diff options
| author | Yuren Hao <yurenh2@illinois.edu> | 2026-07-20 07:10:41 -0500 |
|---|---|---|
| committer | Yuren Hao <yurenh2@illinois.edu> | 2026-07-20 07:10:41 -0500 |
| commit | c700768be62191a038137faec49b1016c7cdb09f (patch) | |
| tree | 5fef2c3772ae15e1636769ca0b169754c15152ed | |
| parent | 6db80927045e6c8ca91df9dd505f07463cc3d73d (diff) | |
棘轮模式(用户设计07-20): beta_simple 1.0=只降不升(越界折半持久,上探×1.0自然无操作,5处>=比较符补丁全断言); 阶梯从上方跟随下沉上界,无形状知识无标定可迁移; fw72m_ratchet王冠排GPU0与bsign并行,两个1.0×候选赛跑
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
| -rw-r--r-- | ep_run/casc_eq_train.py | 10 |
1 files changed, 5 insertions, 5 deletions
diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py index 86a8e06..09a0132 100644 --- a/ep_run/casc_eq_train.py +++ b/ep_run/casc_eq_train.py @@ -497,7 +497,7 @@ def ep_step(x, y): GOV['prev_floor'] = fl beta_t = beta_t * GOV.get('cap', 1.0) # wall-2 loop-gain cap OVERRIDES the floor (the ceiling # can sit below the floor near the wall; survival first) - if args.beta_simple > 1.0: + 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) if args.bsign_rand and torch.rand((), generator=BGEN).item() < 0.5: beta_t = -beta_t @@ -590,7 +590,7 @@ def ep_step(x, y): ok_retry = True if args.beta_ride > 1.0: # the failed trial IS the ceiling measurement GOV['cap'] = max(GOV.get('cap', 1.0) * 0.5 ** (_h + 1), args.cap_floor) - if args.beta_simple > 1.0: # PERSIST the halvings (the ceiling just measured) + if args.beta_simple >= 1.0: # PERSIST the halvings (the ceiling just measured) GOV['bsimp'] = GOV.get('bsimp', 1.0) * 0.5 ** (_h + 1) break if ok_retry: @@ -604,7 +604,7 @@ def ep_step(x, y): drift = _drift(zp, zs_free) gdrift = ddp_max_scalar(drift) ok_retry = _legal(gdrift) - if ok_retry and args.beta_simple > 1.0: + if ok_retry and args.beta_simple >= 1.0: # CODEX FIX: K=8 rescue ran at beta/2^beta_sync — persist those halvings too, # else the next step jumps straight back to the failed beta (upward bias) GOV['bsimp'] = GOV.get('bsimp', 1.0) * 0.5 ** args.beta_sync @@ -618,7 +618,7 @@ def ep_step(x, y): GOV['drift'] = gdrift if args.drift_adapt > 0: GOV['drift_ema'] = 0.95 * GOV.get('drift_ema', gdrift) + 0.05 * gdrift - if args.beta_cap_rho > 0 and GOV.get('rho') is not None and args.beta_simple <= 1.0: + if args.beta_cap_rho > 0 and GOV.get('rho') is not None and args.beta_simple < 1.0: rho_g = ddp_bcast_scalar(GOV['rho']) # rank0's meter rules (identical control on all ranks) res_g = ddp_bcast_scalar(GOV.get('res', 0.0)) # v2: ABSOLUTE-SCALE GATE — rho is only meaningful when the residual is above the noise @@ -718,7 +718,7 @@ def ep_step(x, y): return free_ce, beta_t, GOV.get('kuse', GOV['K']), False for p, g in zip(all_params, gs): p.grad = g - if args.beta_simple > 1.0 and _ok0: + if args.beta_simple >= 1.0 and _ok0: # CODEX FIX: probe upward only on steps that actually COMMIT (after the second-pass # drift and gn-EMA guards) — the old placement raised beta even on later-rejected steps GOV['bsimp'] = GOV.get('bsimp', 1.0) * args.beta_simple |