summaryrefslogtreecommitdiff
path: root/ep_run
AgeCommit message (Collapse)Author
2026-07-14wandb -> team entity eqprop-llm-training (trainers + full-history replay script)Yuren Hao
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-14bcap v2: absolute-scale gate on rho meter (fixes v1 noise-floor starvation), ↵Yuren Hao
gentler attack 0.85/floor 0.05; crown c2 relaunched from s215000 with cap armed Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-14beta-cap-by-loop-gain controller (--beta_cap_rho): rho^ meter from sweep ↵Yuren Hao
residuals, cap OVERRIDES floor near the wall; smoke passed Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-14Adaptive relax (--relax_tol): sweep-to-tolerance + rho-triggered geta ↵Yuren Hao
backtracking + final full-step graphed round (E-read identity fix: mixing leaked iteration residual into E, gn 1e5); kuse telemetry; paired smoke at wall ckpt: adaptive cos 0.771 vs legacy 0.754 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13Estimator arms machinery: --est centered/richardson (two-pass, O(b^2) bias), ↵Yuren Hao
--muon_mom/--adam_b1 knobs; smoke passed (resume+centered cos 0.998) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13Fix swallowed --wandb_run arg (replace-comment bug); wandb default-on intactYuren Hao
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13wandb ON by default in both trainers (project ept-cascade); BP twin gains ↵Yuren Hao
--amp; fw72m_bp launched GPU1 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13fw72m crown run launched (NCCL DDP prod), exact resume (opt state), RESULT ↵Yuren Hao
17, shuffle+prep scripts Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13Manual DDP for cascade-EP: guard-synced (drift/kretry via global max, gn on ↵Yuren Hao
reduced grads), sigma broadcast, per-rank data RNG, None-preserving flat all-reduce, bitwise param sync-check, grad-equivalence test PASSED (cos 0.999999999 vs single-GPU big batch) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13Stage-2 data pipeline: FineWeb-Edu sample-10BT -> 32k BPE -> bins; --data ↵Yuren Hao
flag on both trainers Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-12bf16 mixed-precision (--amp): gate PASSED (cos 0.9682 vs fp32 0.9687), ↵Yuren Hao
trainer flag + 3-seed 4k A/B launched Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-12E-tier wave-1 ledger (RESULT 14 + HW map rows) + cost model v2 recalibrated ↵Yuren Hao
on measured H200 datapoint Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-12RESULT 13: bsign neutral at 4k (1.7382 vs 1.7314); 42M gap-chasing closed at ↵Yuren Hao
noise floor; pivot to Stage-2 pipeline + E-tier
2026-07-11RESULT 11: stage1b sealed (gap 0.050); beta ceiling unfound (3e-3 -> cos ↵Yuren Hao
0.9974); K exonerated on OLMo2; bf16 naive-cast dead at any beta (TF32 stays cost baseline); f3e3 CE-endpoint cont launched Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-10RESULT 10: epoch endpoints (BP 1.2509/1.2750 vs EP 1.4802, +0.22 erosion ↵Yuren Hao
cost); GENERATION GATE PASSED (casc_gen.py, coherent stories, no-backprop 42.75M); stage1b improved-recipe pair launched (Muon cosine via build_hybrid) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-10Sustained dose-response sealed (3e-4/5e-4/1e-3 -> 0.975/0.983/0.990); ship ↵Yuren Hao
--bf_late floor ramp; next-run recipe locked (OLMo2+Muon+late-ramp)
2026-07-10RESULT 6: wall-2 = EP-specific marginal under-convergence at the ↵Yuren Hao
contractivity edge (A skips16/B skips2/C skips1/D-BP clean); ship --kretry (retry marginal batches at K=kmax); launch OLMo2 parity matrix with parity-gated epoch autolaunch Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-10OLMo2-standard architecture (user directive): norm-after-sublayer RMSNorm ↵Yuren Hao
blocks, full-width QK-norm, RoPE(500k), SwiGLU, no-bias, untied head, final RMSNorm, 0.02 init, grouped wd, optional z-loss (plumbed through nudge+readout+gate for EP exactness); EP smoke 42.75M cos=1.0000 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-10AUDIT: retract confounded Muon verdict; tone down parity claims (n=3, ↵Yuren Hao
best-of-noisy-val); scope depth-tax claim to 4k horizon; guard-split skip telemetry (skd/skg); add BP control arm diag_D_bp (+--resume in BP trainer); record resume confounds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-10Stage-1 epoch blowup @12100: sig story REFUTED (only +8%, cos fine till ↵Yuren Hao
after); leading indicator=drift-guard skips -> contractivity bifurcation in nudged relaxation (cascade Hopf wall); add resume/sig0/final_ln; A/B/C diagnostic launched
2026-07-10QK-norm: replace nn.MHA with explicit SDPA attn + --qk_norm (OLMo2-style, ↵Yuren Hao
analog-friendly); cancel epoch, insert 8-run qk validation, stage roadmap TinyStories-epoch->FineWeb-Edu->OLMo2
2026-07-10Add --cosine (warmup->cosine decay to 0.1x lr) to both cascade trainers for ↵Yuren Hao
full-epoch runs
2026-07-10K refuted as L12 lever (cos K-invariant K3==K8); BP 3-seed sealed 1.919; ↵Yuren Hao
mechanism = finite-beta SNR collapse; add --beta_floor/--beta_fixed + launch floor sweep Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09BP-free formal audit (5.6-sol): EP update verified local (108 bitwise ↵Yuren Hao
zero-influence checks); fixes — gate_every<=0 = true off switch, governor reaction now opt-in (--gate_govern, default observe-only), dFdtheta restored as self-sealing invariant-test surface; test_bp_free.py 4/4 green Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09A0.4 precision gate (TF32 harmless cos 0.9946==fp32; pure-bf16 cos 0.9427) + ↵Yuren Hao
Muon hybrid optimizer (--opt muon) wired into both trainers; D1a flagship matrix launched (L12xC512, 3xBP + 3xEP + Muon arms) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09casc_eq_train: exact mode (dtop_every=1) now DEFAULT — bisection showed ↵Yuren Hao
the v7 dedup cost ~4% CE at tuned lr; fast mode kept as opt-in. K3 verdict: exact equilibrium-EP 2.0481 vs BP 2.053+-0.004 = matched-tuning PARITY Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09casc_eq_train v7: two algorithmic dedups — d_top refreshed every other ↵Yuren Hao
round (O(beta^2) error), last-rebuild graphs reused by the theta-readout (kills one full graphed chain per step); fp32 untouched for fair BP twinning Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09casc_eq_train v6: graph-reuse fb (rebuild graphs feed next round's vjps), ↵Yuren Hao
--compile flag, tok-sigma amortized every 25 steps — 2.23 -> 2.68 it/s on 1080 (3.28x BP) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09cascade root cause found: tied readout + default N(0,1) embedding init = ↵Yuren Hao
pathological top-CE stiffness once predictions sharpen (sigma_tok~76 vs GPT-standard 0.02 giving ~1.6); add --untie and --tok_init; diag arms A/B/C Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09casc_eq_train v4: quality-governed estimator — periodic measured ↵Yuren Hao
cos(EP,BP) drives K/beta (spend when quality drops, relax when abundant); guards reduced to sanity-only Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09casc_eq_train v3: adaptive beta (tok-sigma stiffness compensation), ↵Yuren Hao
adaptive-K fb with contraction test, non-contraction bail, in-training cos(EP,BP) telemetry Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09cascade equilibrium solver solved: fb (forward-backward message passing) K=3 ↵Yuren Hao
beta=0.003 — gates 1.0000/0.9990/0.9946 across BP trajectory, L12 0.9999; trainer v2 single-sided EP readout + divergence guard (~5x BP) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09casc_eq_train: equilibrium-mode cascade-EP trainer (GS-reverse solver K ↵Yuren Hao
sweeps, two-phase ±β, EP readout at relaxed states) — the true-EP route C1 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09cascade: retire zil per user directive (EP route = equilibrium two-phase); ↵Yuren Hao
probe gains --init_sweep (state warm-start, EP-clean readout) and --sopt adam; B2 solver sweep Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09cascade: zil interleaved scheme = exact gradients at all depths (cos 1.0000 ↵Yuren Hao
L6-24, trajectory 0.9998+); casc_ep_train zil trainer; naive-relaxation depth failure documented Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09cascade ablation program: plan doc (5 claims, tiers 0-4), probe v2 ↵Yuren Hao
(jacobi/gsf/gsr schemes, full-theta gate, multi-batch), casc_bp_train ckpt producer Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09cascade_probe: layered-energy EP on L distinct standard blocks — free ↵Yuren Hao
equilibrium == plain forward (standard LLM inference), two-phase grad gate cos 0.9968 vs BP (L3 C128) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09lt_ep_train: --gov_keepjr now keeps the jr floor through reg_delay too ↵Yuren Hao
(fastfull arm: jr always-on, resreg post-warmup) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09backfill_wandb: replay historical run logs into W&B (ept-c512 / ept-tol / ↵Yuren Hao
ept-33m) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09lt_ep_train: --rr_floor (resreg floor during free/delay phase) for govfloor armsYuren Hao
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-09lt_ep_train: optional W&B mirroring (--wandb/--wandb_run), failure-proof ↵Yuren Hao
logging of val/ema/best/res/jr/lr/rho/gov state Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
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
2026-07-06dp_ep smoke PASSED: 2-GPU lockstep clean, ~107% linear scaling (allreduce ↵Yuren Hao
negligible vs seconds-long EP steps) val 6.0->5.83/100 steps at effective B48, res nominal, ranks never diverged. Speed package final ledger: pack 1.47x X DP ~linear (X optional AA-v2 1.3x). dip_screen.py: ARPACK screener for the dip-farm trajectories (running on freed 1080s, 2 chains). 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
2026-07-06bp_lm twin control + env forensics probes + holo_a_track_avg ↵Yuren Hao
(semi-convergence fix, ungated) BP+EMA 2x2 (local, healthy env): qknorm {1.9951, 1.9977}, no-qknorm {1.9728, 1.9732} -> parameterization-matched BP twin tops at ~1.97 vs EP 1.7888: the equilibrium computation's iteration/depth dividend = 0.18 CE from identical parameters. External anchor (tuned depth-1 BP, 1.7921): EP at parity. bp_lm on 107/2.3.1 gave 1.9823 ~= local -> plain backward exonerated on the pascal env; the 107 divergence (pair/floss/resreg dead by step 600) narrows to the EP reg/estimator loop. Single-step fingerprints all match (a/b/c/d) -> suspected intermittent kernel issue; discriminators in flight (torch-2.7 env probe + delay_hr2's step-2000 reg-on transition). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-05estimator forensics closed: cos ceiling = adjoint TRUNCATION + ↵Yuren Hao
SEMI-CONVERGENCE (not r, not anchor, not transients) Refutation chain: r-sweep flat (0.02-0.4); deep anchor (res 100x tighter) no gain; kappa brake monotonically harmful. t2sel window sweep at s2000: 40->80 lifts ALL batches (mean 0.889->0.936, truncation confirmed); past 80 SEMI-CONVERGENT (batch-dependent optimum; the inc-argmin t_best rule fails on rotating slow modes -> batch2 degrades 0.956->0.875 at 320). Early stopping IS the regularizer; iteration count = reg parameter. Shipping insight: holofast + sdpa + t2sel80 ~= old default wall-clock with cos 0.89->0.94. warm_fast (record) already ran t2sel=80 vs proven-scratch 40 — a real +0.05-cos hidden difference in the lineage table. Next lever: trend-aware stopping / t_best-neighborhood averaging. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-05r-sweep: cos(EP,BPTT) FLAT across r=0.02..0.4 on both operators — ↵Yuren Hao
nudge-SNR hypothesis refuted The 50% single-shot sensitivity is MULTIPLICATIVE (near-marginal T2 dynamics' state-sensitivity; signal and error scale with r together) — not additive noise divided by 2r. cos ceiling ~0.88 at near-edge states is set by T2 truncation + state sensitivity (0.98 at deeply-contracted states). hr-0.2's empirical wins are NOT estimator SNR. Hardware upside: algorithm indifferent to r across 20x -> nudge amplitude free to fight ADDITIVE readout noise. Surviving accuracy levers: kappa nbrake (Tikhonov, targets the real culprit), tail-window averaging over argmin selection. 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
generated by cgit v1.2.3 (git 2.25.1) at 2026-08-18 21:52:19 +0000