# Short History Attention Debug 本次修改实现了一个默认关闭的 debug hook,用来分析 current token 到 short history 上一帧 token 的对应关系。 ## 背景结论 默认设置下: - current chunk 是 33 个视频帧,对应 9 个 latent frames。 - 默认视频分辨率是 384x640。 - VAE latent 分辨率是 48x80。 - transformer patch 是 `(1, 2, 2)`,所以 current token grid 是 `9 x 24 x 40`。 - short history 经过 `patch_short` 后是 `2 x 24 x 40`。 - 本次只在 short history 的上一帧里找对应 token,也就是每个 current latent frame 的 `24 x 40` tokens 对应 previous short frame 的 `24 x 40` tokens。 ## 已实现内容 新增 `short_attn_debug` pipeline 参数。开启后,attention hook 会在指定 block、chunk、denoise step、current frame 上导出 current token 到 previous short token 的 top-k 匹配。 实现位置: - `helios/diffusers_version/transformer_helios_diffusers.py` - `helios/diffusers_version/pipeline_helios_diffusers.py` - `helios/modules/transformer_helios.py` - `helios/pipelines/pipeline_helios.py` - `tools/visualize_short_attention_matches.py` 导出的 artifact 默认包含: - `match_yx`: 每个 current token 匹配到的 previous short token 坐标 - `query_yx`: current token 自身坐标 - `displacement_yx`: token 位移 - `topk_indices`: top-k 匹配 index - `topk_scores`: top-k raw attention scores - `top1_score` - `top2_score` - `margin`: `top1_score - top2_score` 不会保存完整 `960 x 960` attention matrix,避免显存和磁盘开销过大。 ## 使用示例 在 pipeline 调用里加: ```python output = pipe( prompt=prompt, negative_prompt=negative_prompt, height=384, width=640, num_frames=99, short_attn_debug={ "output_dir": "short_attn_debug", "blocks": [30], "steps": ["last"], "chunks": [0], "current_frame": -1, "prev_short_frame": 1, "pass_names": ["cond"], "topk": 2, "query_chunk_size": 128, }, ).frames[0] ``` 默认推荐先看: - block `30` - denoise last step - chunk `0` - current latent frame `-1` - previous short frame `1` - cond pass ## 可视化 生成图: ```bash python tools/visualize_short_attention_matches.py \ short_attn_debug/short_attn_chunk0_step49_block30_frame8_cond.pt \ --output-dir short_attn_debug \ --stride 4 ``` 输出: - `*_flow_quiver.png`: 稀疏 token flow 箭头图 - `*_displacement_heatmap.png`: 位移大小热图 - `*_confidence_heatmap.png`: `top1 - top2` margin 热图 - `*_top1_score_heatmap.png`: top1 raw score 热图 ## 当前实现细节 匹配计算使用 attn1 里已经 projection、norm、rotary 后的 Q/K: ```text score = Q_current_frame @ K_previous_short_frame.T / sqrt(head_dim) ``` 然后对 heads 做平均,再取 top-k。 为了控制开销,计算按 query chunk 分块进行。默认每次处理 128 个 query tokens。 ## 未做但可继续扩展的方案 1. full softmax attention heatmap 当前保存的是 top-k raw scores,不保存完整 softmax 分布。后续如果要做单点 token viewer,可以只对用户点击的 token 计算完整 `24 x 40` softmax heatmap。 2. click-to-inspect viewer 可以做一个交互式界面:左边点 current token,右边显示它对 previous short frame 的 `24 x 40` score/attention heatmap,并框出 top-k。 3. 多层对比 当前支持配置多个 blocks,例如 `[10, 20, 30, 39]`。后续可以把这些层的 flow 和 confidence 并排画出来,观察 correspondence 在不同深度的变化。 4. 多 denoise step 对比 当前支持 `steps=["first", "last"]` 或具体 step index。后续可以把 first/mid/last step 的图并排,观察匹配是否从噪声期到收敛期变稳定。 5. 多 current latent frame 可视化 当前推荐只看一个 current frame,例如 `-1`。后续可以对 `0..8` 全部导出,生成 grid 或小视频。 6. warped previous frame 可以用 token correspondence 把 previous short frame 的 patch/grid warp 到 current frame,再与 current frame 做差分。这比 flow 图更接近视觉验证,但实现成本更高。 7. mutual nearest neighbor 过滤 当前是 current -> previous short 的单向 top-k。后续可加 previous short -> current,再保留互为 top-1 的 matches,用来过滤不稳定对应。 8. one-to-one matching 如果需要严格的一对一 token matching,可以在 score matrix 上做 Hungarian 或 optimal transport。但 `960 x 960` 会比当前 top-k 更重,而且未必符合 attention 本身的多对一行为。