diff options
| author | Yuren Hao <yurenh2@illinois.edu> | 2026-07-24 18:52:12 -0500 |
|---|---|---|
| committer | Yuren Hao <yurenh2@illinois.edu> | 2026-07-24 18:52:12 -0500 |
| commit | 27560970e2d967e62d22b8693811d7301fbc290d (patch) | |
| tree | eefd61a1eedfcc426f0611608e3ddedc5db5094c /ep_run/casc_eq_train.py | |
| parent | c47abf122a2074250ff565b41f3cb3d47bf9b512 (diff) | |
RESULT 60: s345000走廊空(end075更糟3.4805,LR嫌疑排除)→闭合点275-300k,s250000=最后健康点; --res_gate宽度可配+bsign×beta_simple符号抹除bug修复(COMBO_CERT_PASS); fw135m_ratchet从s250000发射(bsign+棘轮+res_gate 1e-3,27h)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
Diffstat (limited to 'ep_run/casc_eq_train.py')
| -rw-r--r-- | ep_run/casc_eq_train.py | 7 |
1 files changed, 6 insertions, 1 deletions
diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py index eb64082..60b6113 100644 --- a/ep_run/casc_eq_train.py +++ b/ep_run/casc_eq_train.py @@ -49,6 +49,10 @@ ap.add_argument('--bsign_rand', action='store_true') # random-sign beta per ste ap.add_argument('--bf16', action='store_true') # cast model to bf16 (E-accumulation + tok_sigma stay fp32) — the x0.5 cost lever, GATE before production ap.add_argument('--amp', action='store_true') # PROPER mixed precision: autocast(bf16) matmuls, fp32 params/states/d/E — amp_gate.py PASSED 2026-07-12 (cos 0.9682 vs fp32 0.9687); --bf16 naive-cast stays DEAD (state quantization, RESULT 11) ap.add_argument('--dtop_every', type=int, default=1) # 1 = exact (DEFAULT, BP-parity); 2 = fast mode (~20% cheaper, ~4% CE tax at high lr) +ap.add_argument('--res_gate', type=float, default=0.02) # legality residual threshold; 0.02 was calibrated + # at C512 — C768 ran half a schedule semi-converged + # UNDER it (R59b). Per-width rule: ~100x the healthy + # K=3 residual floor measured by the rho probe. ap.add_argument('--watch_every', type=int, default=2000) # wandb-only telemetry cadence: act/weight RMS ap.add_argument('--gate_every', type=int, default=200) # in-training cos(EP,BP) telemetry; <=0 = fully BP-free (no bp_gate at all) ap.add_argument('--gate_govern', action='store_true') # let gate cos adjust K/bscale (default: observe-only => training control is BP-free) @@ -503,6 +507,7 @@ def ep_step(x, y): if args.beta_simple >= 1.0: # FULL ownership: no sigma-scaling, no floor/bf_late, no cap — beta = start * measured multiplier beta_t = args.beta * GOV.get('bsimp', 1.0) + # sign flip LAST — the beta_simple override must not wipe it (bsign+ratchet combo fix) if args.bsign_rand and torch.rand((), generator=BGEN).item() < 0.5: beta_t = -beta_t EST = args.est if args.est_late and GOV['step'] >= args.est_late_at: EST = args.est_late @@ -552,7 +557,7 @@ def ep_step(x, y): # ~1 (the documented v1-cap starvation bug; Codex's OR resurrected it, the SV2 # arm caught it: 6-halving storm every step, beta pinned at 4.7e-5). The 35k # runaway had res >> 0.02 with rho < 0.9 — res-alone provably plugs that hole. - if (not math.isfinite(res_s)) or res_s > 0.02: + if (not math.isfinite(res_s)) or res_s > args.res_gate: return False return True _ok0 = _legal(gdrift) |
