diff options
Diffstat (limited to 'ep_run/casc_eq_train.py')
| -rw-r--r-- | ep_run/casc_eq_train.py | 6 |
1 files changed, 5 insertions, 1 deletions
diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py index 0cd5521..07e2ad4 100644 --- a/ep_run/casc_eq_train.py +++ b/ep_run/casc_eq_train.py @@ -49,6 +49,8 @@ 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('--head_lr_mult', type=float, default=1.0) # W_out Adam-group LR multiplier (C768 + # head-throttle compensation, RESULT 66) 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 @@ -313,7 +315,9 @@ if args.opt == 'muon': from muon import build_hybrid opt, sched = build_hybrid(blocks, all_params, args.lr, args.muon_lr, args.warmup, muon_mom=args.muon_mom, adam_b1=args.adam_b1, - total_steps=(args.steps if args.cosine else 0), lr_min_ratio=args.lr_min_ratio) + total_steps=(args.steps if args.cosine else 0), lr_min_ratio=args.lr_min_ratio, + head_param=(W_out if args.untie or args.olmo2 else None), + head_lr_mult=args.head_lr_mult) else: if args.wd >= 0: # OLMo2-style grouped decay: linear weights + head decay; embeddings/norm-gains none nodecay = {id(p) for p in tok.parameters()} | {id(p) for p in pos.parameters()} | \ |