优化 jepa.py 中通用 rollout 热路径：批量预编码动

作、移除循环内
  torch.cat，并为 history_size==1 与环形缓冲区更新
  添加更轻量实现； 收益不大

jepa.py

@@ -133,35 +133,89 @@ class JEPA(nn.Module):
         """
         with torch.profiler.record_function("lewm.rollout"):
             assert "pixels" in info, "pixels not in info_dict"
+            if history_size < 1:
+                raise ValueError("history_size must be >= 1")
+
             H = info["pixels"].size(2)
             B, S, T = action_sequence.shape[:3]
-            act_0, act_future = torch.split(action_sequence, [H, T - H], dim=2)
+            if T < H:
+                raise ValueError(
+                    f"action_sequence horizon ({T}) must be >= history length ({H})"
+                )
 
             # Cache the encoded initial state across solver iterations.
             init_emb = self._get_cached_init_emb(info)
             HS = history_size
-            emb_hist = init_emb.unsqueeze(1).expand(B, S, -1, -1)
-            emb_hist = emb_hist[..., -HS:, :].reshape(B * S, min(HS, init_emb.size(1)), -1)
+            hist_len = min(HS, init_emb.size(1), H)
+            if hist_len < 1:
+                raise ValueError("rollout requires at least one history step")
 
-            act_hist = act_0[..., -HS:, :].reshape(B * S, min(HS, act_0.size(2)), -1)
-            act_emb_hist = self.action_encoder(act_hist)
-            act_future = act_future.reshape(B * S, act_future.size(2), -1)
+            init_hist = init_emb[:, -hist_len:]
+            init_hist = init_hist.unsqueeze(1).expand(-1, S, -1, -1)
+            init_hist = init_hist.reshape(B * S, hist_len, init_hist.size(-1))
 
-            for t in range(act_future.size(1)):
-                pred_emb = self.predict(emb_hist[:, -HS:], act_emb_hist[:, -HS:])[:, -1:]
-                if HS > 1:
-                    emb_hist = torch.cat([emb_hist[:, -HS + 1 :], pred_emb], dim=1)
+            flat_actions = action_sequence.reshape(B * S, T, -1)
+            action_emb = self.action_encoder(flat_actions)
+            act_hist = action_emb[:, H - hist_len : H]
+            act_future = action_emb[:, H:]
+
+            if HS == 1:
+                emb_hist = init_hist[:, -1:]
+                act_emb_hist = act_hist[:, -1:]
+
+                for t in range(act_future.size(1)):
+                    emb_hist = self.predict(emb_hist, act_emb_hist)[:, -1:]
+                    act_emb_hist = act_future[:, t : t + 1]
+
+                pred_rollout = self.predict(emb_hist, act_emb_hist)[:, -1:]
+            else:
+                emb_hist = init_hist.new_empty((B * S, HS, init_hist.size(-1)))
+                act_emb_hist = action_emb.new_empty((B * S, HS, action_emb.size(-1)))
+                emb_hist[:, :hist_len].copy_(init_hist)
+                act_emb_hist[:, :hist_len].copy_(act_hist)
+
+                history_order = torch.stack(
+                    [
+                        (torch.arange(HS, device=action_emb.device) + offset) % HS
+                        for offset in range(HS)
+                    ]
+                )
+                filled = hist_len
+                next_slot = hist_len % HS
+
+                for t in range(act_future.size(1)):
+                    if filled < HS:
+                        emb_view = emb_hist[:, :filled]
+                        act_view = act_emb_hist[:, :filled]
+                    elif next_slot == 0:
+                        emb_view = emb_hist
+                        act_view = act_emb_hist
+                    else:
+                        order = history_order[next_slot]
+                        emb_view = emb_hist.index_select(1, order)
+                        act_view = act_emb_hist.index_select(1, order)
+
+                    pred_emb = self.predict(emb_view, act_view)[:, -1:]
+                    next_act_emb = act_future[:, t : t + 1]
+                    emb_hist[:, next_slot : next_slot + 1].copy_(pred_emb)
+                    act_emb_hist[:, next_slot : next_slot + 1].copy_(next_act_emb)
+
+                    if filled < HS:
+                        filled += 1
+                    next_slot = (next_slot + 1) % HS
+
+                if filled < HS:
+                    emb_view = emb_hist[:, :filled]
+                    act_view = act_emb_hist[:, :filled]
+                elif next_slot == 0:
+                    emb_view = emb_hist
+                    act_view = act_emb_hist
                 else:
-                    emb_hist = pred_emb
+                    order = history_order[next_slot]
+                    emb_view = emb_hist.index_select(1, order)
+                    act_view = act_emb_hist.index_select(1, order)
 
-                next_act = act_future[:, t : t + 1, :]
-                next_act_emb = self.action_encoder(next_act)
-                if HS > 1:
-                    act_emb_hist = torch.cat([act_emb_hist[:, -HS + 1 :], next_act_emb], dim=1)
-                else:
-                    act_emb_hist = next_act_emb
-
-            pred_rollout = self.predict(emb_hist[:, -HS:], act_emb_hist[:, -HS:])[:, -1:]
+                pred_rollout = self.predict(emb_view, act_view)[:, -1:]
             info["predicted_emb"] = pred_rollout.reshape(B, S, *pred_rollout.shape[1:])
 
             return info

sth.md

@@ -1,192 +1,52 @@
-1. 压 rollout 内环
-  这条最通用，而且基本不改算法语义，只是把实现做对。
+我建议优先做这 4 类，都是跨数据集成立的：
 
-  在 jepa.py:129 这段里，当前问题是：
-
-  - 循环里每步都 action_encoder(next_act)，见 jepa.py:159
-  - history 每步用 torch.cat 重建，见 jepa.py:155 和 jepa.py:162
-  - 每步都走一次很短的 predict()，host 调度比例很高
-
-  通用改法：
+  1. 压 rollout 内环实现
+     见 jepa.py:127。现在每步都在做 action_encoder、切片、torch.cat、小规
+     模 predict 调用，这种碎片化实现对任何任务都亏。
+     通用改法：
 
   - 整条 action_sequence 一次性做 action_encoder
   - emb_hist / act_emb_hist 改成预分配 buffer
-  - 用 ring buffer 或 index rotate 更新历史
-  - 循环里只做 copy_ / 索引覆盖，不做 cat
+  - 循环里只做索引覆盖或 copy_
+  - 去掉循环内 torch.cat
 
-  这个优化对任何数据集都成立，因为它优化的是“rolling inference 实现方
-  式”，不是任务参数。
+  2. 减少热路径里的搬运和同步
+     profile 里 aten::copy_ 很重，这不是 TwoRoom 特有问题。重点看
+     jepa.py:67 和 jepa.py:186。
+     通用目标：
 
-  2. 用 torch.inference_mode()
-  你现在在 eval.py:242 这里只用了 autocast，没有 inference_mode()。
+  - 模型侧张量尽量全程留在 GPU
+  - 避免热路径反复 .to(device) / 隐式 layout 修复
+  - 到必须和环境交互的边界再一次性转 CPU / numpy
+  - 确保进入 predictor 的张量是 contiguous 的，少触发隐式 copy
 
-  建议推理主路径外层直接包：
+  3. 把编译成本移出正式计时
+     现在 torch.compile 默认开在 predictor，见 eval.py:70。102s -> 45s 很
+     像首轮编译预热。
+     通用做法：
 
-  with torch.inference_mode():
-      with inference_ctx:
-          ...
+  - 在正式 start_time 前做一次 dummy predict 或 dummy rollout
+  - 保留只编译 predictor/predict，不要编译整个 solver
 
-  这是纯通用优化，所有数据集都受益。
+  4. 减少临时对象和 shape bookkeeping
+     这是所有任务都会受益的。
+     重点看：
 
-  3. 只编译 predictor / predict，不要编译整个 solver
-  当前热点是大量小 predict() 调用，不是整条 eval graph。
+  - jepa.py:100 到 jepa.py:106
+  - jepa.py:143 到 jepa.py:148
+    方向是：
+  - 能循环外做的 reshape，不放循环里
+  - 能原地更新，不新建张量
+  - 少做 dict 字段增删和中间容器组装
 
-  通用建议：
+  不建议优先做的通用性较差方案：
 
-  - 先只编译 self.predictor
-  - 或只编译 JEPA.predict()
-  - 模式优先试 reduce-overhead
+  - 调 TwoRoom 专属 cache 规则
+  - 改数据集采样逻辑
+  - 按小数据集特点缩短 horizon
+  - 直接改 CEM 超参当“优化”
 
-  不要先编译整个 WorldModelPolicy 或 CEM solver；那通常图不稳定，泛化收益
-  反而差。
+  如果你要我直接开始改，我建议第一批只做两件事：
 
-  4. 减少循环里的张量形状重排和临时对象
-  这也是实现层通用优化。
-
-  可以继续查：
-
-  - rearrange 是否能前移到循环外
-  - 是否有重复的 slice/view 触发隐式拷贝
-  - pred_proj(rearrange(...)) 这类 reshape 往返是否能合并
-
-  这类优化对所有任务都有效，因为是在降 Python 和 tensor bookkeeping 成
-  本。
-
-  5. 再考虑结构级优化，但放后面
-  比如 predictor 深度、MLP 宽度、heads 数量。这也通用，但已经开始碰模型容
-  量和精度，不该是第一刀。
-
-  不建议优先做的
-  这些更偏任务/数据集相关，不算你要的“泛用优化”：
-
-  - 先调 num_samples/topk/n_steps
-  - 先缩 horizon
-  - 先按 tworoom 特性做 shortcut
-  - 先针对某个 dataset 做 cache 规则
-
-  一句话判断
-  你现在最像是“算法没错，但 rollout 实现过于碎片化”，所以第一优先级应该
-  是：
-  一次性 action encode + 预分配历史 buffer + 去掉循环内 torch.cat +
-  inference_mode + compile predictor
-
-  如果你要，我下一步就直接改 jepa.py 做这套通用优化，不碰任何数据集特化逻
-  辑。
-
-  除开 CEM solve 本体，剩下这些杂项可以这样优化。
-
-  最高优先级
-
-  1. 保证传给环境的是 numpy，不要让 Gym 代转
-     你日志已经说明 env step() 收到了 torch.Tensor。这会带来拷贝、同步、
-     checker 额外开销。
-     做法：
-
-  - 在 policy 输出动作、准备喂给 env 的那一层，显式转成
-    action.detach().cpu().numpy()
-  - 最好一次性转好，别在 env 内部或 wrapper 内隐式转换
-
-  收益：
-
-  - 去掉 Gym warning
-  - 减少同步和类型检查开销
-  - 通常是最直接的非模型提速点
-
-  2. 关掉 Gym passive checker
-     这些 warning 本身就说明 checker 在持续检查类型和空间匹配。
-     做法：
-
-  - 尽量用禁 checker 的构造方式
-  - 或在你自己的 env wrapper 里保证输入输出符合 Gym 预期，避免它每步检查
-
-  收益：
-
-  - 每步少一层 Python 校验
-  - 对长 episode / 多 episode 累积明显
-
-  中优先级
-
-  3. 把预处理前移，避免每步重复做能缓存的东西
-     如果 goal、初始条件、某些 dataset 字段在 episode 里不变，就不要每次
-     都重新组织。
-     做法：
-
-  - goal 相关 embedding 已经有缓存，继续扩展到更多静态字段
-  - 固定的 callables 参数尽量预解析
-  - 能在 episode 开头准备好的，不要放在 step 循环里
-
-  收益：
-
-  - 降低 Python dict 操作和小张量处理开销
-
-  4. 避免频繁 CPU/GPU 来回切
-     如果模型在 GPU，但环境在 CPU，就要非常小心中间格式。
-     做法：
-
-  - 模型侧尽量连续留在 GPU
-  - 到真正 env step 前再一次性转 CPU numpy
-  - 不要中间反复 .cpu() / .to(device) / np.array(...)
-
-  收益：
-
-  - 减少隐式同步
-  - 稳定延迟
-
-  5. 缩减 Python 层对象操作
-     dict 组装、字段拷贝、wrapper 嵌套太多时，端到端会慢。
-     做法：
-
-  - 关键热路径里少做深拷贝
-  - 少重复构造新的 info / obs 容器
-  - 固定结构优先原地更新
-
-  收益：
-
-  - 对小步高频调用路径有效
-
-  如果你要继续压评测时间
-
-  6. 降低日志和 warning 输出
-     频繁 warning 会拖慢，而且污染 timing。
-     做法：
-
-  - 修掉类型不匹配后 warning 自然消失
-  - 非必要的 print 尤其是 step 内 print 要去掉
-
-  7. 针对环境 step 做批量化检查
-     如果 num_envs=50，尽量确认 env wrapper 没有在内部退化成逐环境 Python
-     for-loop。
-     做法：
-
-  - 查 world.evaluate_from_dataset() 到 env step() 之间是不是 batch 接口
-  - 如果 batch env 里还有逐个 env 转换/检查，尽量前移或向量化
-
-  收益：
-
-  - 这类经常能解释“为什么 solver 时间之外还有很多时间”
-
-  8. 把 callables 的执行成本单独看
-     你这里有：
-
-  - _set_state
-  - _set_goal_state
-
-
-  - 确认它们只在必要时执行
-  - 能批量设置就别逐条 Python 调
-  2. 消掉 Gym warning
-  3. 单独量 env.step 总时间
-  4. 检查是否有反复 CPU/GPU 转换
-  5. 再看 wrapper / callable / obs 组装
-
-  一句话总结
-  剩下的杂项优化，核心不是“再多上几张卡”，而是：
-
-  - 去掉隐式类型转换
-  - 去掉多余检查
-  - 去掉重复数据整理
-  - 减少 CPU/GPU 往返
-  - 减少 Python 高频小开销
-
-  如果你要，我下一步可以直接帮你定位“动作是在哪一层以 torch.Tensor 传进
-  env 的”，给你指出具体应该改哪个函数。
+  - 重写 jepa.py:127 这段 rollout，去掉循环内 action_encoder + cat
+  - 在 eval.py:306 前加 compile warmup

tworoom_results.txt

@@ -1282,3 +1282,489 @@ evaluation_time: 44.974061727523804 seconds
 inference_precision: fp16
 inference_compile_target: predictor
 inference_compile_mode: reduce-overhead
+
+==== CONFIG ====
+cache_dir: null
+solver:
+  _target_: stable_worldmodel.solver.CEMSolver
+  model: ???
+  batch_size: 1
+  num_samples: 300
+  var_scale: 1.0
+  n_steps: 30
+  topk: 30
+  device: cuda
+  seed: ${seed}
+world:
+  env_name: swm/TwoRoom-v1
+  num_envs: ${eval.num_eval}
+  max_episode_steps: 100
+  history_size: 1
+  frame_skip: 1
+seed: 42
+policy: two-room/tworoom/lejepa
+inference_precision: fp16
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+  - action
+  - proprio
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: tworoom
+  callables:
+  - method: _set_state
+    args:
+      state:
+        value: proprio
+  - method: _set_goal_state
+    args:
+      goal_state:
+        value: goal_proprio
+output:
+  filename: tworoom_results.txt
+
+==== RESULTS ====
+metrics: {'success_rate': 88.0, 'episode_successes': array([ True, False,  True, False,  True,  True,  True,  True, False,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True, False,  True,  True,  True,  True,
+        True,  True, False, False,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True]), 'seeds': None}
+evaluation_time: 102.31317353248596 seconds
+inference_precision: fp16
+inference_compile_target: predictor
+inference_compile_mode: reduce-overhead
+
+==== CONFIG ====
+cache_dir: null
+solver:
+  _target_: stable_worldmodel.solver.CEMSolver
+  model: ???
+  batch_size: 1
+  num_samples: 300
+  var_scale: 1.0
+  n_steps: 30
+  topk: 30
+  device: cuda
+  seed: ${seed}
+world:
+  env_name: swm/TwoRoom-v1
+  num_envs: ${eval.num_eval}
+  max_episode_steps: 100
+  history_size: 1
+  frame_skip: 1
+seed: 42
+policy: two-room/tworoom/lejepa
+inference_precision: fp16
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+  - action
+  - proprio
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: tworoom
+  callables:
+  - method: _set_state
+    args:
+      state:
+        value: proprio
+  - method: _set_goal_state
+    args:
+      goal_state:
+        value: goal_proprio
+output:
+  filename: tworoom_results.txt
+
+==== RESULTS ====
+metrics: {'success_rate': 88.0, 'episode_successes': array([ True, False,  True, False,  True,  True,  True,  True, False,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True, False,  True,  True,  True,  True,
+        True,  True, False, False,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True]), 'seeds': None}
+evaluation_time: 45.355348110198975 seconds
+inference_precision: fp16
+inference_compile_target: predictor
+inference_compile_mode: reduce-overhead
+
+==== CONFIG ====
+cache_dir: null
+solver:
+  _target_: stable_worldmodel.solver.CEMSolver
+  model: ???
+  batch_size: 1
+  num_samples: 300
+  var_scale: 1.0
+  n_steps: 30
+  topk: 30
+  device: cuda
+  seed: ${seed}
+world:
+  env_name: swm/TwoRoom-v1
+  num_envs: ${eval.num_eval}
+  max_episode_steps: 100
+  history_size: 1
+  frame_skip: 1
+seed: 42
+policy: two-room/tworoom/lejepa
+inference_precision: fp16
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+  - action
+  - proprio
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: tworoom
+  callables:
+  - method: _set_state
+    args:
+      state:
+        value: proprio
+  - method: _set_goal_state
+    args:
+      goal_state:
+        value: goal_proprio
+output:
+  filename: tworoom_results.txt
+profile:
+  enabled: true
+  export_tensorboard: false
+  export_chrome_trace: false
+
+==== RESULTS ====
+metrics: {'success_rate': 88.0, 'episode_successes': array([ True, False,  True, False,  True,  True,  True,  True, False,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True, False,  True,  True,  True,  True,
+        True,  True, False, False,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True]), 'seeds': None}
+evaluation_time: 110.91939687728882 seconds
+inference_precision: fp16
+inference_compile_target: predictor
+inference_compile_mode: reduce-overhead
+profile_dir: /mnt/ASC1637/lewm_baseline/le-wm/torch_profile
+profile_summary: /mnt/ASC1637/lewm_baseline/le-wm/torch_profile/key_averages.txt
+
+==== CONFIG ====
+cache_dir: null
+solver:
+  _target_: stable_worldmodel.solver.CEMSolver
+  model: ???
+  batch_size: 1
+  num_samples: 300
+  var_scale: 1.0
+  n_steps: 30
+  topk: 30
+  device: cuda
+  seed: ${seed}
+world:
+  env_name: swm/TwoRoom-v1
+  num_envs: ${eval.num_eval}
+  max_episode_steps: 100
+  history_size: 1
+  frame_skip: 1
+seed: 42
+policy: two-room/tworoom/lejepa
+inference_precision: fp16
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+  - action
+  - proprio
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: tworoom
+  callables:
+  - method: _set_state
+    args:
+      state:
+        value: proprio
+  - method: _set_goal_state
+    args:
+      goal_state:
+        value: goal_proprio
+output:
+  filename: tworoom_results.txt
+
+==== RESULTS ====
+metrics: {'success_rate': 88.0, 'episode_successes': array([ True, False,  True, False,  True,  True,  True,  True, False,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True, False,  True,  True,  True,  True,  True,
+        True,  True,  True,  True, False,  True,  True,  True,  True,
+        True,  True, False,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True]), 'seeds': None}
+evaluation_time: 54.21496343612671 seconds
+inference_precision: fp16
+inference_compile_target: predictor
+inference_compile_mode: reduce-overhead
+
+==== CONFIG ====
+cache_dir: null
+solver:
+  _target_: stable_worldmodel.solver.CEMSolver
+  model: ???
+  batch_size: 1
+  num_samples: 300
+  var_scale: 1.0
+  n_steps: 30
+  topk: 30
+  device: cuda
+  seed: ${seed}
+world:
+  env_name: swm/TwoRoom-v1
+  num_envs: ${eval.num_eval}
+  max_episode_steps: 100
+  history_size: 1
+  frame_skip: 1
+seed: 42
+policy: two-room/tworoom/lejepa
+inference_precision: fp16
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+  - action
+  - proprio
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: tworoom
+  callables:
+  - method: _set_state
+    args:
+      state:
+        value: proprio
+  - method: _set_goal_state
+    args:
+      goal_state:
+        value: goal_proprio
+output:
+  filename: tworoom_results.txt
+
+==== RESULTS ====
+metrics: {'success_rate': 90.0, 'episode_successes': array([ True, False,  True, False,  True,  True,  True,  True, False,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True, False,  True,  True,  True,  True,  True,
+        True,  True,  True,  True, False,  True,  True,  True,  True,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True]), 'seeds': None}
+evaluation_time: 43.69562244415283 seconds
+inference_precision: fp16
+inference_compile_target: predictor
+inference_compile_mode: reduce-overhead
+
+==== CONFIG ====
+cache_dir: null
+solver:
+  _target_: stable_worldmodel.solver.CEMSolver
+  model: ???
+  batch_size: 1
+  num_samples: 300
+  var_scale: 1.0
+  n_steps: 30
+  topk: 30
+  device: cuda
+  seed: ${seed}
+world:
+  env_name: swm/TwoRoom-v1
+  num_envs: ${eval.num_eval}
+  max_episode_steps: 100
+  history_size: 1
+  frame_skip: 1
+seed: 42
+policy: two-room/tworoom/lejepa
+inference_precision: fp16
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+  - action
+  - proprio
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: tworoom
+  callables:
+  - method: _set_state
+    args:
+      state:
+        value: proprio
+  - method: _set_goal_state
+    args:
+      goal_state:
+        value: goal_proprio
+output:
+  filename: tworoom_results.txt
+
+==== RESULTS ====
+metrics: {'success_rate': 88.0, 'episode_successes': array([ True, False,  True, False,  True,  True,  True,  True, False,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True, False,  True,  True,  True,  True,  True,
+        True,  True,  True,  True, False,  True,  True,  True,  True,
+        True,  True, False,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True]), 'seeds': None}
+evaluation_time: 42.99847435951233 seconds
+inference_precision: fp16
+inference_compile_target: predictor
+inference_compile_mode: reduce-overhead
+
+==== CONFIG ====
+cache_dir: null
+solver:
+  _target_: stable_worldmodel.solver.CEMSolver
+  model: ???
+  batch_size: 1
+  num_samples: 300
+  var_scale: 1.0
+  n_steps: 30
+  topk: 30
+  device: cuda
+  seed: ${seed}
+world:
+  env_name: swm/TwoRoom-v1
+  num_envs: ${eval.num_eval}
+  max_episode_steps: 100
+  history_size: 1
+  frame_skip: 1
+seed: 42
+policy: two-room/tworoom/lejepa
+inference_precision: fp16
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+  - action
+  - proprio
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: tworoom
+  callables:
+  - method: _set_state
+    args:
+      state:
+        value: proprio
+  - method: _set_goal_state
+    args:
+      goal_state:
+        value: goal_proprio
+output:
+  filename: tworoom_results.txt
+
+==== RESULTS ====
+metrics: {'success_rate': 88.0, 'episode_successes': array([ True, False,  True, False,  True,  True,  True,  True, False,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True, False,  True,  True,  True,  True,  True,
+        True,  True,  True,  True, False,  True,  True,  True,  True,
+        True,  True, False,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True]), 'seeds': None}
+evaluation_time: 43.14276576042175 seconds
+inference_precision: fp16
+inference_compile_target: predictor
+inference_compile_mode: reduce-overhead
+
+==== CONFIG ====
+cache_dir: null
+solver:
+  _target_: stable_worldmodel.solver.CEMSolver
+  model: ???
+  batch_size: 1
+  num_samples: 300
+  var_scale: 1.0
+  n_steps: 30
+  topk: 30
+  device: cuda
+  seed: ${seed}
+world:
+  env_name: swm/TwoRoom-v1
+  num_envs: ${eval.num_eval}
+  max_episode_steps: 100
+  history_size: 1
+  frame_skip: 1
+seed: 42
+policy: two-room/tworoom/lejepa
+inference_precision: fp16
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+  - action
+  - proprio
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: tworoom
+  callables:
+  - method: _set_state
+    args:
+      state:
+        value: proprio
+  - method: _set_goal_state
+    args:
+      goal_state:
+        value: goal_proprio
+output:
+  filename: tworoom_results.txt
+
+==== RESULTS ====
+metrics: {'success_rate': 88.0, 'episode_successes': array([ True, False,  True, False,  True,  True,  True,  True, False,
+        True,  True,  True,  True,  True,  True,  True,  True,  True,
+        True,  True,  True, False,  True,  True,  True,  True,  True,
+        True,  True,  True,  True, False,  True,  True,  True,  True,
+        True,  True, False,  True,  True,  True,  True,  True,  True,
+        True,  True,  True,  True,  True]), 'seeds': None}
+evaluation_time: 43.71034002304077 seconds
+inference_precision: fp16
+inference_compile_target: predictor
+inference_compile_mode: reduce-overhead