diff options
| author | Yuren Hao <yurenh2@illinois.edu> | 2026-07-06 07:15:30 -0500 |
|---|---|---|
| committer | Yuren Hao <yurenh2@illinois.edu> | 2026-07-06 07:15:30 -0500 |
| commit | 488c50e1bdbf8f420b2ad5b4021a7f950d121835 (patch) | |
| tree | 3605b3c92af8fd7c0c18ab5982bee084a858b015 /ep_run/lt_ep_train.py | |
| parent | 8f43de671cf03ee20d5d652b6cd3ba575982a436 (diff) | |
--holoavg shipped (gate 0.913->0.936 @t2sel160, batch2 +0.051) + farm wave-2 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
Diffstat (limited to 'ep_run/lt_ep_train.py')
| -rw-r--r-- | ep_run/lt_ep_train.py | 7 |
1 files changed, 6 insertions, 1 deletions
diff --git a/ep_run/lt_ep_train.py b/ep_run/lt_ep_train.py index 1307702..2807be3 100644 --- a/ep_run/lt_ep_train.py +++ b/ep_run/lt_ep_train.py @@ -192,7 +192,10 @@ def ep_step(blk, idx, y, T1, T2, eps, beta, jacreg=0.0, holo=0, hr=0.02, t1max=0 acc = None for _ in range(K): if getattr(blk, 'track', False): # common-mode-tracking AEP (loose-tolerant) - _tr = holo_a_track_fast if getattr(blk, 'holofast', False) else holo_a_track + if getattr(blk, 'holoavg', False): # trend-aware stop + plateau averaging (semi-convergence fix) + from holo_ep import holo_a_track_avg as _tr + else: + _tr = holo_a_track_fast if getattr(blk, 'holofast', False) else holo_a_track ai, _ = _tr(blk, zs, xin0, y, hr, t2sel, eps) # fast = exact halved-jvp mirror (1.55x) else: ai, _ = holo_a_select2(blk, zs, xin0, y, hr, t2sel, eps, li=getattr(blk, 'li_avg', 0)) @@ -429,6 +432,7 @@ def main(): ap.add_argument('--track', action='store_true') # common-mode-tracking AEP correction ap.add_argument('--holofast', action='store_true') # exact halved-jvp track (1.55x nudged phase; parity = FD noise floor) ap.add_argument('--sdpa', action='store_true') # fused flash attention in the no_grad relax loop + ap.add_argument('--holoavg', action='store_true') # trend-aware stop + plateau-avg track (gate: 0.913->0.936 @t2sel160) ap.add_argument('--rt_final', type=float, default=0.0) # anneal res_target to this (0=off), 25%-75% of run ap.add_argument('--nudge_brake', type=float, default=0.0) # kappa: anchor spring during nudge (Tikhonov adjoint) ap.add_argument('--init_ckpt', type=str, default='') # warm-start weights from a saved ckpt @@ -505,6 +509,7 @@ def main(): blk.track = cfg.track blk.holofast = cfg.holofast blk.sdpa = cfg.sdpa + blk.holoavg = cfg.holoavg blk.nbrake = cfg.nudge_brake blk.qknorm = cfg.qknorm if cfg.resinit != 1.0: # near-identity block at init (contractive) -> stable big-width start |