From 807c6ba3bab8d9afe18ca7d8fb67038bc2cfaf73 Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Wed, 5 Aug 2026 09:09:14 -0500 Subject: =?UTF-8?q?RESULT=2090:=20=E4=BC=98=E5=8C=96=E5=99=A8=E7=AD=9B?= =?UTF-8?q?=E9=80=89=E6=94=B6=E6=95=9B=20=E2=80=94=20Muon=204.7227=20?= =?UTF-8?q?=E9=A2=86=E8=B7=91,=20NS=E6=B6=88=E8=9E=8D=E5=88=A4=E5=86=B3(?= =?UTF-8?q?=E6=91=98NS=E4=BB=A3=E4=BB=B70.45,=20nanoGPT=E7=89=88=E6=9C=AA?= =?UTF-8?q?=E8=B0=83=E7=94=A8NS=E4=B9=8B=E8=B0=9C=E5=BD=92=E4=BB=96?= =?UTF-8?q?=E4=BB=AC),=20sign=E5=AE=B6=E6=97=8F=E8=81=9A=E5=9C=A85.13-5.25?= =?UTF-8?q?;=20k=E4=BB=B7=E5=80=BC=E6=9B=B2=E7=BA=BF=E5=9C=A8=E8=B7=91;=20?= =?UTF-8?q?=E6=9B=B2=E7=8E=87=E9=80=9A=E9=81=93=E6=AD=BB=E5=88=91+?= =?UTF-8?q?=E5=81=8F=E7=BD=AE=E5=85=AC=E7=90=86=E5=85=A5=E6=A1=A3?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn --- ep_run/casc_eq_train.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'ep_run/casc_eq_train.py') diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py index e29048c..d19504d 100644 --- a/ep_run/casc_eq_train.py +++ b/ep_run/casc_eq_train.py @@ -22,7 +22,7 @@ ap.add_argument('--wandb_run', default='') ap.add_argument('--kmax', type=int, default=8) # adaptive fb rounds cap ap.add_argument('--noguard', action='store_true') # diagnosis: skip only non-finite grads ap.add_argument('--untie', action='store_true') # separate readout matrix (untied from tok) -ap.add_argument('--opt', choices=['adamw', 'muon', 'sgdm', 'lion', 'olion', 'adafactor', 'signline', 'conslion', 'ditherlion', 'cautlion'], default='adamw') +ap.add_argument('--opt', choices=['adamw', 'muon', 'sgdm', 'lion', 'olion', 'adafactor', 'signline', 'conslion', 'ditherlion', 'cautlion', 'olionns', 'olionk1', 'olionk2', 'olionk3'], default='adamw') ap.add_argument('--muon_lr', type=float, default=0.02) ap.add_argument('--tok_init', type=float, default=0.0) # >0: init tok/pos with this std (GPT-standard 0.02) ap.add_argument('--compile', action='store_true') # torch.compile each block (free speed where supported) @@ -345,7 +345,7 @@ if args.bf16: if args.untie: with torch.no_grad(): W_out.data = W_out.data.to(torch.bfloat16) print('[bf16] model cast to bfloat16 (E-accum + sigma stay fp32)', flush=True) -if args.opt in ('sgdm', 'lion', 'olion', 'adafactor', 'signline', 'conslion', 'ditherlion', 'cautlion'): +if args.opt in ('sgdm', 'lion', 'olion', 'adafactor', 'signline', 'conslion', 'ditherlion', 'cautlion', 'olionns', 'olionk1', 'olionk2', 'olionk3'): from muon import build_alt opt, sched = build_alt(args.opt, blocks, all_params, args.lr, args.warmup, total_steps=(args.steps if args.cosine else 0), lr_min_ratio=args.lr_min_ratio, -- cgit v1.2.3