第八轮:monitor 状态三项被注释吞掉的 bug;R9 改判稳态 planWarmP99 + 运动中饥饿 + 认证超时占比;APG steps 8000(batch 48 保持总任务量);r4 不再静默退回 CDN;自检副本含未跟踪文件;显存文案按 19–20 GB;审查文档精简为实测/判据/待数据/三段训练

This commit is contained in:
Hark
2026-10-18 06:34:22 +08:00
committed by cbb
parent 466a71ecea
commit 34f19aac61
14 changed files with 81 additions and 298 deletions
+2 -1
View File
@@ -1,8 +1,9 @@
{
"apg": {
"batch": 48, "lr": 1e-4, "steps": 3000, "grad_clip": 1.0, "amp": "bf16",
"batch": 48, "lr": 1e-4, "steps": 8000, "grad_clip": 1.0, "amp": "bf16",
"plans": 60, "plans_max": 150, "seg_plans": 10,
"_vram": "batch 由 128 降到 48:2026-10-18 本机实测(d=192 同结构、每进程 19.2 GB 硬帽)batch=128 无论 plans 取 60 还是 150 都 OOM(崩前已分配 17.85 GiB);batch=48/plans=150 峰值 13.08 GiB、plans=60 为 7.67 GiB,加 Windows+4K 桌面常驻 4.4 GB 仍在 19–20 GB 预算内。batch=64/plans=150 要 17.66 GiB(不取)。展开长度 plans 没动——降 plans 会把长任务的绕出段截掉。",
"_steps": "2026-10-18 第八轮:batch 128→48 后 steps 3000→8000,保持总任务量(3000×128 ≈ 8000×48)不变;每步展开代价按 batch 线性降,总机时与原估同量级(约 4 h),学习率不动。",
"margin_train_mm": 3.0,
"w_tgt": 1.0, "w_rot": 0.3, "w_col": 2.0, "w_smooth": 0.02, "w_len": 0.001, "w_bc": 0.5,
"cat_weights": {"A": 0.4, "B": 1.0, "C": 1.6, "D": 0.6, "Dr": 0.6},