summaryrefslogtreecommitdiff
path: root/ep_run/casc_eq_train.py
AgeCommit message (Collapse)Author
13 days优化器价目表实装: Lion/OLion/Adafactor2D/build_alt + --opt 六选项Yuren Hao
OLion按论文算法框(慢动量+Nesterov混合→NS→逐元素sign→γ=0.2定标+解耦wd; betas取Lion默认 0.9/0.99并注明为假设); Adafactor2D=行/列因子化二阶矩+RMS-1裁剪(07-11审计的"模拟原生Adam"); build_alt与build_hybrid同分割(矩阵走被测规则+其余AdamW), sgdm例外全参数SGD(全局部基线)。 四路CPU冒烟全过(损失下降, gate cos 1.0000)。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-08-04RESULT 89终局 + 交付物: 135M完赛 EP 3.2087 vs BP 3.2074(+0.1% ppl), ↵Yuren Hao
旧读出+45.2% 窗口敏感性诚实记录: 97%窗给-0.0030(seed间距0.0059), 完赛大窗给+0.0013(间距0.0005, cosine 衰减后BP收拢); 以大窗为准且不得称"统计不可分"(2.6倍间距), 正确表述=差0.1%且处分辨极限, seed数(BP n=2/EP n=1)不足以给区间。 交付: --gen模式(复用模型定义, EP/BP同代码路径)、fig_135m_curves(四曲线/横轴token/BP细线在上)、 EPT_135M_samples.ipynb(同prompt同种子并排, 输出已烤入, 代码真可跑)。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-30RESULT 78: ↵Yuren Hao
read_lin梯度级验证==fp64(fp32下泄漏塌零,读出舍入=全部病因); 附属谱仪:前沿随训练下压+seed轴证'比值决定非层号'+阈值带2-5e-7; 修复入训练路径; CE级仪器终审在飞(预注册~+100%关闭) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-30RESULT 77: 地板理论三箭全中(fp64塌零/β平移/β×8痊愈) — ↵Yuren Hao
物理泄漏改判fp32伪影; --read_lin零成本修复实装(读出用满精度d替代舍入的z−o), C768验证在飞; geo电池撤除; dgain重新定性为数值workaround Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-30RESULT 76: 谱仪读出 — ↵Yuren Hao
逐块泄漏剖面(C768上6块-0.70..-0.23,C512块8-11也病)+偶性实锤(-β同号)+disp_ratio阈值跨宽度塌缩(~2e-7≈2×fp32 eps)+fp32地板假说(一举解释全部约束); 预注册双证伪已发射(β-scan平移+fp64决定臂); --probe_f64实装 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-30dgain几何profile + M1泄漏谱仪(用户设计,07-30)Yuren Hao
--dgain_geo r: 每层读出位移增益=r^l(首层恒1,不含总深度L,加深自动延伸); --dgain_geo_cap封顶。 用户提案"首层不变逐层x2"的实装,三臂电池(2^l | 2^l cap128 | 1.414^l)已排队于仪器ckpt。 --probe_dgspec N: 逐块顽固度谱仪。L_l(g)=mean_b[gEP_l(g)-gBP_l]同批配对差分(batch采样噪声 精确对消,残余=一致性偏差+估计器噪声/sqrtN); 主指标=在BP方向的带符号投影(范数在1e7维被噪声 膨胀,冒烟实测g8读700%,投影无此病且恰为致害分量); 附(-beta,g=1)臂分离奇偶(beta奇函数FD偏差 翻号,位移阈值泄漏为|位移|偶函数不翻); 同时记录每块自由相位移比RMS(d_l)/RMS(z_l)=无量纲律 候选横轴。fp32仪器级(bf16地板致盲,b15教训)。GOV['dgprofile']钩子仅探针置位,训练路径不触。 排队: dgspec C768(泄漏态s250000)+C512对照(应读~0) on GPU1, ladder后补扫C320/C192/C128 =5宽度x12块60点做g*_l x 位移比塌缩检验。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-29RESULT 72: 剂量曲线通顶 — ↵Yuren Hao
dg128关闭97%(距all-BP杆0.0015),物理泄漏完全可赎回; 对数律全程成立34→97%; dgrand败于固定档; 满剂量验证fw135m_dg128三卡发射(~22h,判据=终点回72M gap带); 遗留问题=剂量-宽度标度(270M答) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-29RESULT 69: ↵Yuren Hao
饱和盲区被自己的治疗测试证伪(cap30/50关闭0.0000,证书cap=1爆模型); cap价格≈0(用户担忧定价); 存活约束=顶半+attn+1/β+11项免疫; dgain位移臂在飞(cos0.83的代价由CE裁决)+mixffn补分解 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-28RESULT 66: head被节流到60%(位移微探针:全族1.00-1.02唯W_out 0.60) ↵Yuren Hao
— 用户等价LR假说命中并定位到单矩阵; Adam岛m/√v节流,对centered/fp32/K8全不敏感=EP头读共享结构; --head_lr_mult实装; 电池2(headmix诊断+hlr1.67/2.5补偿)在飞 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-24RESULT 60b: ↵Yuren Hao
res表在amp下宽度盲(bf16地板把fp32四个量级压成1.2×,棘轮首步死亡螺旋); 带内自适应在C768+amp判死,仪器=离线fp32探针; fw135m_b15(定点1.5e-3+bsign)+ckpt哨戒探针在飞; res/rho入wandb Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-24RESULT 60: ↵Yuren Hao
s345000走廊空(end075更糟3.4805,LR嫌疑排除)→闭合点275-300k,s250000=最后健康点; --res_gate宽度可配+bsign×beta_simple符号抹除bug修复(COMBO_CERT_PASS); fw135m_ratchet从s250000发射(bsign+棘轮+res_gate 1e-3,27h) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-22遥测扩展(用户规格): wandb新键 gn/clip_norm/clip_fired + ↵Yuren Hao
watch_every(2k)act_rms/w_rms + gate处mag_ep/mag_bp — 打印行格式冻结(sync正则兼容),BP孪生镜像; 冒烟认证LOGCERT_PASS(8键API读回,gn=clip=mag_ep互证) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-22resume NameError修复: ↵Yuren Hao
bsimp恢复写在GOV定义之前(仅新ckpt含bsimp键时触发,135M DDP3迁移三rank齐炸=开火证书log:656-666); 缓冲变量两行修复,GOV定义后应用 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-20棘轮模式(用户设计07-20): beta_simple ↵Yuren Hao
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
2026-07-20RESULT 54: ↵Yuren Hao
上界∝1/σ²下沉(σ242→473翻倍,β*从0.12跌到<0.03),固定β=停滞非安全(plain2收官89%skip,best其实是200k成绩); K=30密探针推翻Codex的β*=0.7(K=8误判慢发散为收敛); σ-scaling原设计对但被floor钉死+早期σ暴涨张力; --beta_cos_min实装(按进程cosine降β跟踪下沉上界); betacos王冠2e-2→8e-4在飞 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-20guards_silent(用户令07-20): ↵Yuren Hao
守卫只报不拦(skip全部放行提交,唯一例外=非有限梯度仍跳),证书流程废除;唯一执法=控制器自身折半;判官=用户盯CE; simplev2王冠直发链排plain2后(--guards_silent --kretry 0,无cap/wsync) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-19门最终修正(SV2验证臂拦截): ↵Yuren Hao
res>0.02单独判非法即封35k逃逸孔(该段res>>0.02,AND靠ρ̂≤0.9放行);ρ̂降级纯遥测——Codex的OR复活了v2注释里记载的噪声地板bug(ρ̂≈noise/noise≈1单独成立→每步6连折半,β钉4.7e-5),验证证书流程拦下;sv2b重验证在跑 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-19CODEX审计修复×4: ↵Yuren Hao
_legal门AND→OR+NaN非法+ddp_max(600步连续放行的逃生孔,主因); kretry救活持久化折半; up-step移到commit点后; bsimp进ckpt — codex裁决:主因=门代码缺陷,AIMD算术无错(0.01423≈ln1.01/ln2,392.37=1.01^600精确),用户'是bug'判定成立 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-19beta_simple(用户规格,复杂机构全拆): ↵Yuren Hao
β=起始×bsimp,干净步×1.01,非法步sync折半持久;σ缩放/floor/bf_late/cap/ride/servo/wsync全部不触β; 开火证书PASS(近边态锯齿0.8-2.4e-3,18刀,爬升段2.7×,零崩溃); fw72m_simple王冠已发射 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-19beta_servo: ↵Yuren Hao
贴上界伺服(亮区反演cap*=servo·cap_rho/ρ̂双向一步落点,暗区复用ride爬升;常数仅0.8安全系数) — 开火证书:plain s195000死亡区2.5k步CERT_PASS(β双向移动25/26行,best 3.5703→3.4260,零崩溃;贴边开销+50%墙钟为最恶劣工况); fw72m_servo王冠从头跑已发射(servo+ride4+cap_floor0+wsync3) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-18governor接线修复:res/rho非法判据常开(beta_cap_rho>0即生效),crow ↵Yuren Hao
n-3的18k垃圾步正是从'表只接beta_sync'这个洞进的账;用户判定=adaptive beta bug,成立 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-18wsync: 同步权重步接受(用户风格重设计,knee撤回) — ↵Yuren Hao
opt.step前快照params+动量,下步nudged relax过同一_legal门,非法即p←(p+snap)/2折半重放(预算复用beta_sync语义),折半到底整步回滚+skip;零新常数,模型函数零改动;42M引爆预检+死亡窗口重放已排c190后 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-18RESULT 49: ↵Yuren Hao
回路解剖=23段乘积近临界(下行×72/上行×13k/顶读1e-6),plain只热b8一段(h6 logit 94 vs cent 75)边际越线,灾变=硬底响应打伤b1(入口vjp 1.3→14.9); batch噪声对照0.39/0.45=Q1封口; --logit_knee 80实装+knee臂排c190后 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-18RESULT 43+44: crown-3 ↵Yuren Hao
死因=cap硬底钉在塌落天花板之上(变相wall-2); endgame A/B平局(3.4245/3.4244)估计器无罪; bpmix定位gap全在上半blocks(77%)attn/ffn均分,底半=0; --cap_floor补丁+fw72m_plain_nf无硬底190k重放在飞 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-18--bpmix: gap-source ablation (overwrite EP grads with true BP grads per ↵Yuren Hao
module group); 5-arm endgame ablation chained after A/B Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-17--beta_sync: synchronous acceptance (this-step relax telemetry gates the ↵Yuren Hao
commit; reject -> halve beta, retry same batch; failed trial doubles as ceiling measurement for the ride cap) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-17ride-v2 (b)/(c) flag-gated (--ride_ema/--ride_cool, default off); isolation ↵Yuren Hao
probe design: control|a|a+b|a+c|a+d|full at 72M mid-stage Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-17ride-v2: (a) floor-jump cap continuity (the 0.09 bug), (b) rho-EMA ↵Yuren Hao
decisions, (c) 200-step post-attack cooldown, (d) adaptive drift ceiling (--drift_adapt); smoke: continuous beta across floor jump Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-17save final ckpt at last step (fixed-eval protocol); fw72m_ride crown-3 ↵Yuren Hao
launched (plain+ride governor); ride/BP seed campaign on GPU0 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-16--beta_ride ceiling-riding governor (two-sided bcap: climb in calm, attack ↵Yuren Hao
near threshold); smoke: 3e-3 -> 2.4e-2 hover at ride cap, rho calm at 42M tail Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-16RESULT 29: qcomp8 tax-free (+0.0018) = T64 8-bit compute green light; ↵Yuren Hao
--centmirror ships (cos 1.000000000, 1.72x -> 1.39x) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-16Stage-0 gate #1b: --qcomp_bits (compute on DAC-grid weights, fp32 master = ↵Yuren Hao
T64 word-streaming / shadow accumulation); qcomp8/6/4 launched Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-15Stage-0 gate #1: --qup_bits (absolute-grid stochastic-rounding weight ↵Yuren Hao
quantization = finite analog cell levels); qup8/qup6/qup10 arms launched on GPU0 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-15centfast (doubled-batch +/-beta centered, cos 1.000000000 vs sequential, ↵Yuren Hao
DDP-verified) + est_late/est_late_at (tail-only centered); quiet bench: 1.72x -> 1.64x, est_late@80% -> ~1.13x amortized Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-14wandb reorganized: per-regime team projects (ept-tinystories-42m / ↵Yuren Hao
ept-fineweb-72m), auto project routing in trainers, split replay + reports Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-14wandb -> team entity eqprop-llm-training (trainers + full-history replay script)Yuren Hao
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-14bcap v2: absolute-scale gate on rho meter (fixes v1 noise-floor starvation), ↵Yuren Hao
gentler attack 0.85/floor 0.05; crown c2 relaunched from s215000 with cap armed Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-14beta-cap-by-loop-gain controller (--beta_cap_rho): rho^ meter from sweep ↵Yuren Hao
residuals, cap OVERRIDES floor near the wall; smoke passed Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-14Adaptive relax (--relax_tol): sweep-to-tolerance + rho-triggered geta ↵Yuren Hao
backtracking + final full-step graphed round (E-read identity fix: mixing leaked iteration residual into E, gn 1e5); kuse telemetry; paired smoke at wall ckpt: adaptive cos 0.771 vs legacy 0.754 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13Estimator arms machinery: --est centered/richardson (two-pass, O(b^2) bias), ↵Yuren Hao
--muon_mom/--adam_b1 knobs; smoke passed (resume+centered cos 0.998) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13Fix swallowed --wandb_run arg (replace-comment bug); wandb default-on intactYuren Hao
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13wandb ON by default in both trainers (project ept-cascade); BP twin gains ↵Yuren Hao
--amp; fw72m_bp launched GPU1 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13fw72m crown run launched (NCCL DDP prod), exact resume (opt state), RESULT ↵Yuren Hao
17, shuffle+prep scripts Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13Manual DDP for cascade-EP: guard-synced (drift/kretry via global max, gn on ↵Yuren Hao
reduced grads), sigma broadcast, per-rank data RNG, None-preserving flat all-reduce, bitwise param sync-check, grad-equivalence test PASSED (cos 0.999999999 vs single-GPU big batch) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-13Stage-2 data pipeline: FineWeb-Edu sample-10BT -> 32k BPE -> bins; --data ↵Yuren Hao
flag on both trainers Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-12bf16 mixed-precision (--amp): gate PASSED (cos 0.9682 vs fp32 0.9687), ↵Yuren Hao
trainer flag + 3-seed 4k A/B launched Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-12RESULT 13: bsign neutral at 4k (1.7382 vs 1.7314); 42M gap-chasing closed at ↵Yuren Hao
noise floor; pivot to Stage-2 pipeline + E-tier
2026-07-11RESULT 11: stage1b sealed (gap 0.050); beta ceiling unfound (3e-3 -> cos ↵Yuren Hao
0.9974); K exonerated on OLMo2; bf16 naive-cast dead at any beta (TF32 stays cost baseline); f3e3 CE-endpoint cont launched Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-10RESULT 10: epoch endpoints (BP 1.2509/1.2750 vs EP 1.4802, +0.22 erosion ↵Yuren Hao
cost); GENERATION GATE PASSED (casc_gen.py, coherent stories, no-backprop 42.75M); stage1b improved-recipe pair launched (Muon cosine via build_hybrid) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
2026-07-10Sustained dose-response sealed (3e-4/5e-4/1e-3 -> 0.975/0.983/0.990); ship ↵Yuren Hao
--bf_late floor ramp; next-run recipe locked (OLMo2+Muon+late-ramp)
generated by cgit v1.2.3 (git 2.25.1) at 2026-08-18 19:39:55 +0000