diag: ml/eval/diag_cl.py 闭环失败定位 + fix(dagger): collect.py kin.gaps 分块 1024 点、cap=gap_min+1（原 4096 点 cap 40 吃满显存/内存）

ml/README.md

@@ -129,6 +129,7 @@ bash ml/pipeline.sh train                           # MeanFlow 头 → ml/runs/b
 ```
 
 开训前先跑老师轨迹回放自检（`ml/runs/replay_teacher.json`，约 1 分钟）：老师轨迹放进 PyTorch 仿真口径，到达 ≥ 95%、碰撞 ≤ 1% 才开训，否则 rc 5 停——口径对不上时闭环指标不可信。
+闭环指标长期 0% 而开环（ol_fk_mm / ol_best_mm / ol_div）正常时，跑 `PYTHONUTF8=1 python -m ml.eval.diag_cl --ckpt ml/runs/base/last.pt --n 256`（约 2–4 分钟，可与训练并行）→ `ml/runs/diag_cl.json`：A 起点特征 Python 与 JS 逐组差、B 两份特征下动作差、C 首次碰撞时刻 / 组对 / 最小目标误差 / 开安全层到达率。
 
 先跑 2000 步测吞吐（`ml/runs/bench_base/log.jsonl` 的 `samples_per_s`，`eta_h` 是剩余小时数），再正式训练到 60000 步或闭环到达率连续 5 次不升为止。
 

ml/dagger/collect.py

@@ -16,6 +16,8 @@ from ml.data import load_tasks
 from ml.model.policy import build
 from ml.sim.closed_loop import rollout, goal_err, tasks_to_tensors, DT
 
+GAP_CHUNK = 1024      # 每次 kin.gaps 的查询点数上限
+
 
 def main():
     ap = argparse.ArgumentParser()
@@ -41,8 +43,14 @@ def main():
         tr = per['traj'].to(a.device)                                   # [B,T,6]
         B, Tn, _ = tr.shape
         _, T, R, _, _ = tasks_to_tensors(tb, a.device)
+        # 2026-10-19 修（用户实测）：原来一次喂 64 题 × 64 步 = 4096 点、cap=40 mm 给 kin.gaps，40 mm 内幸存的胶囊对太多，
+        # 峰值把显存 / 内存吃满。这里只需判断"≤0（碰撞）"和"> gap_min（起点地板）"，cap 取 gap_min+1 即可，
+        # 且按 GAP_CHUNK 个点分块，用完即释放。
+        cap = gap_min + 1.0
+        flat = tr.reshape(-1, 6)
         with torch.no_grad():
-            g = torch.cat([kin.gaps(tr[:, s:s + 64].reshape(-1, 6), 40.0)['gm'].min(-1).values.view(B, -1) for s in range(0, Tn, 64)], 1)
+            g = torch.cat([kin.gaps(flat[s:s + GAP_CHUNK], cap)['gm'].min(-1).values for s in range(0, flat.shape[0], GAP_CHUNK)]).view(B, Tn)
+        del flat
         for b, t in enumerate(tb):
             hit = (g[b] <= 0).nonzero()
             c = int(hit[0]) if hit.numel() else Tn
@@ -58,6 +66,8 @@ def main():
                 if e.item() < 5 and (not t.get('R') or oe.item() < 5): stat['rej_near'] += 1; continue
                 out.append({'q0': [round(v, 3) for v in q[0].tolist()], 'T': t['T'], 'R': t.get('R'), 'cat': t.get('cat', 'A'),
                             'src': {'tid': t['tid'], 'k': k, 'col': bool(hit.numel())}})
+        del tr, g, per
+        if a.device.startswith('cuda'): torch.cuda.empty_cache()
         print(f'[dagger] {min(i + a.bs, len(tasks))}/{len(tasks)} 题，起点 {len(out)}', flush=True)
     stat['states'] = len(out)
     rng.shuffle(out)

ml/eval/diag_cl.py

@@ -0,0 +1,122 @@
+#!/usr/bin/env python
+"""闭环失败定位（只读，手动运行，约 2–4 分钟；训练进行中也可以另开终端跑，显存约 2–3 GB）。
+
+2026-10-19 新增：第二次预训练（979a581）开环已经正常（候选0 末端误差 ~6 mm、K 中最好 ~2.9 mm、多样性 0.65），
+闭环却仍 100% 碰撞、0% 到达。开环用的是 JS 生成的特征 X，闭环用的是 Python features() 现算的特征，
+两者之间只要有一处口径不同，模型在闭环里看到的就是没见过的输入。本脚本回答三件事：
+  A. 特征口径：val 任务起点（非增广、老师状态）上，Python features() 与数据集 X 的逐组最大差；
+  B. 同一噪声下，用两份特征得到的动作差（16 步末端关节差，度）——A 有差但 B 很小说明差异无害；
+  C. 闭环失败形态：首次碰撞时刻分布、碰撞步数、首次碰撞的组对、全程最小目标误差；
+     以及同一批题开安全层（safety=True，近似部署）的到达率 / 刹停次数。
+用法（仓库根目录）：
+  PYTHONUTF8=1 python -m ml.eval.diag_cl --ckpt ml/runs/base/last.pt --n 256 --out ml/runs/diag_cl.json
+"""
+import argparse, json, os
+
+import torch
+
+from ml.common.kin_torch import Kin
+from ml.common.features import features, STEP, H
+from ml.data import Samples, load_tasks
+from ml.model.policy import build
+from ml.sim.closed_loop import rollout, goal_err, DT
+
+# 83 维特征分组（与 features.py 拼接顺序一致）
+GROUPS = [('sincos', 0, 12), ('joint_norm', 12, 15), ('dq', 15, 21), ('p', 21, 24), ('Rq', 24, 30), ('T', 30, 33),
+          ('R', 33, 39), ('T-p', 39, 42), ('T-p_clamp', 42, 45), ('mode', 45, 47), ('prefix', 47, 65), ('mask', 65, 68),
+          ('sub', 68, 74), ('sub_flag', 74, 75), ('gaps', 75, 83)]
+
+
+def pct(v, p):
+    v = sorted(v)
+    return v[min(len(v) - 1, int(len(v) * p))] if v else None
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument('--ckpt', required=True); ap.add_argument('--samples', default='ml/data/samples')
+    ap.add_argument('--n', type=int, default=256); ap.add_argument('--bs', type=int, default=64)
+    ap.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu'); ap.add_argument('--out', default='ml/runs/diag_cl.json')
+    a = ap.parse_args()
+    dev = a.device
+    ck = torch.load(a.ckpt, map_location=dev)
+    model = build(ck['cfg']).to(dev); model.load_state_dict(ck['model']); model.eval()
+    kin = Kin(device=dev)
+    out = {'ckpt': a.ckpt, 'step': ck.get('step')}
+    tasks = [t for t in load_tasks([a.samples]) if t['split'] == 'val'][:a.n]
+    by_tid = {t['tid']: t for t in tasks}
+
+    # ---------- A / B：起点特征口径 ----------
+    vd = Samples(a.samples, device=dev, split='val')
+    tid = vd.TID.long()
+    first = torch.ones_like(tid, dtype=torch.bool); first[1:] = tid[1:] != tid[:-1]     # 每题第一行 = t=0 样本
+    idx = [i for i in first.nonzero().flatten().tolist() if int(tid[i]) in by_tid]
+    X0 = vd.X[idx].float()
+    tk = [by_tid[int(tid[i])] for i in idx]
+    q0 = torch.tensor([t['q0'] for t in tk], dtype=torch.float32, device=dev)
+    qx = torch.atan2(X0[:, 0:12:2], X0[:, 1:12:2]) * 180 / torch.pi
+    dqx = torch.remainder(qx - q0 + 180, 360) - 180
+    clean = dqx.abs().max(-1).values < 1e-2                  # 未被噪声增广的起点
+    out['A_rows'] = int(clean.sum()); out['A_rows_aug_skipped'] = int((~clean).sum())
+    if clean.any():
+        Xc = X0[clean]; qc = q0[clean]; tc = [t for t, c in zip(tk, clean.tolist()) if c]
+        T = torch.tensor([t['T'] for t in tc], dtype=torch.float32, device=dev)
+        R = torch.tensor([t['R'] if t.get('R') else [0.] * 9 for t in tc], dtype=torch.float32, device=dev)
+        B = qc.shape[0]
+        pre = Xc[:, 47:65].view(B, 3, 6) * torch.tensor(STEP, device=dev)       # 前缀 / 子目标直接取数据集里的，只比状态相关部分
+        pl = Xc[:, 65:68].sum(-1).round()
+        sf = Xc[:, 74]
+        sub = qc + Xc[:, 68:74] * 180
+        Fp = features(kin, qc, qc, T, R, pre, pl, sub, sf)
+        out['A_maxdiff'] = {g: (Fp[:, s:e] - Xc[:, s:e]).abs().max().item() for g, s, e in GROUPS}
+        out['A_meandiff'] = {g: (Fp[:, s:e] - Xc[:, s:e]).abs().mean().item() for g, s, e in GROUPS}
+        with torch.no_grad():
+            z = torch.randn(B * 8, H, 6, device=dev, generator=torch.Generator(device=dev).manual_seed(7))
+            a1, _, _ = model.sample_k(Xc, 8, z); a2, _, _ = model.sample_k(Fp, 8, z)
+        stepv = torch.tensor(STEP, device=dev)
+        d = ((a1.clamp(-1, 1) - a2.clamp(-1, 1)) * stepv).sum(2).abs().max(-1).values     # [B,K] 16 步末端关节差（度）
+        mag = (a1.clamp(-1, 1) * stepv).sum(2).abs().max(-1).values
+        out['B_act_diff_deg'] = {'mean': d.mean().item(), 'p90': d.flatten().kthvalue(max(1, int(0.9 * d.numel()))).values.item(),
+                                 'ref_motion_deg_mean': mag.mean().item()}
+
+    # ---------- C：闭环失败形态 ----------
+    first_t, col_steps, pairs, min_err, reach = [], [], {}, [], 0
+    reach_s, brakes_s, col_s = 0, [], 0
+    for i in range(0, len(tasks), a.bs):
+        ch = tasks[i:i + a.bs]
+        with torch.no_grad():
+            _, per = rollout(model, kin, ch, dev, K=8, safety=False, seed=i, return_traj=True)
+            tr = per['traj'].to(dev)                                            # [B,Tn,6]
+            B, Tn, _ = tr.shape
+            flat = tr.reshape(-1, 6)                                            # 分块 ≤1024 点（同 dagger/collect.py 的显存修复）
+            gm = torch.cat([kin.gaps(flat[s:s + 1024], 0.5)['gm'] for s in range(0, flat.shape[0], 1024)]).view(B, Tn, 49)
+            hit = gm.min(-1).values <= 0                                        # [B,Tn]
+            Tt = torch.tensor([t['T'] for t in ch], dtype=torch.float32, device=dev)
+            Rt = torch.tensor([t['R'] if t.get('R') else [0.] * 9 for t in ch], dtype=torch.float32, device=dev)
+            e, _ = goal_err(kin, tr.reshape(-1, 6), Tt.repeat_interleave(Tn, 0), Rt.repeat_interleave(Tn, 0))
+            e = e.view(B, Tn)
+            # 到达后轨迹被冻结（d=0），之后不再计
+            for b in range(B):
+                if per['done'][b]: reach += 1
+                min_err.append(e[b].min().item()); col_steps.append(int(per['col'][b]))
+                h = hit[b].nonzero().flatten()
+                if len(h):
+                    k = int(h[0]); first_t.append(k * DT)
+                    pr = int(gm[b, k].argmin()); key = '%d-%d' % (pr // 7, pr % 7); pairs[key] = pairs.get(key, 0) + 1
+            rs, ps = rollout(model, kin, ch, dev, K=8, safety=True, seed=i)
+            reach_s += int(ps['done'].sum()); col_s += int((ps['col'] > 0).sum()); brakes_s += ps['brakes'].tolist()
+    n = len(tasks)
+    out['C'] = {'n': n, 'reach': reach / n, 'collided': len(first_t) / n,
+                'first_hit_s': {'p10': pct(first_t, .1), 'p50': pct(first_t, .5), 'p90': pct(first_t, .9),
+                                'within_0.1s': sum(t <= 0.1 for t in first_t) / max(1, len(first_t))},
+                'col_steps': {'p50': pct(col_steps, .5), 'p90': pct(col_steps, .9)},
+                'first_hit_pair(group a-b)': dict(sorted(pairs.items(), key=lambda kv: -kv[1])[:8]),
+                'min_goal_err_mm': {'p10': pct(min_err, .1), 'p50': pct(min_err, .5), 'p90': pct(min_err, .9)},
+                'safety_on': {'reach': reach_s / n, 'collided': col_s / n, 'brakes_p50': pct(brakes_s, .5)}}
+    s = json.dumps(out, indent=1, ensure_ascii=False); print(s)
+    os.makedirs(os.path.dirname(a.out) or '.', exist_ok=True)
+    open(a.out, 'w', encoding='utf-8').write(s)
+
+
+if __name__ == '__main__':
+    main()