summaryrefslogtreecommitdiff
path: root/ep_run/muon.py
diff options
context:
space:
mode:
authorYuren Hao <yurenh2@illinois.edu>2026-08-05 09:09:14 -0500
committerYuren Hao <yurenh2@illinois.edu>2026-08-05 09:09:14 -0500
commit807c6ba3bab8d9afe18ca7d8fb67038bc2cfaf73 (patch)
tree39122335f5c283685838cf3780718d2d85825cb1 /ep_run/muon.py
parent68ac1f9d1fa8944a216964479e49e813de6370ac (diff)
RESULT 90: 优化器筛选收敛 — Muon 4.7227 领跑, NS消融判决(摘NS代价0.45, nanoGPT版未调用NS之谜归他们), sign家族聚在5.13-5.25; k价值曲线在跑; 曲率通道死刑+偏置公理入档
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
Diffstat (limited to 'ep_run/muon.py')
-rw-r--r--ep_run/muon.py6
1 files changed, 5 insertions, 1 deletions
diff --git a/ep_run/muon.py b/ep_run/muon.py
index 1289620..7bad679 100644
--- a/ep_run/muon.py
+++ b/ep_run/muon.py
@@ -177,7 +177,11 @@ def build_alt(opt_name, blocks, other_params, lr, warmup, total_steps=0, lr_min_
'signline': lambda: SignLine(mats, lr=lm, wd=wd),
'conslion': lambda: ConsensusLion(mats, lr=lm, wd=wd),
'ditherlion': lambda: DitherLion(mats, lr=lm, wd=wd),
- 'cautlion': lambda: CautiousLion(mats, lr=lm, wd=wd)}[opt_name]()
+ 'cautlion': lambda: CautiousLion(mats, lr=lm, wd=wd),
+ 'olionns': lambda: OLion(mats, lr=lm, wd=wd, ns_steps=0),
+ 'olionk1': lambda: OLion(mats, lr=lm, wd=wd, ns_steps=1),
+ 'olionk2': lambda: OLion(mats, lr=lm, wd=wd, ns_steps=2),
+ 'olionk3': lambda: OLion(mats, lr=lm, wd=wd, ns_steps=3)}[opt_name]()
oa = torch.optim.AdamW(rest, lr=lr, weight_decay=1e-4)
opts = [om, oa]
if total_steps > 0: