summaryrefslogtreecommitdiff
path: root/ep_run/casc_eq_train.py
AgeCommit message (Expand)Author
13 days偏置放大公理(用户08-06指出): 逐坐标归一化把持久小偏置...Yuren Hao
13 days第二代原语优化器: SignLine/ConsensusLion/DitherLion 实装+本地筛...Yuren Hao
13 daysEqOLion R4(真实轨迹判决): 冻结极因子几十步退相关至~0.53=慢...Yuren Hao
13 days优化器价目表实装: Lion/OLion/Adafactor2D/build_alt + --opt 六选项Yuren Hao
2026-08-04RESULT 89终局 + 交付物: 135M完赛 EP 3.2087 vs BP 3.2074(+0.1% ppl), ...Yuren Hao
2026-07-30RESULT 78: read_lin梯度级验证==fp64(fp32下泄漏塌零,读出舍入=全...Yuren Hao
2026-07-30RESULT 77: 地板理论三箭全中(fp64塌零/β平移/β×8痊愈) — 物...Yuren Hao
2026-07-30RESULT 76: 谱仪读出 — 逐块泄漏剖面(C768上6块-0.70..-0.23,C512...Yuren Hao
2026-07-30dgain几何profile + M1泄漏谱仪(用户设计,07-30)Yuren Hao
2026-07-29RESULT 72: 剂量曲线通顶 — dg128关闭97%(距all-BP杆0.0015),物理...Yuren Hao
2026-07-29RESULT 69: 饱和盲区被自己的治疗测试证伪(cap30/50关闭0.0000,...Yuren Hao
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
2026-07-24RESULT 60b: res表在amp下宽度盲(bf16地板把fp32四个量级压成1.2×...Yuren Hao
2026-07-24RESULT 60: s345000走廊空(end075更糟3.4805,LR嫌疑排除)→闭合点275...Yuren Hao
2026-07-22遥测扩展(用户规格): wandb新键 gn/clip_norm/clip_fired + watch_every...Yuren Hao
2026-07-22resume NameError修复: bsimp恢复写在GOV定义之前(仅新ckpt含bsimp...Yuren Hao
2026-07-20棘轮模式(用户设计07-20): beta_simple 1.0=只降不升(越界折半持...Yuren Hao
2026-07-20RESULT 54: 上界∝1/σ²下沉(σ242→473翻倍,β*从0.12跌到<0.03),固...Yuren Hao
2026-07-20guards_silent(用户令07-20): 守卫只报不拦(skip全部放行提交,唯...Yuren Hao
2026-07-19门最终修正(SV2验证臂拦截): res>0.02单独判非法即封35k逃逸...Yuren Hao
2026-07-19CODEX审计修复×4: _legal门AND→OR+NaN非法+ddp_max(600步连续放行...Yuren Hao
2026-07-19beta_simple(用户规格,复杂机构全拆): β=起始×bsimp,干净步×1.0...Yuren Hao
2026-07-19beta_servo: 贴上界伺服(亮区反演cap*=servo·cap_rho/ρ̂双向一步...Yuren Hao
2026-07-18governor接线修复:res/rho非法判据常开(beta_cap_rho>0即生效),crown...Yuren Hao
2026-07-18wsync: 同步权重步接受(用户风格重设计,knee撤回) — opt.step...Yuren Hao
2026-07-18RESULT 49: 回路解剖=23段乘积近临界(下行×72/上行×13k/顶读1e-...Yuren Hao
2026-07-18RESULT 43+44: crown-3 死因=cap硬底钉在塌落天花板之上(变相wall-...Yuren Hao
2026-07-18--bpmix: gap-source ablation (overwrite EP grads with true BP grads per modul...Yuren Hao
2026-07-17--beta_sync: synchronous acceptance (this-step relax telemetry gates the comm...Yuren Hao
2026-07-17ride-v2 (b)/(c) flag-gated (--ride_ema/--ride_cool, default off); isolation p...Yuren Hao
2026-07-17ride-v2: (a) floor-jump cap continuity (the 0.09 bug), (b) rho-EMA decisions,...Yuren Hao
2026-07-17save final ckpt at last step (fixed-eval protocol); fw72m_ride crown-3 launch...Yuren Hao
2026-07-16--beta_ride ceiling-riding governor (two-sided bcap: climb in calm, attack ne...Yuren Hao
2026-07-16RESULT 29: qcomp8 tax-free (+0.0018) = T64 8-bit compute green light; --centm...Yuren Hao
2026-07-16Stage-0 gate #1b: --qcomp_bits (compute on DAC-grid weights, fp32 master = T6...Yuren Hao
2026-07-15Stage-0 gate #1: --qup_bits (absolute-grid stochastic-rounding weight quantiz...Yuren Hao
2026-07-15centfast (doubled-batch +/-beta centered, cos 1.000000000 vs sequential, DDP-...Yuren Hao
2026-07-14wandb reorganized: per-regime team projects (ept-tinystories-42m / ept-finewe...Yuren Hao
2026-07-14wandb -> team entity eqprop-llm-training (trainers + full-history replay script)Yuren Hao
2026-07-14bcap v2: absolute-scale gate on rho meter (fixes v1 noise-floor starvation), ...Yuren Hao
2026-07-14beta-cap-by-loop-gain controller (--beta_cap_rho): rho^ meter from sweep resi...Yuren Hao
2026-07-14Adaptive relax (--relax_tol): sweep-to-tolerance + rho-triggered geta backtra...Yuren Hao
2026-07-13Estimator arms machinery: --est centered/richardson (two-pass, O(b^2) bias), ...Yuren Hao
2026-07-13Fix swallowed --wandb_run arg (replace-comment bug); wandb default-on intactYuren Hao
2026-07-13wandb ON by default in both trainers (project ept-cascade); BP twin gains --a...Yuren Hao
2026-07-13fw72m crown run launched (NCCL DDP prod), exact resume (opt state), RESULT 17...Yuren Hao
2026-07-13Manual DDP for cascade-EP: guard-synced (drift/kretry via global max, gn on r...Yuren Hao
2026-07-13Stage-2 data pipeline: FineWeb-Edu sample-10BT -> 32k BPE -> bins; --data fla...Yuren Hao
2026-07-12bf16 mixed-precision (--amp): gate PASSED (cos 0.9682 vs fp32 0.9687), traine...Yuren Hao
2026-07-12RESULT 13: bsign neutral at 4k (1.7382 vs 1.7314); 42M gap-chasing closed at ...Yuren Hao