 2026-08-18 21:13:56 +0000