feat: add pusht imf full-attention config

docs/superpowers/plans/2026-03-27-pusht-imf-fullattn-implementation.md

@@ -0,0 +1,60 @@
+# PushT iMF Full-Attention Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Add a separate full-attention PushT image iMF config, commit/push it on a new branch, and launch the 9-run 350-epoch architecture sweep across 3 GPUs.
+
+**Architecture:** Keep the existing causal iMF path untouched and add a standalone full-attention config that only flips `policy.causal_attn=false` while retaining one-step iMF inference and SwanLab-safe naming. Reuse the previous 9-run architecture matrix and balanced 3-queue scheduling across local 5090 plus 5880 GPU0/GPU1.
+
+**Tech Stack:** Hydra, Diffusion Policy iMF image workspace, SwanLab, uv env, local shell + trusted remote 5880 over SSH.
+
+---
+
+### Task 1: Add full-attention iMF config with TDD
+
+**Files:**
+- Create: `image_pusht_diffusion_policy_dit_imf_fullattn.yaml`
+- Modify: `tests/test_pusht_swanlab_config.py`
+
+- [ ] Write a failing config regression test asserting the new config uses SwanLab-safe naming and `policy.causal_attn == False`.
+- [ ] Run the targeted pytest command and verify it fails because the config does not exist yet.
+- [ ] Add the minimal full-attention config by composing from the existing PushT image iMF config and overriding only `exp_name` and `policy.causal_attn=false`.
+- [ ] Re-run the targeted pytest and verify it passes.
+
+### Task 2: Verify the new config
+
+**Files:**
+- Read: `image_pusht_diffusion_policy_dit_imf_fullattn.yaml`
+
+- [ ] Run `train.py --help` for the new config.
+- [ ] Run a real `training.debug=true` smoke test locally to confirm the training path is valid.
+
+### Task 3: Commit and push the new branch
+
+**Files:**
+- Commit only the new config/test/plan files needed for the full-attention experiment chain.
+
+- [ ] Run verification commands again before commit.
+- [ ] Commit with a focused message.
+- [ ] Push `feat/pusht-imf-fullattn` to origin.
+
+### Task 4: Launch the 9-run sweep
+
+**Files:**
+- Write queue scripts and logs under `data/run_logs/` locally and on 5880.
+- Write outputs under `data/outputs/` locally and on 5880.
+
+- [ ] Use the same matrix as the prior iMF sweep: `n_emb ∈ {128,256,384}`, `n_layer ∈ {6,12,18}`, `seed=42`.
+- [ ] Set `training.num_epochs=350` for all 9 runs.
+- [ ] Encode `fullattn` in every `exp_name`, `logging.name`, and run directory to avoid collisions.
+- [ ] Balance the 9 runs across local 5090, 5880 GPU0, and 5880 GPU1 as three serial queues.
+- [ ] Sync the new config to the remote smoke repo before launching remote queues.
+
+### Task 5: Monitor and auto-summarize
+
+**Files:**
+- Read local and remote pid files, logs, outputs, checkpoints.
+
+- [ ] Start an xhigh monitoring agent that polls all three queues.
+- [ ] On completion, parse all 9 `logs.json.txt` files and rank by max `test_mean_score`.
+- [ ] Report embedding/layer trends and the best configuration.

docs/superpowers/specs/2026-03-27-pusht-imf-fullattn-design.md

@@ -0,0 +1,107 @@
+# PushT Image iMF Full-Attention Sweep Design
+
+## Goal
+在一个独立新分支上，为 PushT 图像 iMF 路线新增 **full-attention** 变体（关闭因果注意力），并按与之前相同的架构扫描网格运行 **9 组实验**，每组训练 **350 epochs**。所有实验完成后，提取每组 **`max(test_mean_score)`** 并输出完整排名和趋势总结。
+
+## Scope
+本次工作仅覆盖：
+1. 在不影响现有因果版 iMF 路线的前提下，新增 full-attention 实验链路；
+2. 对 `n_emb ∈ {128, 256, 384}` 与 `n_layer ∈ {6, 12, 18}` 的 9 组组合做 350-epoch 扫描；
+3. 在本机 5090 与 5880 双卡上做三路并行调度；
+4. 在全部实验完成后自动汇总结果并直接向用户汇报。
+
+不在本次范围内：
+- 不替换或删除现有因果版 iMF 配置；
+- 不改动已有 DiT baseline 实现；
+- 不做多 seed 扩展；
+- 不额外增加视频记录。
+
+## Design Choice
+采用“**新增独立配置 + 新分支**”的方式，而不是覆盖现有 iMF 默认配置。
+
+原因：
+- 现有因果版 iMF 已完成实验与结果记录，保持不动更利于对照；
+- full-attention 作为新的实验链路，使用独立配置更易复现；
+- 运行时只需要通过配置切换 `policy.causal_attn=false`，不需要重新设计 iMF 算法本身。
+
+## Configuration Design
+新增一个独立配置文件，例如：
+- `image_pusht_diffusion_policy_dit_imf_fullattn.yaml`
+
+其职责：
+- 继承当前 PushT image iMF 配置链路；
+- 保持 iMF 单步推理、SwanLab 标量记录、无视频记录；
+- 显式设置：
+  - `policy.causal_attn=false`
+  - `policy.n_head=1`
+- 保持其余 iMF 训练语义不变。
+
+SwanLab 命名延续当前修复后的策略：
+- `logging.name=${exp_name}`
+- `logging.resume=false`
+- `logging.id=null`
+- `logging.group=${exp_name}` 或统一 sweep group override
+
+## Code Change Strategy
+优先最小改动：
+- 若当前 `IMFTransformerForDiffusion` 已支持 `causal_attn=False` 分支，则不改核心算法，仅通过新配置关闭因果 mask；
+- 如需补充回归验证，则新增针对 full-attention 配置/掩码行为的最小测试；
+- 不改变已有因果版实验配置和已有测试语义。
+
+## Experiment Matrix
+实验网格固定为：
+
+- `n_emb=128, n_layer=6`
+- `n_emb=128, n_layer=12`
+- `n_emb=128, n_layer=18`
+- `n_emb=256, n_layer=6`
+- `n_emb=256, n_layer=12`
+- `n_emb=256, n_layer=18`
+- `n_emb=384, n_layer=6`
+- `n_emb=384, n_layer=12`
+- `n_emb=384, n_layer=18`
+
+统一设置：
+- `training.num_epochs=350`
+- `training.resume=false`
+- `seed=42`
+- PushT image 数据路径不变
+- 指标以 **`logs.json.txt` 中 `test_mean_score` 的最大值** 为准
+
+## Scheduling Design
+使用三路串行队列并行执行 9 个实验：
+
+- 本机 5090：1 个顺序队列
+- 5880 GPU0：1 个顺序队列
+- 5880 GPU1：1 个顺序队列
+
+分配原则：
+- 延续按 `n_emb × n_layer` 近似平衡工作量；
+- 每张卡同一时刻只跑 1 个实验；
+- 队列脚本负责“前一个结束后自动启动下一个”。
+
+## Monitoring Design
+继续采用“**训练队列脚本 + 监控 agent**”双层机制：
+
+1. **实际调度**由本地/远端队列脚本负责；
+2. **监控**由一个 xhigh 子 agent 轮询：
+   - 读取 pid 状态
+   - 检查 master log
+   - 检查每个 run 的 `logs.json.txt`
+   - 判断是否卡死/失败/全部完成
+3. 一旦全部完成，监控 agent 直接返回：
+   - 9 组实验的最终 epoch
+   - 每组 `max(test_mean_score)`
+   - 排名表
+   - embedding / layer 趋势总结
+
+本次要求下，agent 在收到全部完成信号后应直接向主会话回报结果，不等待用户再次提醒。
+
+## Success Criteria
+满足以下条件即视为完成：
+1. full-attention iMF 配置在新分支上可运行；
+2. 9 组 350-epoch 实验全部完成；
+3. 不记录仿真视频，只记录标量；
+4. SwanLab 运行命名不冲突；
+5. 输出 9 组实验 `max(test_mean_score)` 的完整汇总与结论；
+6. 全部实验结束后主会话可直接给用户最终总结。

image_pusht_diffusion_policy_dit_imf_fullattn.yaml

@@ -0,0 +1,33 @@
+defaults:
+  - diffusion_policy/config/train_diffusion_transformer_hybrid_workspace@_here_
+  - override /diffusion_policy/config/task@task: pusht_image
+  - _self_
+
+exp_name: pusht_image_dit_imf_fullattn
+
+policy:
+  _target_: diffusion_policy.policy.imf_transformer_hybrid_image_policy.IMFTransformerHybridImagePolicy
+  num_inference_steps: 1
+  n_head: 1
+  causal_attn: false
+
+logging:
+  backend: swanlab
+  mode: online
+  name: ${exp_name}
+  resume: false
+  tags: ["${name}", "${task_name}", "${exp_name}", "swanlab"]
+  id: null
+  group: ${exp_name}
+
+dataloader:
+  num_workers: 0
+
+val_dataloader:
+  num_workers: 0
+
+task:
+  env_runner:
+    n_envs: 1
+    n_test_vis: 0
+    n_train_vis: 0

tests/test_pusht_swanlab_config.py

@@ -30,3 +30,15 @@ def test_image_pusht_dit_imf_swanlab_config_uses_exp_name_and_no_resume_collisio
     assert cfg.logging.resume is False
     assert cfg.logging.id is None
     assert cfg.logging.group == cfg.exp_name
+
+
+def test_image_pusht_dit_imf_fullattn_config_uses_exp_name_and_disables_causal_attention():
+    cfg = _load_cfg('image_pusht_diffusion_policy_dit_imf_fullattn.yaml')
+
+    assert cfg.logging.backend == 'swanlab'
+    assert cfg.logging.mode == 'online'
+    assert cfg.logging.name == cfg.exp_name
+    assert cfg.logging.resume is False
+    assert cfg.logging.id is None
+    assert cfg.logging.group == cfg.exp_name
+    assert cfg.policy.causal_attn is False