From b83947778e2c776f757a07d4719b7ce961d7ed55 Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Fri, 3 Jul 2026 05:56:50 -0500 Subject: =?UTF-8?q?Initial=20commit:=20ept=20=E2=80=94=20backprop-free=20e?= =?UTF-8?q?quilibrium=20transformer=20(EP)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Code (ep_run/), organized docs (docs/{method,campaign,hardware,outreach,paper}), analysis scripts (scripts/), ONBOARDING.md entry point. Large data/checkpoints git-ignored (share separately). Co-Authored-By: Claude Opus 4.8 Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn --- ep_run/spec_rho_vs_c.py | 29 +++++++++++++++++++++++++++++ 1 file changed, 29 insertions(+) create mode 100644 ep_run/spec_rho_vs_c.py (limited to 'ep_run/spec_rho_vs_c.py') diff --git a/ep_run/spec_rho_vs_c.py b/ep_run/spec_rho_vs_c.py new file mode 100644 index 0000000..47661e0 --- /dev/null +++ b/ep_run/spec_rho_vs_c.py @@ -0,0 +1,29 @@ +import torch, math, pickle +from pathlib import Path +import lt_ep_train as L +from lt_ep_train import EQBlock +L.DD=Path('data/tinystories_bpe'); L.vocab=pickle.load(open(L.DD/'meta.pkl','rb'))['vocab_size'] +dev='cuda'; eps=0.1; B=8; T=256; N=800 +torch.manual_seed(1234); idx,y=L.get_batch('val',B,T) +idx=idx.to(dev) if hasattr(idx,'to') else idx +def measure_c(ckpt,c): + blk=EQBlock(512,16,256,256,s=1.0,c=c,attn_mode='thick'); blk.qknorm=True + ck=torch.load(ckpt,map_location=dev) + with torch.no_grad(): + for p,w in zip(blk.allp,ck['allp']): p.copy_(w.to(dev)) + xin=blk.embed(idx).detach(); z=xin.clone(); ress=[] + for t in range(N): + z2=z+eps*blk.force(z,xin).detach() + r=(z2-z).norm().item()/(z.norm().item()+1e-9); ress.append(r); z=z2 + if not math.isfinite(r) or r>1e2: break + win=[ress[i] for i in range(len(ress)) if 1e-60] + rho=math.exp(sum(math.log(x) for x in rats)/len(rats)) if rats else float('nan') + return rho, ress[-1] +print("=== rho vs damping c — does more c pull the operator off the rho=1 threshold? ===") +print("(weights trained at c=1; this is eval-time c — a margin indicator, not the trained answer)") +for ck,lab in [('runs/redx_traj/s3200.pt','redx-s3200 (val2.74, marginal)'),('runs/bptt_final.pt','BPTT (1.83)')]: + for c in [1.0,1.5,2.0,3.0,4.0]: + try: rho,fr=measure_c(ck,c); print(f" {lab} c={c}: rho={rho:.4f} final_res={fr:.2e}") + except Exception as e: print(f" {lab} c={c}: ERR {repr(e)[:60]}") +print("=== DONE ===") -- cgit v1.2.3