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/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)