保存的优化

attention.py — 3处改动：
  1. __init__ 添加 _kv_fused = False 标志
  2.新增 fuse_kv() 方法：将 to_k + to_v → to_kv，同时处理 _ip/_as/_aa 辅助 KV 对
  2. bmm_forward 两个分支加_kv_fused 判断，用to_kv().chunk(2, dim=-1) 替代分别调用

.claude/settings.local.json

@@ -0,0 +1,15 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(conda env list:*)",
+      "Bash(mamba env:*)",
+      "Bash(micromamba env list:*)",
+      "Bash(echo:*)",
+      "Bash(git show:*)",
+      "Bash(nvidia-smi:*)",
+      "Bash(conda activate unifolm-wma)",
+      "Bash(conda info:*)",
+      "Bash(direnv allow:*)"
+    ]
+  }
+}

.envrc

@@ -0,0 +1,2 @@
+eval "$(conda shell.bash hook 2>/dev/null)"
+conda activate unifolm-wma

.gitignore

@@ -55,7 +55,6 @@ coverage.xml
 *.pot
 
 # Django stuff:
-*.log
 local_settings.py
 db.sqlite3
 
@@ -121,10 +120,17 @@ localTest/
 fig/
 figure/
 *.mp4
-*.json
+
 Data/ControlVAE.yml
 Data/Misc
 Data/Pretrained
 Data/utils.py
 Experiment/checkpoint
 Experiment/log
+
+*.ckpt
+
+*.0
+ckpts/unifolm_wma_dual.ckpt.prepared.pt
+trt_engines/video_backbone.engine
+trt_engines/video_backbone.onnx

ckpts/LICENSE

@@ -0,0 +1,439 @@
+Attribution-NonCommercial-ShareAlike 4.0 International
+
+Copyright (c) 2016-2025 HangZhou YuShu TECHNOLOGY CO.,LTD. ("Unitree Robotics")
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+    wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More considerations
+     for the public:
+    wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
+Public License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial-ShareAlike 4.0 International Public License
+("Public License"). To the extent this Public License may be
+interpreted as a contract, You are granted the Licensed Rights in
+consideration of Your acceptance of these terms and conditions, and the
+Licensor grants You such rights in consideration of benefits the
+Licensor receives from making the Licensed Material available under
+these terms and conditions.
+
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. BY-NC-SA Compatible License means a license listed at
+     creativecommons.org/compatiblelicenses, approved by Creative
+     Commons as essentially the equivalent of this Public License.
+
+  d. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+
+  e. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  f. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  g. License Elements means the license attributes listed in the name
+     of a Creative Commons Public License. The License Elements of this
+     Public License are Attribution, NonCommercial, and ShareAlike.
+
+  h. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  i. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  j. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  k. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  l. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  m. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  n. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. Additional offer from the Licensor -- Adapted Material.
+               Every recipient of Adapted Material from You
+               automatically receives an offer from the Licensor to
+               exercise the Licensed Rights in the Adapted Material
+               under the conditions of the Adapter's License You apply.
+
+            c. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+  b. ShareAlike.
+
+     In addition to the conditions in Section 3(a), if You Share
+     Adapted Material You produce, the following conditions also apply.
+
+       1. The Adapter's License You apply must be a Creative Commons
+          license with the same License Elements, this version or
+          later, or a BY-NC-SA Compatible License.
+
+       2. You must include the text of, or the URI or hyperlink to, the
+          Adapter's License You apply. You may satisfy this condition
+          in any reasonable manner based on the medium, means, and
+          context in which You Share Adapted Material.
+
+       3. You may not offer or impose any additional or different terms
+          or conditions on, or apply any Effective Technological
+          Measures to, Adapted Material that restrict exercise of the
+          rights granted under the Adapter's License You apply.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material,
+     including for purposes of Section 3(b); and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

ckpts/README.md

@@ -0,0 +1,38 @@
+---
+tags:
+- robotics
+---
+
+# UnifoLM-WMA-0: A World-Model-Action (WMA) Framework under UnifoLM Family
+<p style="font-size: 1.2em;">
+    <a href="https://unigen-x.github.io/unifolm-world-model-action.github.io"><strong>Project Page</strong></a> | 
+    <a href="https://github.com/unitreerobotics/unifolm-world-model-action"><strong>Code</strong></a> |
+    <a href="https://huggingface.co/unitreerobotics/datasets"><strong>Dataset</strong></a> 
+  </p>
+<div align="center">
+  <div align="justify">
+    <b>UnifoLM-WMA-0</b> is Unitree‘s first open-source world-model–action architecture spanning multiple types of robotic embodiments, designed specifically for general-purpose robot learning. Its core component is a world-model capable of understanding the physical interactions between robots and the environments. This world-model provides two key functions: (a) <b>Simulation Engine</b> – operates as an interactive simulator to generate synthetic data for robot learning; (b) <b>Policy Enhancement</b> – connects with an action head and, by predicting future interaction processes with the world-model, further optimizes decision-making performance.
+  </div>
+</div>
+
+## 🦾 Real Robot Deployment
+| <img src="assets/real_z1_stackbox.gif" style="border:none;box-shadow:none;margin:0;padding:0;" /> | <img src="assets/real_dual_stackbox.gif" style="border:none;box-shadow:none;margin:0;padding:0;" /> |
+|:---:|:---:|
+| <img src="assets/real_cleanup_pencils.gif" style="border:none;box-shadow:none;margin:0;padding:0;" /> | <img src="assets/real_g1_pack_camera.gif" style="border:none;box-shadow:none;margin:0;padding:0;" /> |
+
+**Note: the top-right window shows the world model’s prediction of future environmental changes.**
+
+## License
+The model is released under the CC BY-NC-SA 4.0 license as found in the [LICENSE](https://huggingface.co/unitreerobotics/UnifoLM-WMA-0/blob/main/LICENSE). You are responsible for ensuring that your use of Unitree AI Models complies with all applicable laws.
+
+## Model Architecture
+![Demo](assets/world_model_interaction.gif)
+
+## Citation
+```
+@misc{unifolm-wma-0,
+  author = {Unitree},
+  title  = {UnifoLM-WMA-0: A World-Model-Action (WMA) Framework under UnifoLM Family},
+  year   = {2025},
+}
+```

configs/inference/world_model_interaction.yaml

@@ -222,7 +222,7 @@ data:
     test:
       target: unifolm_wma.data.wma_data.WMAData
       params:
-        data_dir: '/path/to/unifolm-world-model-action/examples/world_model_interaction_prompts'
+        data_dir: '/home/qhy/unifolm-world-model-action/examples/world_model_interaction_prompts'
         video_length: ${model.params.wma_config.params.temporal_length}
         frame_stride: 2
         load_raw_resolution: True

psnr_score_for_challenge.py

@@ -0,0 +1,89 @@
+import os
+import glob
+import numpy as np
+import json
+from argparse import ArgumentParser, ArgumentDefaultsHelpFormatter
+from tqdm import tqdm
+from moviepy.video.io.VideoFileClip import VideoFileClip
+import PIL.Image
+
+
+def calculate_psnr(img1, img2):
+    mse = np.mean((img1.astype(np.float64) - img2.astype(np.float64)) ** 2)
+    if mse == 0:
+        return float('inf')
+    max_pixel = 255.0
+    psnr = 20 * np.log10(max_pixel / np.sqrt(mse))
+    return psnr
+
+
+def process_video_psnr(gt_path, pred_path):
+    try:
+        clip_gt = VideoFileClip(gt_path)
+        clip_pred = VideoFileClip(pred_path)
+
+        fps = min(clip_gt.fps, clip_pred.fps)
+        duration = min(clip_gt.duration, clip_pred.duration)
+
+        time_points = np.arange(0, duration, 1.0 / fps)
+
+        video_psnrs = []
+
+        for t in time_points:
+            frame_gt = clip_gt.get_frame(t)
+            frame_pred = clip_pred.get_frame(t)
+
+            img_gt = PIL.Image.fromarray(frame_gt).resize((256, 256), PIL.Image.Resampling.BILINEAR)
+            img_pred = PIL.Image.fromarray(frame_pred).resize((256, 256), PIL.Image.Resampling.BILINEAR)
+
+            psnr = calculate_psnr(np.array(img_gt), np.array(img_pred))
+            video_psnrs.append(psnr)
+
+        clip_gt.close()
+        clip_pred.close()
+
+        return np.mean(video_psnrs) if video_psnrs else 0.0
+
+    except Exception as e:
+        print(f"Error processing {os.path.basename(gt_path)}: {e}")
+        return None
+
+
+def main():
+    parser = ArgumentParser(formatter_class=ArgumentDefaultsHelpFormatter)
+    parser.add_argument('--gt_video', type=str, required=True, help='path to reference videos')
+    parser.add_argument('--pred_video', type=str, required=True, help='path to pred videos')
+    parser.add_argument('--output_file', type=str, default=None, help='path to output file')
+    args = parser.parse_args()
+
+    if not os.path.exists(args.gt_video):
+        print(f"Error: GT video not found at {args.gt_video}")
+        return
+    if not os.path.exists(args.pred_video):
+        print(f"Error: Pred video not found at {args.pred_video}")
+        return
+
+    print(f"Comparing:\nRef: {args.gt_video}\nPred: {args.pred_video}")
+
+    v_psnr = process_video_psnr(args.gt_video, args.pred_video)
+
+    if v_psnr is not None:
+        print("-" * 30)
+        print(f"Video PSNR: {v_psnr:.4f} dB")
+        print("-" * 30)
+
+        if args.output_file:
+            result = {
+                "gt_video": args.gt_video,
+                "pred_video": args.pred_video,
+                "psnr": v_psnr
+            }
+            with open(args.output_file, 'w') as f:
+                json.dump(result, f, indent=4)
+            print(f"Result saved to {args.output_file}")
+    else:
+        print("Failed to calculate PSNR.")
+
+
+if __name__ == '__main__':
+    main()

pyproject.toml

@@ -19,13 +19,13 @@ dependencies = [
     "pytorch-lightning==1.9.3",
     "pyyaml==6.0",
     "setuptools==65.6.3",
-    "torch==2.3.1",
+    #"torch==2.3.1",
-    "torchvision==0.18.1",
+    #"torchvision==0.18.1",
     "tqdm==4.66.5",
     "transformers==4.40.1",
     "moviepy==1.0.3",
     "av==12.3.0",
-    "xformers==0.0.27",
+    #"xformers==0.0.27",
     "gradio==4.39.0",
     "timm==0.9.10",
     "scikit-learn==1.5.1",

record.md

@@ -0,0 +1,208 @@
+ 好，下面是完整的 decode 优化讲解。                                                                                                    
+                                                                                                                                        
+  ---                                                                                                                                   
+  背景：原始 decode 流程的问题
+                                                                                                                                        
+  原始代码中，每轮迭代的 world model 交互会调用 image_guided_synthesis_sim_mode，内部对DDIM 采样得到的 latent 做全量 VAE decode：     
+
+  # ddpms.py decode_core, perframe_ae=True
+  for index in range(z.shape[0]):  # 逐帧循环，16帧就调16次
+      frame_z = 1. / self.scale_factor * z[index:index+1]
+      frame_result = self.first_stage_model.decode(frame_z)
+      results.append(frame_result)
+
+  假设 8 轮迭代、每轮 16 帧，WM 路径 decode 128 次，policy 路径再 decode 128 次，共 256 次 VAE decode。每次都是同步阻塞的。
+
+  decode 完的视频还会被逐轮写成 mp4 + tensorboard，产生大量磁盘 I/O。最后还要把所有轮的 decoded video 在内存中torch.cat
+  拼接，再写一次完整视频。
+
+  ---
+  优化1：decode_video 开关——按需跳过 decode
+
+  文件： world_model_interaction.py函数 image_guided_synthesis_sim_mode
+
+  改动： 给函数加decode_video 参数（默认 False），返回值增加 raw samples：
+
+  def image_guided_synthesis_sim_mode(...,
+          decode_video: bool = False,  # 新增
+          ...) -> tuple[Tensor | None, Tensor, Tensor, Tensor | None]:
+
+      samples = None
+      if ddim_sampler is not None:
+          samples, actions, states, intermedia = ddim_sampler.sample(...)if decode_video:  # 条件 decode
+              batch_images = model.decode_first_stage(samples)
+              batch_variants = batch_images
+
+      return batch_variants, actions, states, samples# 多返回 samples
+
+  调用侧：
+  - Policy 路径：由 CLI 参数 --fast_policy_no_decode 控制，只需要 action 时可跳过 decode
+  - WM 交互路径：传decode_video=False，只拿 raw latent
+
+  效果： WM 路径每轮省掉 16 帧全量 decode。
+
+  ---
+  优化2：只decode observation 需要的帧
+
+  问题： WM 跳过了全量 decode，但下一轮的CLIP embedding 需要 pixel-space 图像做 observation。
+
+  改动： 只decode exe_steps 帧（通常 1帧），而不是全部 16 帧：
+
+  # WM 调用，不做全量 decode
+  pred_videos_1, _, pred_states, wm_samples = image_guided_synthesis_sim_mode(
+      ..., decode_video=False)
+
+  # 只 decode exe_steps 帧给 observation
+  obs_pixels = model.decode_first_stage(
+      wm_samples[:, :, :args.exe_steps, :, :])
+
+  for idx in range(args.exe_steps):
+      observation = {
+          'observation.images.top':obs_pixels[0, :, idx:idx + 1].permute(1, 0, 2, 3),
+          ...
+      }
+      cond_obs_queues = populate_queues(cond_obs_queues, observation)
+
+  关键细节： 必须逐帧填充 observation queue（idx:idx+1），不能全用最后一帧，否则 CLIP embedding 输入变了会影响精度。
+
+  效果： 每轮从 decode 16 帧降到 decode exe_steps 帧（省15 帧/轮）。
+
+  ---
+  优化3：decode stream——GPU 上并行 decode 和 UNet
+
+  问题： 写入最终视频仍需要完整 segment 的 pixel，这部分 decode 还是要做。
+
+  思路： 用独立 CUDA stream 做 segment decode，和下一轮 UNet 推断在 GPU 上并行。
+
+  改动：
+
+  初始化：
+  decode_stream = torch.cuda.Stream(device=device)
+  pending_decode = None
+
+  循环尾部：
+  # 收集上一轮 decode 结果
+  if pending_decode is not None:
+      decode_stream.synchronize()
+      write_q.put(pending_decode.cpu())
+      pending_decode = None
+
+  # 在 decode stream 上启动当前轮 segment decode（不阻塞主线程）
+  latent_slice = wm_samples[:, :, :args.exe_steps]
+  decode_stream.wait_stream(torch.cuda.current_stream())  # 确保 latent 就绪
+  with torch.cuda.stream(decode_stream):
+      pending_decode = model.decode_first_stage(latent_slice)
+  # 主线程立即进入下一轮 UNet
+
+  循环结束后收集最后一轮：
+  if pending_decode is not None:
+      decode_stream.synchronize()
+      write_q.put(pending_decode.cpu())
+
+  原理： decode_stream.wait_stream() 建立 stream间依赖，确保 latent 产出后才开始 decode。两个 stream 的 kernel 可以被GPU
+  调度器交错执行。
+
+  效果： segment decode 时间被下一轮 UNet 推断掩盖。
+
+  ---
+  优化4：Writer 进程——CPU 工作跨进程并行
+
+  问题： decode 完的tensor 需要转numpy + cv2 编码写盘，这是 CPU 密集型操作，Python GIL 限制线程并行。
+
+  改动：
+
+  辅助函数（主进程和子进程都能调用）：
+  def _video_tensor_to_frames(video: Tensor) -> np.ndarray:
+      video = torch.clamp(video.float(), -1., 1.)
+      n = video.shape[0]
+      video = video.permute(2, 0, 1, 3, 4)
+      frame_grids = [
+          torchvision.utils.make_grid(f, nrow=int(n), padding=0) for f in video
+      ]
+      grid = torch.stack(frame_grids, dim=0)
+      grid = ((grid + 1.0) / 2.0 * 255).to(torch.uint8).permute(0, 2, 3, 1)
+      return grid.numpy()[:, :, :, ::-1]  # RGB → BGR
+
+  Writer 进程：
+  def _video_writer_process(q: mp.Queue, filename: str, fps: int):
+      vwriter = None
+      while True:
+          item = q.get()
+          if item is None:  # sentinel，退出
+              break
+          frames = _video_tensor_to_frames(item)
+          if vwriter is None:
+              h, w = frames.shape[1], frames.shape[2]
+              fourcc = cv2.VideoWriter_fourcc(*'mp4v')
+              vwriter = cv2.VideoWriter(filename, fourcc, fps, (w, h))
+          for f in frames:
+              vwriter.write(f)
+      if vwriter is not None:
+          vwriter.release()
+
+  主进程启动 writer：
+  write_q = mp.Queue()
+  writer_proc = mp.Process(target=_video_writer_process,
+                           args=(write_q, sample_full_video_file, args.save_fps))
+  writer_proc.start()
+
+  主进程通过 write_q.put(tensor.cpu()) 发送数据，循环结束发None sentinel 并join()。
+
+  效果：
+  - tensor→numpy 转换和cv2 编码不占主进程 CPU 时间
+  - 不受 GIL 限制
+  - cv2.VideoWriter 增量写入，不攒内存，不做最终 torch.cat
+  - Queue 自带背压，writer 处理不过来时 put 自然阻塞，不会 OOM
+
+  ---
+  同时删除的冗余 I/O
+
+  - 每轮迭代的 WM 中间tensorboard log（log_to_tensorboard_async）
+  - 每轮迭代的 WM 中间 mp4（save_results_async）
+  - 每轮迭代的 policy 中间 mp4
+  - 最终的 torch.cat + 一次性 save_results_async 写完整视频
+
+  ---
+  总结：优化前后对比
+
+  ┌─────────────────────┬────────────────────────┬───────────────────────────┐
+  │                     │          原始          │          优化后           │
+  ├─────────────────────┼────────────────────────┼───────────────────────────┤
+  │ WM decode帧数/轮    │ 16                     │ exe_steps（通常1）        │
+  ├─────────────────────┼────────────────────────┼───────────────────────────┤
+  │ segment decode 阻塞 │ 同步阻塞主线程         │ decode stream，GPU并行    │
+  ├─────────────────────┼────────────────────────┼───────────────────────────┤
+  │ tensor→numpy+写盘   │ 主线程同步/线程池(GIL) │ 独立进程，无GIL           │
+  ├─────────────────────┼────────────────────────┼───────────────────────────┤
+  │ 视频写入方式        │ 攒内存→cat→一次写      │ cv2增量追加写             │
+  ├─────────────────────┼────────────────────────┼───────────────────────────┤
+  │ 中间I/O             │ 每轮2×mp4+2×tb         │ 无                        │
+  ├─────────────────────┼────────────────────────┼───────────────────────────┤
+  │ 内存占用            │ 随迭代数线性增长       │ 恒定（不攒decoded video） │
+  └─────────────────────┴────────────────────────┴───────────────────────────┘
+
+✻ Cooked for 1m 31s                                                                                                                     
+  
+❯ 目前哪些优化是默认开启的                                                                                                              
+                                                                                                                                      
+● 让我检查各个优化的开关状态。                                                                                                          
+                                                                                                                                        
+● Searched for 2 patterns (ctrl+o to expand)                                                                                            
+                                                                                                                                      
+● 当前各优化的默认状态：                                                                                                                
+                                                                                                                                        
+  ┌───────────────────────────────────┬────────────┬────────────────────────────────────────────────────────┐                           
+  │               优化                │ 默认开启？ │                        控制方式                        │                         
+  ├───────────────────────────────────┼────────────┼────────────────────────────────────────────────────────┤                           
+  │ WM路径跳过全量decode              │ ✅ 是      │ 硬编码 decode_video=False（789行）                     │                           
+  ├───────────────────────────────────┼────────────┼────────────────────────────────────────────────────────┤
+  │ 只decode exe_steps帧给observation │ ✅ 是      │ 无开关，直接生效（792行）                              │
+  ├───────────────────────────────────┼────────────┼────────────────────────────────────────────────────────┤
+  │ decode stream GPU并行             │ ✅ 是      │ 无开关，直接生效（824-827行）                          │
+  ├───────────────────────────────────┼────────────┼────────────────────────────────────────────────────────┤
+  │ Writer进程增量写盘                │ ✅ 是      │ 无开关，直接生效（672-675行）                          │
+  ├───────────────────────────────────┼────────────┼────────────────────────────────────────────────────────┤
+  │ 删除中间wm mp4/tb保存             │ ✅ 是      │ 代码已删除                                             │
+  ├───────────────────────────────────┼────────────┼────────────────────────────────────────────────────────┤
+  │ Policy路径跳过decode              │ ❌ 否      │ 需传--fast_policy_no_decode（默认False，即默认decode） │
+  └───────────────────────────────────┴────────────┴────────────────────────────────────────────────────────┘

run_all_case.sh

@@ -0,0 +1,114 @@
+#!/bin/bash
+
+# 自动执行所有场景的所有case
+# 总共5个场景，每个场景4个case，共20个case
+# 设置环境变量（离线模式）
+export HF_HUB_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+# 颜色定义
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
+NC='\033[0m' # No Color
+
+# 定义所有场景
+SCENARIOS=(
+    "unitree_g1_pack_camera"
+    "unitree_z1_dual_arm_cleanup_pencils"
+    "unitree_z1_dual_arm_stackbox"
+    "unitree_z1_dual_arm_stackbox_v2"
+    "unitree_z1_stackbox"
+)
+
+# 定义case数量
+CASES=(1 2 3 4)
+
+# 记录开始时间
+START_TIME=$(date +%s)
+LOG_FILE="run_all_cases_$(date +%Y%m%d_%H%M%S).log"
+
+echo -e "${BLUE}========================================${NC}"
+echo -e "${BLUE}开始执行所有场景的case${NC}"
+echo -e "${BLUE}总共: ${#SCENARIOS[@]} 个场景 x ${#CASES[@]} 个case = $((${#SCENARIOS[@]} * ${#CASES[@]})) 个任务${NC}"
+echo -e "${BLUE}日志文件: ${LOG_FILE}${NC}"
+echo -e "${BLUE}========================================${NC}"
+echo ""
+
+# 初始化计数器
+TOTAL_CASES=$((${#SCENARIOS[@]} * ${#CASES[@]}))
+CURRENT_CASE=0
+SUCCESS_COUNT=0
+FAIL_COUNT=0
+
+# 记录失败的case
+declare -a FAILED_CASES
+
+# 遍历所有场景
+for scenario in "${SCENARIOS[@]}"; do
+    echo -e "${YELLOW}>>> 场景: ${scenario}${NC}"
+
+    # 遍历所有case
+    for case_num in "${CASES[@]}"; do
+        CURRENT_CASE=$((CURRENT_CASE + 1))
+        case_dir="${scenario}/case${case_num}"
+        script_path="${case_dir}/run_world_model_interaction.sh"
+
+        echo -e "${BLUE}[${CURRENT_CASE}/${TOTAL_CASES}] 执行: ${case_dir}${NC}"
+
+        # 检查脚本是否存在
+        if [ ! -f "${script_path}" ]; then
+            echo -e "${RED}错误: 脚本不存在 ${script_path}${NC}"
+            FAIL_COUNT=$((FAIL_COUNT + 1))
+            FAILED_CASES+=("${case_dir} (脚本不存在)")
+            continue
+        fi
+
+        # 执行脚本
+        echo "开始时间: $(date '+%Y-%m-%d %H:%M:%S')"
+
+        if bash "${script_path}" >> "${LOG_FILE}" 2>&1; then
+            echo -e "${GREEN}✓ 成功: ${case_dir}${NC}"
+            SUCCESS_COUNT=$((SUCCESS_COUNT + 1))
+        else
+            echo -e "${RED}✗ 失败: ${case_dir}${NC}"
+            FAIL_COUNT=$((FAIL_COUNT + 1))
+            FAILED_CASES+=("${case_dir}")
+        fi
+
+        echo "结束时间: $(date '+%Y-%m-%d %H:%M:%S')"
+        echo ""
+    done
+
+    echo ""
+done
+
+# 计算总耗时
+END_TIME=$(date +%s)
+DURATION=$((END_TIME - START_TIME))
+HOURS=$((DURATION / 3600))
+MINUTES=$(((DURATION % 3600) / 60))
+SECONDS=$((DURATION % 60))
+
+# 输出总结
+echo -e "${BLUE}========================================${NC}"
+echo -e "${BLUE}执行完成！${NC}"
+echo -e "${BLUE}========================================${NC}"
+echo -e "总任务数: ${TOTAL_CASES}"
+echo -e "${GREEN}成功: ${SUCCESS_COUNT}${NC}"
+echo -e "${RED}失败: ${FAIL_COUNT}${NC}"
+echo -e "总耗时: ${HOURS}小时 ${MINUTES}分钟 ${SECONDS}秒"
+echo -e "详细日志: ${LOG_FILE}"
+echo ""
+
+# 如果有失败的case，列出来
+if [ ${FAIL_COUNT} -gt 0 ]; then
+    echo -e "${RED}失败的case列表:${NC}"
+    for failed_case in "${FAILED_CASES[@]}"; do
+        echo -e "${RED}  - ${failed_case}${NC}"
+    done
+    echo ""
+fi
+
+echo -e "${BLUE}========================================${NC}"

run_all_cases_20260218_190150.log

@@ -0,0 +1,504 @@
+2026-02-18 19:01:56.891895: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-02-18 19:01:56.940243: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+2026-02-18 19:01:56.940285: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+2026-02-18 19:01:56.941395: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+2026-02-18 19:01:56.948327: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-02-18 19:01:57.870809: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+Global seed set to 123
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+INFO:root:***** Configing Data *****
+>>> unitree_z1_stackbox: 1 data samples loaded.
+>>> unitree_z1_stackbox: data stats loaded.
+>>> unitree_z1_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox_v2: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: normalizer initiated.
+>>> unitree_z1_dual_arm_cleanup_pencils: 1 data samples loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: data stats loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: normalizer initiated.
+>>> unitree_g1_pack_camera: 1 data samples loaded.
+>>> unitree_g1_pack_camera: data stats loaded.
+>>> unitree_g1_pack_camera: normalizer initiated.
+>>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
+    TRT output 'y': [1, 4, 16, 40, 64] torch.float32
+    TRT output 'hs_a_0': [1, 16, 320, 40, 64] torch.float32
+    TRT output 'hs_a_1': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_2': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_3': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_4': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_5': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_6': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_7': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_8': [1, 16, 320, 40, 64] torch.float32
+>>> TRT backbone loaded from /home/qhy/unifolm-world-model-action/scripts/evaluation/../../trt_engines/video_backbone.engine
+>>> Generate 16 frames under each generation ...
+DEBUG:h5py._conv:Creating converter from 3 to 5
+DEBUG:PIL.PngImagePlugin:STREAM b'IHDR' 16 13
+DEBUG:PIL.PngImagePlugin:STREAM b'pHYs' 41 9
+DEBUG:PIL.PngImagePlugin:STREAM b'IDAT' 62 4096
+
+  0%|          | 0/11 [00:00<?, ?it/s][02/18/2026-19:02:10] [TRT] [W] Using default stream in enqueueV3() may lead to performance issues due to additional calls to cudaStreamSynchronize() by TensorRT to ensure correct synchronization. Please use non-default stream instead.
+
+  9%|▉         | 1/11 [00:17<02:51, 17.15s/it]>>> Step 0: generating actions ...
+>>> Step 0: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 1: generating actions ...
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Image: failed to import FpxImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Image: failed to import MicImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XVThumbImagePlugin
+DEBUG:PIL.Image:Importing XVThumbImagePlugin
+
+ 18%|█▊        | 2/11 [00:33<02:31, 16.87s/it]
+ 27%|██▋       | 3/11 [00:50<02:14, 16.76s/it]
+ 36%|███▋      | 4/11 [01:07<01:57, 16.81s/it]
+ 45%|████▌     | 5/11 [01:24<01:41, 16.85s/it]
+ 55%|█████▍    | 6/11 [01:41<01:24, 16.82s/it]
+ 64%|██████▎   | 7/11 [01:57<01:07, 16.82s/it]
+ 73%|███████▎  | 8/11 [02:14<00:50, 16.83s/it]
+ 82%|████████▏ | 9/11 [02:31<00:33, 16.80s/it]
+ 91%|█████████ | 10/11 [02:48<00:16, 16.81s/it]
+100%|██████████| 11/11 [03:05<00:00, 16.81s/it]
+100%|██████████| 11/11 [03:05<00:00, 16.83s/it]
+>>> Step 1: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 2: generating actions ...
+>>> Step 2: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 3: generating actions ...
+>>> Step 3: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 4: generating actions ...
+>>> Step 4: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...
+>>> Step 5: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 6: generating actions ...
+>>> Step 6: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 8: generating actions ...
+>>> Step 8: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 9: generating actions ...
+>>> Step 9: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 10: generating actions ...
+>>> Step 10: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+
+real	3m49.072s
+user	4m16.055s
+sys	0m44.636s
+2026-02-18 19:05:45.956647: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-02-18 19:05:46.004149: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+2026-02-18 19:05:46.004193: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+2026-02-18 19:05:46.005265: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+2026-02-18 19:05:46.012074: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-02-18 19:05:46.932966: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+Global seed set to 123
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+INFO:root:***** Configing Data *****
+>>> unitree_z1_stackbox: 1 data samples loaded.
+>>> unitree_z1_stackbox: data stats loaded.
+>>> unitree_z1_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox_v2: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: normalizer initiated.
+>>> unitree_z1_dual_arm_cleanup_pencils: 1 data samples loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: data stats loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: normalizer initiated.
+>>> unitree_g1_pack_camera: 1 data samples loaded.
+>>> unitree_g1_pack_camera: data stats loaded.
+>>> unitree_g1_pack_camera: normalizer initiated.
+>>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
+    TRT output 'y': [1, 4, 16, 40, 64] torch.float32
+    TRT output 'hs_a_0': [1, 16, 320, 40, 64] torch.float32
+    TRT output 'hs_a_1': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_2': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_3': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_4': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_5': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_6': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_7': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_8': [1, 16, 320, 40, 64] torch.float32
+>>> TRT backbone loaded from /home/qhy/unifolm-world-model-action/scripts/evaluation/../../trt_engines/video_backbone.engine
+>>> Generate 16 frames under each generation ...
+DEBUG:h5py._conv:Creating converter from 3 to 5
+DEBUG:PIL.PngImagePlugin:STREAM b'IHDR' 16 13
+DEBUG:PIL.PngImagePlugin:STREAM b'pHYs' 41 9
+DEBUG:PIL.PngImagePlugin:STREAM b'IDAT' 62 4096
+
+  0%|          | 0/11 [00:00<?, ?it/s][02/18/2026-19:05:59] [TRT] [W] Using default stream in enqueueV3() may lead to performance issues due to additional calls to cudaStreamSynchronize() by TensorRT to ensure correct synchronization. Please use non-default stream instead.
+
+  9%|▉         | 1/11 [00:16<02:47, 16.71s/it]>>> Step 0: generating actions ...
+>>> Step 0: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 1: generating actions ...
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Image: failed to import FpxImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Image: failed to import MicImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XVThumbImagePlugin
+DEBUG:PIL.Image:Importing XVThumbImagePlugin
+
+ 18%|█▊        | 2/11 [00:33<02:30, 16.75s/it]
+ 27%|██▋       | 3/11 [00:50<02:15, 16.91s/it]
+ 36%|███▋      | 4/11 [01:07<01:59, 17.02s/it]
+ 45%|████▌     | 5/11 [01:24<01:41, 16.98s/it]
+ 55%|█████▍    | 6/11 [01:41<01:24, 16.94s/it]
+ 64%|██████▎   | 7/11 [01:58<01:07, 16.90s/it]
+ 73%|███████▎  | 8/11 [02:15<00:50, 16.83s/it]
+ 82%|████████▏ | 9/11 [02:31<00:33, 16.80s/it]
+ 91%|█████████ | 10/11 [02:49<00:16, 16.94s/it]
+100%|██████████| 11/11 [03:06<00:00, 16.97s/it]
+100%|██████████| 11/11 [03:06<00:00, 16.91s/it]
+>>> Step 1: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 2: generating actions ...
+>>> Step 2: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 3: generating actions ...
+>>> Step 3: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 4: generating actions ...
+>>> Step 4: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...
+>>> Step 5: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 6: generating actions ...
+>>> Step 6: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 8: generating actions ...
+>>> Step 8: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 9: generating actions ...
+>>> Step 9: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 10: generating actions ...
+>>> Step 10: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+
+real	3m49.162s
+user	4m12.814s
+sys	0m45.565s
+2026-02-18 19:09:35.113634: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-02-18 19:09:35.161428: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+2026-02-18 19:09:35.161474: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+2026-02-18 19:09:35.162551: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+2026-02-18 19:09:35.169325: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-02-18 19:09:36.089250: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+Global seed set to 123
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+INFO:root:***** Configing Data *****
+>>> unitree_z1_stackbox: 1 data samples loaded.
+>>> unitree_z1_stackbox: data stats loaded.
+>>> unitree_z1_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox_v2: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: normalizer initiated.
+>>> unitree_z1_dual_arm_cleanup_pencils: 1 data samples loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: data stats loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: normalizer initiated.
+>>> unitree_g1_pack_camera: 1 data samples loaded.
+>>> unitree_g1_pack_camera: data stats loaded.
+>>> unitree_g1_pack_camera: normalizer initiated.
+>>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
+    TRT output 'y': [1, 4, 16, 40, 64] torch.float32
+    TRT output 'hs_a_0': [1, 16, 320, 40, 64] torch.float32
+    TRT output 'hs_a_1': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_2': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_3': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_4': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_5': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_6': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_7': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_8': [1, 16, 320, 40, 64] torch.float32
+>>> TRT backbone loaded from /home/qhy/unifolm-world-model-action/scripts/evaluation/../../trt_engines/video_backbone.engine
+>>> Generate 16 frames under each generation ...
+DEBUG:h5py._conv:Creating converter from 3 to 5
+DEBUG:PIL.PngImagePlugin:STREAM b'IHDR' 16 13
+DEBUG:PIL.PngImagePlugin:STREAM b'pHYs' 41 9
+DEBUG:PIL.PngImagePlugin:STREAM b'IDAT' 62 4096
+
+  0%|          | 0/11 [00:00<?, ?it/s][02/18/2026-19:09:49] [TRT] [W] Using default stream in enqueueV3() may lead to performance issues due to additional calls to cudaStreamSynchronize() by TensorRT to ensure correct synchronization. Please use non-default stream instead.
+
+  9%|▉         | 1/11 [00:16<02:45, 16.53s/it]>>> Step 0: generating actions ...
+>>> Step 0: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 1: generating actions ...
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Image: failed to import FpxImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Image: failed to import MicImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin

run_all_psnr.sh

@@ -0,0 +1,61 @@
+#!/bin/bash
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+cd "$SCRIPT_DIR"
+
+SCENARIOS=(
+    unitree_g1_pack_camera
+    unitree_z1_dual_arm_cleanup_pencils
+    unitree_z1_dual_arm_stackbox
+    unitree_z1_dual_arm_stackbox_v2
+    unitree_z1_stackbox
+)
+
+CASES=(case1 case2 case3 case4)
+
+total=0
+success=0
+fail=0
+
+for scenario in "${SCENARIOS[@]}"; do
+    for case in "${CASES[@]}"; do
+        case_dir="${scenario}/${case}"
+        gt_video="${case_dir}/${scenario}_${case}.mp4"
+        pred_video=$(ls "${case_dir}"/output/inference/*_full_fs*.mp4 2>/dev/null | head -1)
+        output_file="${case_dir}/psnr_result.json"
+
+        total=$((total + 1))
+        echo "=========================================="
+        echo "[${total}/20] ${case_dir}"
+
+        if [ ! -f "$gt_video" ]; then
+            echo "  SKIP: GT video not found: $gt_video"
+            fail=$((fail + 1))
+            continue
+        fi
+        if [ -z "$pred_video" ]; then
+            echo "  SKIP: pred video not found in ${case_dir}/output/inference/"
+            fail=$((fail + 1))
+            continue
+        fi
+
+        echo "  GT:   $gt_video"
+        echo "  Pred: $pred_video"
+        echo "  Out:  $output_file"
+
+        if python3 psnr_score_for_challenge.py \
+            --gt_video "$gt_video" \
+            --pred_video "$pred_video" \
+            --output_file "$output_file"; then
+            success=$((success + 1))
+            echo "  DONE"
+        else
+            fail=$((fail + 1))
+            echo "  FAILED"
+        fi
+    done
+done
+
+echo "=========================================="
+echo "Finished: ${success} success, ${fail} fail, ${total} total"

scripts/evaluation/base_model_inference.py

@@ -16,6 +16,9 @@ from collections import OrderedDict
 from unifolm_wma.models.samplers.ddim import DDIMSampler
 from unifolm_wma.utils.utils import instantiate_from_config
 
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+
 
 def get_filelist(data_dir: str, postfixes: list[str]) -> list[str]:
     """

scripts/evaluation/real_eval_server.py

@@ -19,6 +19,9 @@ from fastapi.responses import JSONResponse
 from typing import Any, Dict, Optional, Tuple, List
 from datetime import datetime
 
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+
 from unifolm_wma.utils.utils import instantiate_from_config
 from unifolm_wma.models.samplers.ddim import DDIMSampler
 

scripts/evaluation/world_model_interaction.py

@@ -9,6 +9,8 @@ import logging
 import einops
 import warnings
 import imageio
+import atexit
+from concurrent.futures import ThreadPoolExecutor
 
 from pytorch_lightning import seed_everything
 from omegaconf import OmegaConf
@@ -16,8 +18,12 @@ from tqdm import tqdm
 from einops import rearrange, repeat
 from collections import OrderedDict
 from torch import nn
-from eval_utils import populate_queues, log_to_tensorboard
+from eval_utils import populate_queues
 from collections import deque
+from typing import Optional, List, Any
+
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
 from torch import Tensor
 from torch.utils.tensorboard import SummaryWriter
 from PIL import Image
@@ -150,6 +156,81 @@ def save_results(video: Tensor, filename: str, fps: int = 8) -> None:
                                options={'crf': '10'})
 
 
+# ========== Async I/O ==========
+_io_executor: Optional[ThreadPoolExecutor] = None
+_io_futures: List[Any] = []
+
+
+def _get_io_executor() -> ThreadPoolExecutor:
+    global _io_executor
+    if _io_executor is None:
+        _io_executor = ThreadPoolExecutor(max_workers=2)
+    return _io_executor
+
+
+def _flush_io():
+    """Wait for all pending async I/O to finish."""
+    global _io_futures
+    for fut in _io_futures:
+        try:
+            fut.result()
+        except Exception as e:
+            print(f">>> [async I/O] error: {e}")
+    _io_futures.clear()
+
+
+atexit.register(_flush_io)
+
+
+def _save_results_sync(video_cpu: Tensor, filename: str, fps: int) -> None:
+    """Synchronous save on CPU tensor (runs in background thread)."""
+    video = torch.clamp(video_cpu.float(), -1., 1.)
+    n = video.shape[0]
+    video = video.permute(2, 0, 1, 3, 4)
+    frame_grids = [
+        torchvision.utils.make_grid(framesheet, nrow=int(n), padding=0)
+        for framesheet in video
+    ]
+    grid = torch.stack(frame_grids, dim=0)
+    grid = (grid + 1.0) / 2.0
+    grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
+    torchvision.io.write_video(filename,
+                               grid,
+                               fps=fps,
+                               video_codec='h264',
+                               options={'crf': '10'})
+
+
+def save_results_async(video: Tensor, filename: str, fps: int = 8) -> None:
+    """Submit video saving to background thread pool."""
+    video_cpu = video.detach().cpu()
+    fut = _get_io_executor().submit(_save_results_sync, video_cpu, filename, fps)
+    _io_futures.append(fut)
+
+
+def _log_to_tb_sync(writer, video_cpu: Tensor, tag: str, fps: int) -> None:
+    """Synchronous TensorBoard log on CPU tensor (runs in background thread)."""
+    if video_cpu.dim() == 5:
+        n = video_cpu.shape[0]
+        video = video_cpu.permute(2, 0, 1, 3, 4)
+        frame_grids = [
+            torchvision.utils.make_grid(framesheet, nrow=int(n), padding=0)
+            for framesheet in video
+        ]
+        grid = torch.stack(frame_grids, dim=0)
+        grid = (grid + 1.0) / 2.0
+        grid = grid.unsqueeze(dim=0)
+        writer.add_video(tag, grid, fps=fps)
+
+
+def log_to_tensorboard_async(writer, data: Tensor, tag: str, fps: int = 10) -> None:
+    """Submit TensorBoard logging to background thread pool."""
+    if isinstance(data, torch.Tensor) and data.dim() == 5:
+        data_cpu = data.detach().cpu()
+        fut = _get_io_executor().submit(_log_to_tb_sync, writer, data_cpu, tag, fps)
+        _io_futures.append(fut)
+
+
 def get_init_frame_path(data_dir: str, sample: dict) -> str:
     """Construct the init_frame path from directory and sample metadata.
 
@@ -327,7 +408,8 @@ def image_guided_synthesis_sim_mode(
         timestep_spacing: str = 'uniform',
         guidance_rescale: float = 0.0,
         sim_mode: bool = True,
-        **kwargs) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        decode_video: bool = True,
+        **kwargs) -> tuple[torch.Tensor | None, torch.Tensor, torch.Tensor]:
     """
     Performs image-guided video generation in a simulation-style mode with optional multimodal guidance (image, state, action, text).
 
@@ -350,10 +432,13 @@ def image_guided_synthesis_sim_mode(
         timestep_spacing (str): Timestep sampling method in DDIM sampler. Typically "uniform" or "linspace".
         guidance_rescale (float): Guidance rescaling factor to mitigate overexposure from classifier-free guidance.
         sim_mode (bool): Whether to perform world-model interaction or decision-making using the world-model.
+        decode_video (bool): Whether to decode latent samples to pixel-space video.
+            Set to False to skip VAE decode for speed when only actions/states are needed.
         **kwargs: Additional arguments passed to the DDIM sampler.
 
     Returns:
-        batch_variants (torch.Tensor): Predicted pixel-space video frames [B, C, T, H, W].
+        batch_variants (torch.Tensor | None): Predicted pixel-space video frames [B, C, T, H, W],
+            or None when decode_video=False.
         actions (torch.Tensor): Predicted action sequences [B, T, D] from diffusion decoding.
         states (torch.Tensor): Predicted state sequences [B, T, D] from diffusion decoding.
     """
@@ -406,6 +491,7 @@ def image_guided_synthesis_sim_mode(
     kwargs.update({"unconditional_conditioning_img_nonetext": None})
     cond_mask = None
     cond_z0 = None
+    batch_variants = None
     if ddim_sampler is not None:
         samples, actions, states, intermedia = ddim_sampler.sample(
             S=ddim_steps,
@@ -424,9 +510,10 @@ def image_guided_synthesis_sim_mode(
             guidance_rescale=guidance_rescale,
             **kwargs)
 
-        # Reconstruct from latent to pixel space
+        if decode_video:
-        batch_images = model.decode_first_stage(samples)
+            # Reconstruct from latent to pixel space
-        batch_variants = batch_images
+            batch_images = model.decode_first_stage(samples)
+            batch_variants = batch_images
 
     return batch_variants, actions, states
 
@@ -453,26 +540,56 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
     csv_path = os.path.join(args.prompt_dir, f"{args.dataset}.csv")
     df = pd.read_csv(csv_path)
 
-    # Load config
+    # Load config (always needed for data setup)
     config = OmegaConf.load(args.config)
-    config['model']['params']['wma_config']['params'][
-        'use_checkpoint'] = False
-    model = instantiate_from_config(config.model)
-    model.perframe_ae = args.perframe_ae
-    assert os.path.exists(args.ckpt_path), "Error: checkpoint Not Found!"
-    model = load_model_checkpoint(model, args.ckpt_path)
-    model.eval()
-    print(f'>>> Load pre-trained model ...')
 
-    # Build unnomalizer
+    prepared_path = args.ckpt_path + ".prepared.pt"
+    if os.path.exists(prepared_path):
+        # ---- Fast path: load the fully-prepared model ----
+        print(f">>> Loading prepared model from {prepared_path} ...")
+        model = torch.load(prepared_path,
+                           map_location=f"cuda:{gpu_no}",
+                           weights_only=False,
+                           mmap=True)
+        model.eval()
+        print(f">>> Prepared model loaded.")
+    else:
+        # ---- Normal path: construct + load checkpoint ----
+        config['model']['params']['wma_config']['params'][
+            'use_checkpoint'] = False
+        model = instantiate_from_config(config.model)
+        model.perframe_ae = args.perframe_ae
+
+        assert os.path.exists(args.ckpt_path), "Error: checkpoint Not Found!"
+        model = load_model_checkpoint(model, args.ckpt_path)
+        model.eval()
+        model = model.cuda(gpu_no)
+        print(f'>>> Load pre-trained model ...')
+
+        # Save prepared model for fast loading next time
+        print(f">>> Saving prepared model to {prepared_path} ...")
+        torch.save(model, prepared_path)
+        print(f">>> Prepared model saved ({os.path.getsize(prepared_path) / 1024**3:.1f} GB).")
+
+    # Build normalizer (always needed, independent of model loading path)
     logging.info("***** Configing Data *****")
     data = instantiate_from_config(config.data)
     data.setup()
     print(">>> Dataset is successfully loaded ...")
 
-    model = model.cuda(gpu_no)
     device = get_device_from_parameters(model)
 
+    # Fuse KV projections in attention layers (to_k + to_v → to_kv)
+    from unifolm_wma.modules.attention import CrossAttention
+    kv_count = sum(1 for m in model.modules()
+                   if isinstance(m, CrossAttention) and m.fuse_kv())
+    print(f"    ✓ KV fused: {kv_count} attention layers")
+
+    # Load TRT backbone if engine exists
+    trt_engine_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), '..', '..', 'trt_engines', 'video_backbone.engine')
+    if os.path.exists(trt_engine_path):
+        model.model.diffusion_model.load_trt_backbone(trt_engine_path)
+
     # Run over data
     assert (args.height % 16 == 0) and (
         args.width % 16
@@ -587,7 +704,8 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
                     fs=model_input_fs,
                     timestep_spacing=args.timestep_spacing,
                     guidance_rescale=args.guidance_rescale,
-                    sim_mode=False)
+                    sim_mode=False,
+                    decode_video=not args.fast_policy_no_decode)
 
                 # Update future actions in the observation queues
                 for idx in range(len(pred_actions[0])):
@@ -644,29 +762,31 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
                     cond_obs_queues = populate_queues(cond_obs_queues,
                                                       observation)
 
-                # Save the imagen videos for decision-making
+                # Save the imagen videos for decision-making (async)
-                sample_tag = f"{args.dataset}-vid{sample['videoid']}-dm-fs-{fs}/itr-{itr}"
+                if pred_videos_0 is not None:
-                log_to_tensorboard(writer,
+                    sample_tag = f"{args.dataset}-vid{sample['videoid']}-dm-fs-{fs}/itr-{itr}"
-                                   pred_videos_0,
+                    log_to_tensorboard_async(writer,
-                                   sample_tag,
+                                             pred_videos_0,
-                                   fps=args.save_fps)
+                                             sample_tag,
+                                             fps=args.save_fps)
                 # Save videos environment changes via world-model interaction
                 sample_tag = f"{args.dataset}-vid{sample['videoid']}-wd-fs-{fs}/itr-{itr}"
-                log_to_tensorboard(writer,
+                log_to_tensorboard_async(writer,
-                                   pred_videos_1,
+                                         pred_videos_1,
-                                   sample_tag,
+                                         sample_tag,
-                                   fps=args.save_fps)
+                                         fps=args.save_fps)
 
                 # Save the imagen videos for decision-making
-                sample_video_file = f'{video_save_dir}/dm/{fs}/itr-{itr}.mp4'
+                if pred_videos_0 is not None:
-                save_results(pred_videos_0.cpu(),
+                    sample_video_file = f'{video_save_dir}/dm/{fs}/itr-{itr}.mp4'
-                             sample_video_file,
+                    save_results_async(pred_videos_0,
-                             fps=args.save_fps)
+                                       sample_video_file,
+                                       fps=args.save_fps)
                 # Save videos environment changes via world-model interaction
                 sample_video_file = f'{video_save_dir}/wm/{fs}/itr-{itr}.mp4'
-                save_results(pred_videos_1.cpu(),
+                save_results_async(pred_videos_1,
-                             sample_video_file,
+                                   sample_video_file,
-                             fps=args.save_fps)
+                                   fps=args.save_fps)
 
                 print('>' * 24)
                 # Collect the result of world-model interactions
@@ -674,12 +794,15 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
 
             full_video = torch.cat(wm_video, dim=2)
             sample_tag = f"{args.dataset}-vid{sample['videoid']}-wd-fs-{fs}/full"
-            log_to_tensorboard(writer,
+            log_to_tensorboard_async(writer,
-                               full_video,
+                                     full_video,
-                               sample_tag,
+                                     sample_tag,
-                               fps=args.save_fps)
+                                     fps=args.save_fps)
             sample_full_video_file = f"{video_save_dir}/../{sample['videoid']}_full_fs{fs}.mp4"
-            save_results(full_video, sample_full_video_file, fps=args.save_fps)
+            save_results_async(full_video, sample_full_video_file, fps=args.save_fps)
+
+    # Wait for all async I/O to complete
+    _flush_io()
 
 
 def get_parser():
@@ -794,6 +917,11 @@ def get_parser():
                         action='store_true',
                         default=False,
                         help="not using the predicted states as comparison")
+    parser.add_argument(
+        "--fast_policy_no_decode",
+        action='store_true',
+        default=False,
+        help="Speed mode: policy pass only predicts actions, skip policy video decode/log/save.")
     parser.add_argument("--save_fps",
                         type=int,
                         default=8,

scripts/export_trt.py

@@ -0,0 +1,87 @@
+"""Export video UNet backbone to ONNX, then convert to TensorRT engine.
+
+Usage:
+    python scripts/export_trt.py \
+        --ckpt ckpts/unifolm_wma_dual.ckpt.prepared.pt \
+        --config configs/inference/world_model_interaction.yaml \
+        --out_dir trt_engines
+"""
+
+import os
+import sys
+import argparse
+
+import torch
+import tensorrt as trt
+from omegaconf import OmegaConf
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'src'))
+from unifolm_wma.utils.utils import instantiate_from_config
+from unifolm_wma.trt_utils import export_backbone_onnx
+
+
+def load_model(config_path, ckpt_path):
+    if ckpt_path.endswith('.prepared.pt'):
+        model = torch.load(ckpt_path, map_location='cpu')
+    else:
+        config = OmegaConf.load(config_path)
+        model = instantiate_from_config(config.model)
+        state_dict = torch.load(ckpt_path, map_location='cpu')
+        if 'state_dict' in state_dict:
+            state_dict = state_dict['state_dict']
+        model.load_state_dict(state_dict, strict=False)
+    model.eval().cuda()
+    return model
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--ckpt', required=True)
+    parser.add_argument('--config', default='configs/inference/world_model_interaction.yaml')
+    parser.add_argument('--out_dir', default='trt_engines')
+    parser.add_argument('--context_len', type=int, default=95)
+    parser.add_argument('--fp16', action='store_true', default=True)
+    args = parser.parse_args()
+
+    os.makedirs(args.out_dir, exist_ok=True)
+    onnx_path = os.path.join(args.out_dir, 'video_backbone.onnx')
+    engine_path = os.path.join(args.out_dir, 'video_backbone.engine')
+
+    if os.path.exists(onnx_path):
+        print(f">>> ONNX already exists at {onnx_path}, skipping export.")
+        n_outputs = 10
+    else:
+        print(">>> Loading model ...")
+        model = load_model(args.config, args.ckpt)
+        print(">>> Exporting ONNX ...")
+        with torch.no_grad():
+            n_outputs = export_backbone_onnx(model, onnx_path, context_len=args.context_len)
+        del model
+        torch.cuda.empty_cache()
+
+    print(">>> Converting ONNX -> TensorRT engine ...")
+    logger = trt.Logger(trt.Logger.WARNING)
+    builder = trt.Builder(logger)
+    network = builder.create_network(1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
+    parser = trt.OnnxParser(network, logger)
+
+    if not parser.parse_from_file(os.path.abspath(onnx_path)):
+        for i in range(parser.num_errors):
+            print(f"    ONNX parse error: {parser.get_error(i)}")
+        raise RuntimeError("ONNX parsing failed")
+
+    config = builder.create_builder_config()
+    config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, 16 << 30)
+    if args.fp16:
+        config.set_flag(trt.BuilderFlag.FP16)
+
+    engine_bytes = builder.build_serialized_network(network, config)
+    with open(engine_path, 'wb') as f:
+        f.write(engine_bytes)
+
+    print(f"\n>>> Done! Engine saved to {engine_path}")
+    print(f"    Outputs: 1 y + {n_outputs - 1} hs_a tensors")
+
+
+if __name__ == '__main__':
+    main()

scripts/trainer.py

@@ -11,6 +11,9 @@ from unifolm_wma.utils.utils import instantiate_from_config
 from unifolm_wma.utils.train import get_trainer_callbacks, get_trainer_logger, get_trainer_strategy
 from unifolm_wma.utils.train import set_logger, init_workspace, load_checkpoints, get_num_parameters
 
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+
 
 def get_parser(**parser_kwargs):
     parser = argparse.ArgumentParser(**parser_kwargs)

src/unifolm_wma/models/diffusion_head/conditional_unet1d.py

@@ -501,6 +501,10 @@ class ConditionalUnet1D(nn.Module):
         self.last_frame_only = last_frame_only
         self.horizon = horizon
 
+        # Context precomputation cache
+        self._global_cond_cache_enabled = False
+        self._global_cond_cache = {}
+
     def forward(self,
                 sample: torch.Tensor,
                 timestep: Union[torch.Tensor, float, int],
@@ -530,14 +534,20 @@ class ConditionalUnet1D(nn.Module):
         B, T, D = sample.shape
         if self.use_linear_act_proj:
             sample = self.proj_in_action(sample.unsqueeze(-1))
-            global_cond = self.obs_encoder(cond)
+            _gc_key = (cond['image'].data_ptr(), cond['agent_pos'].data_ptr())
-            global_cond = rearrange(global_cond,
+            if self._global_cond_cache_enabled and _gc_key in self._global_cond_cache:
-                                    '(b t) d -> b 1 (t d)',
+                global_cond = self._global_cond_cache[_gc_key]
-                                    b=B,
+            else:
-                                    t=self.n_obs_steps)
+                global_cond = self.obs_encoder(cond)
-            global_cond = repeat(global_cond,
+                global_cond = rearrange(global_cond,
-                                 'b c d -> b (repeat c) d',
+                                        '(b t) d -> b 1 (t d)',
-                                 repeat=T)
+                                        b=B,
+                                        t=self.n_obs_steps)
+                global_cond = repeat(global_cond,
+                                     'b c d -> b (repeat c) d',
+                                     repeat=T)
+                if self._global_cond_cache_enabled:
+                    self._global_cond_cache[_gc_key] = global_cond
         else:
             sample = einops.rearrange(sample, 'b h t -> b t h')
             sample = self.proj_in_horizon(sample)

src/unifolm_wma/models/samplers/ddim.py

@@ -6,6 +6,8 @@ from unifolm_wma.utils.diffusion import make_ddim_sampling_parameters, make_ddim
 from unifolm_wma.utils.common import noise_like
 from unifolm_wma.utils.common import extract_into_tensor
 from tqdm import tqdm
+from unifolm_wma.modules.attention import enable_cross_attn_kv_cache, disable_cross_attn_kv_cache
+from unifolm_wma.modules.networks.wma_model import enable_ctx_cache, disable_ctx_cache
 
 
 class DDIMSampler(object):
@@ -67,11 +69,12 @@ class DDIMSampler(object):
             ddim_timesteps=self.ddim_timesteps,
             eta=ddim_eta,
             verbose=verbose)
-        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        # Ensure tensors are on correct device for efficient indexing
-        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_sigmas', to_torch(torch.as_tensor(ddim_sigmas)))
-        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_alphas', to_torch(torch.as_tensor(ddim_alphas)))
+        self.register_buffer('ddim_alphas_prev', to_torch(torch.as_tensor(ddim_alphas_prev)))
         self.register_buffer('ddim_sqrt_one_minus_alphas',
-                             np.sqrt(1. - ddim_alphas))
+                             to_torch(torch.as_tensor(np.sqrt(1. - ddim_alphas))))
         sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
             (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) *
             (1 - self.alphas_cumprod / self.alphas_cumprod_prev))
@@ -241,63 +244,70 @@ class DDIMSampler(object):
 
         dp_ddim_scheduler_action.set_timesteps(len(timesteps))
         dp_ddim_scheduler_state.set_timesteps(len(timesteps))
-        for i, step in enumerate(iterator):
+        ts = torch.empty((b, ), device=device, dtype=torch.long)
-            index = total_steps - i - 1
+        enable_cross_attn_kv_cache(self.model)
-            ts = torch.full((b, ), step, device=device, dtype=torch.long)
+        enable_ctx_cache(self.model)
+        try:
+            for i, step in enumerate(iterator):
+                index = total_steps - i - 1
+                ts.fill_(step)
 
-            # Use mask to blend noised original latent (img_orig) & new sampled latent (img)
+                # Use mask to blend noised original latent (img_orig) & new sampled latent (img)
-            if mask is not None:
+                if mask is not None:
-                assert x0 is not None
+                    assert x0 is not None
-                if clean_cond:
+                    if clean_cond:
-                    img_orig = x0
+                        img_orig = x0
-                else:
+                    else:
-                    img_orig = self.model.q_sample(x0, ts)
+                        img_orig = self.model.q_sample(x0, ts)
-                img = img_orig * mask + (1. - mask) * img
+                    img = img_orig * mask + (1. - mask) * img
 
-            outs = self.p_sample_ddim(
+                outs = self.p_sample_ddim(
-                img,
+                    img,
-                action,
+                    action,
-                state,
+                    state,
-                cond,
+                    cond,
-                ts,
+                    ts,
-                index=index,
+                    index=index,
-                use_original_steps=ddim_use_original_steps,
+                    use_original_steps=ddim_use_original_steps,
-                quantize_denoised=quantize_denoised,
+                    quantize_denoised=quantize_denoised,
-                temperature=temperature,
+                    temperature=temperature,
-                noise_dropout=noise_dropout,
+                    noise_dropout=noise_dropout,
-                score_corrector=score_corrector,
+                    score_corrector=score_corrector,
-                corrector_kwargs=corrector_kwargs,
+                    corrector_kwargs=corrector_kwargs,
-                unconditional_guidance_scale=unconditional_guidance_scale,
+                    unconditional_guidance_scale=unconditional_guidance_scale,
-                unconditional_conditioning=unconditional_conditioning,
+                    unconditional_conditioning=unconditional_conditioning,
-                mask=mask,
+                    mask=mask,
-                x0=x0,
+                    x0=x0,
-                fs=fs,
+                    fs=fs,
-                guidance_rescale=guidance_rescale,
+                    guidance_rescale=guidance_rescale,
-                **kwargs)
+                    **kwargs)
 
-            img, pred_x0, model_output_action, model_output_state = outs
+                img, pred_x0, model_output_action, model_output_state = outs
 
-            action = dp_ddim_scheduler_action.step(
+                action = dp_ddim_scheduler_action.step(
-                model_output_action,
+                    model_output_action,
-                step,
+                    step,
-                action,
+                    action,
-                generator=None,
+                    generator=None,
-            ).prev_sample
+                ).prev_sample
-            state = dp_ddim_scheduler_state.step(
+                state = dp_ddim_scheduler_state.step(
-                model_output_state,
+                    model_output_state,
-                step,
+                    step,
-                state,
+                    state,
-                generator=None,
+                    generator=None,
-            ).prev_sample
+                ).prev_sample
 
-            if callback: callback(i)
+                if callback: callback(i)
-            if img_callback: img_callback(pred_x0, i)
+                if img_callback: img_callback(pred_x0, i)
 
-            if index % log_every_t == 0 or index == total_steps - 1:
+                if index % log_every_t == 0 or index == total_steps - 1:
-                intermediates['x_inter'].append(img)
+                    intermediates['x_inter'].append(img)
-                intermediates['pred_x0'].append(pred_x0)
+                    intermediates['pred_x0'].append(pred_x0)
-                intermediates['x_inter_action'].append(action)
+                    intermediates['x_inter_action'].append(action)
-                intermediates['x_inter_state'].append(state)
+                    intermediates['x_inter_state'].append(state)
+        finally:
+            disable_cross_attn_kv_cache(self.model)
+            disable_ctx_cache(self.model)
 
         return img, action, state, intermediates
 
@@ -325,10 +335,6 @@ class DDIMSampler(object):
                       guidance_rescale=0.0,
                       **kwargs):
         b, *_, device = *x.shape, x.device
-        if x.dim() == 5:
-            is_video = True
-        else:
-            is_video = False
 
         if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
             model_output, model_output_action, model_output_state = self.model.apply_model(
@@ -377,17 +383,11 @@ class DDIMSampler(object):
         sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
         sigmas = self.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
 
-        if is_video:
+        # Use 0-d tensors directly (already on device); broadcasting handles shape
-            size = (b, 1, 1, 1, 1)
+        a_t = alphas[index]
-        else:
+        a_prev = alphas_prev[index]
-            size = (b, 1, 1, 1)
+        sigma_t = sigmas[index]
-
+        sqrt_one_minus_at = sqrt_one_minus_alphas[index]
-        a_t = torch.full(size, alphas[index], device=device)
-        a_prev = torch.full(size, alphas_prev[index], device=device)
-        sigma_t = torch.full(size, sigmas[index], device=device)
-        sqrt_one_minus_at = torch.full(size,
-                                       sqrt_one_minus_alphas[index],
-                                       device=device)
 
         if self.model.parameterization != "v":
             pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
@@ -395,12 +395,8 @@ class DDIMSampler(object):
             pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)
 
         if self.model.use_dynamic_rescale:
-            scale_t = torch.full(size,
+            scale_t = self.ddim_scale_arr[index]
-                                 self.ddim_scale_arr[index],
+            prev_scale_t = self.ddim_scale_arr_prev[index]
-                                 device=device)
-            prev_scale_t = torch.full(size,
-                                      self.ddim_scale_arr_prev[index],
-                                      device=device)
             rescale = (prev_scale_t / scale_t)
             pred_x0 *= rescale
 

src/unifolm_wma/modules/attention.py

@@ -98,6 +98,10 @@ class CrossAttention(nn.Module):
         self.text_context_len = text_context_len
         self.agent_state_context_len = agent_state_context_len
         self.agent_action_context_len = agent_action_context_len
+        self._kv_cache = {}
+        self._kv_cache_enabled = False
+        self._kv_fused = False
+
         self.cross_attention_scale_learnable = cross_attention_scale_learnable
         if self.image_cross_attention:
             self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
@@ -114,6 +118,27 @@ class CrossAttention(nn.Module):
                 self.register_parameter('alpha_caa',
                                         nn.Parameter(torch.tensor(0.)))
 
+    def fuse_kv(self):
+        """Fuse to_k/to_v into to_kv (2 Linear → 1). Works for all layers."""
+        k_w = self.to_k.weight  # (inner_dim, context_dim)
+        v_w = self.to_v.weight
+        self.to_kv = nn.Linear(k_w.shape[1], k_w.shape[0] * 2, bias=False)
+        self.to_kv.weight = nn.Parameter(torch.cat([k_w, v_w], dim=0))
+        del self.to_k, self.to_v
+        if self.image_cross_attention:
+            for suffix in ('_ip', '_as', '_aa'):
+                k_attr = f'to_k{suffix}'
+                v_attr = f'to_v{suffix}'
+                kw = getattr(self, k_attr).weight
+                vw = getattr(self, v_attr).weight
+                fused = nn.Linear(kw.shape[1], kw.shape[0] * 2, bias=False)
+                fused.weight = nn.Parameter(torch.cat([kw, vw], dim=0))
+                setattr(self, f'to_kv{suffix}', fused)
+                delattr(self, k_attr)
+                delattr(self, v_attr)
+        self._kv_fused = True
+        return True
+
     def forward(self, x, context=None, mask=None):
         spatial_self_attn = (context is None)
         k_ip, v_ip, out_ip = None, None, None
@@ -140,19 +165,28 @@ class CrossAttention(nn.Module):
                                     self.agent_action_context_len +
                                     self.text_context_len:, :]
 
-            k = self.to_k(context_ins)
+            if self._kv_fused:
-            v = self.to_v(context_ins)
+                k, v = self.to_kv(context_ins).chunk(2, dim=-1)
-            k_ip = self.to_k_ip(context_image)
+                k_ip, v_ip = self.to_kv_ip(context_image).chunk(2, dim=-1)
-            v_ip = self.to_v_ip(context_image)
+                k_as, v_as = self.to_kv_as(context_agent_state).chunk(2, dim=-1)
-            k_as = self.to_k_as(context_agent_state)
+                k_aa, v_aa = self.to_kv_aa(context_agent_action).chunk(2, dim=-1)
-            v_as = self.to_v_as(context_agent_state)
+            else:
-            k_aa = self.to_k_aa(context_agent_action)
+                k = self.to_k(context_ins)
-            v_aa = self.to_v_aa(context_agent_action)
+                v = self.to_v(context_ins)
+                k_ip = self.to_k_ip(context_image)
+                v_ip = self.to_v_ip(context_image)
+                k_as = self.to_k_as(context_agent_state)
+                v_as = self.to_v_as(context_agent_state)
+                k_aa = self.to_k_aa(context_agent_action)
+                v_aa = self.to_v_aa(context_agent_action)
         else:
             if not spatial_self_attn:
                 context = context[:, :self.text_context_len, :]
-            k = self.to_k(context)
+            if self._kv_fused:
-            v = self.to_v(context)
+                k, v = self.to_kv(context).chunk(2, dim=-1)
+            else:
+                k = self.to_k(context)
+                v = self.to_v(context)
 
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h),
                       (q, k, v))
@@ -236,134 +270,162 @@ class CrossAttention(nn.Module):
         k_ip, v_ip, out_ip = None, None, None
         k_as, v_as, out_as = None, None, None
         k_aa, v_aa, out_aa = None, None, None
+        attn_mask_aa = None
 
+        h = self.heads
         q = self.to_q(x)
         context = default(context, x)
 
-        if self.image_cross_attention and not spatial_self_attn:
+        b, _, _ = q.shape
+        q = q.unsqueeze(3).reshape(b, q.shape[1], h, self.dim_head).permute(0, 2, 1, 3).reshape(b * h, q.shape[1], self.dim_head).contiguous()
+
+        def _reshape_kv(t):
+            return t.unsqueeze(3).reshape(b, t.shape[1], h, self.dim_head).permute(0, 2, 1, 3).reshape(b * h, t.shape[1], self.dim_head).contiguous()
+
+        use_cache = self._kv_cache_enabled and not spatial_self_attn
+        cache_hit = use_cache and len(self._kv_cache) > 0
+
+        if cache_hit:
+            k = self._kv_cache['k']
+            v = self._kv_cache['v']
+            k_ip = self._kv_cache.get('k_ip')
+            v_ip = self._kv_cache.get('v_ip')
+            k_as = self._kv_cache.get('k_as')
+            v_as = self._kv_cache.get('v_as')
+            k_aa = self._kv_cache.get('k_aa')
+            v_aa = self._kv_cache.get('v_aa')
+            attn_mask_aa = self._kv_cache.get('attn_mask_aa')
+        elif self.image_cross_attention and not spatial_self_attn:
             if context.shape[1] == self.text_context_len + self.video_length:
                 context_ins, context_image = context[:, :self.text_context_len, :], context[:,self.text_context_len:, :]
-                k = self.to_k(context)
+                if self._kv_fused:
-                v = self.to_v(context)
+                    k, v = self.to_kv(context).chunk(2, dim=-1)
-                k_ip = self.to_k_ip(context_image)
+                    k_ip, v_ip = self.to_kv_ip(context_image).chunk(2, dim=-1)
-                v_ip = self.to_v_ip(context_image)
+                else:
+                    k = self.to_k(context)
+                    v = self.to_v(context)
+                    k_ip = self.to_k_ip(context_image)
+                    v_ip = self.to_v_ip(context_image)
+                k, v = map(_reshape_kv, (k, v))
+                k_ip, v_ip = map(_reshape_kv, (k_ip, v_ip))
+                if use_cache:
+                    self._kv_cache = {'k': k, 'v': v, 'k_ip': k_ip, 'v_ip': v_ip}
             elif context.shape[1] == self.agent_state_context_len + self.text_context_len + self.video_length:
                 context_agent_state = context[:, :self.agent_state_context_len, :]
                 context_ins = context[:, self.agent_state_context_len:self.agent_state_context_len+self.text_context_len, :]
                 context_image = context[:, self.agent_state_context_len+self.text_context_len:, :]
-                k = self.to_k(context_ins)
+                if self._kv_fused:
-                v = self.to_v(context_ins)
+                    k, v = self.to_kv(context_ins).chunk(2, dim=-1)
-                k_ip = self.to_k_ip(context_image)
+                    k_ip, v_ip = self.to_kv_ip(context_image).chunk(2, dim=-1)
-                v_ip = self.to_v_ip(context_image)
+                    k_as, v_as = self.to_kv_as(context_agent_state).chunk(2, dim=-1)
-                k_as = self.to_k_as(context_agent_state)
+                else:
-                v_as = self.to_v_as(context_agent_state)
+                    k = self.to_k(context_ins)
+                    v = self.to_v(context_ins)
+                    k_ip = self.to_k_ip(context_image)
+                    v_ip = self.to_v_ip(context_image)
+                    k_as = self.to_k_as(context_agent_state)
+                    v_as = self.to_v_as(context_agent_state)
+                k, v = map(_reshape_kv, (k, v))
+                k_ip, v_ip = map(_reshape_kv, (k_ip, v_ip))
+                k_as, v_as = map(_reshape_kv, (k_as, v_as))
+                if use_cache:
+                    self._kv_cache = {'k': k, 'v': v, 'k_ip': k_ip, 'v_ip': v_ip, 'k_as': k_as, 'v_as': v_as}
             else:
                 context_agent_state = context[:, :self.agent_state_context_len, :]
                 context_agent_action = context[:, self.agent_state_context_len:self.agent_state_context_len+self.agent_action_context_len, :]
                 context_ins = context[:, self.agent_state_context_len+self.agent_action_context_len:self.agent_state_context_len+self.agent_action_context_len+self.text_context_len, :]
                 context_image = context[:, self.agent_state_context_len+self.agent_action_context_len+self.text_context_len:, :]
 
-                k = self.to_k(context_ins)
+                if self._kv_fused:
-                v = self.to_v(context_ins)
+                    k, v = self.to_kv(context_ins).chunk(2, dim=-1)
-                k_ip = self.to_k_ip(context_image)
+                    k_ip, v_ip = self.to_kv_ip(context_image).chunk(2, dim=-1)
-                v_ip = self.to_v_ip(context_image)
+                    k_as, v_as = self.to_kv_as(context_agent_state).chunk(2, dim=-1)
-                k_as = self.to_k_as(context_agent_state)
+                    k_aa, v_aa = self.to_kv_aa(context_agent_action).chunk(2, dim=-1)
-                v_as = self.to_v_as(context_agent_state)
+                else:
-                k_aa = self.to_k_aa(context_agent_action)
+                    k = self.to_k(context_ins)
-                v_aa = self.to_v_aa(context_agent_action)
+                    v = self.to_v(context_ins)
+                    k_ip = self.to_k_ip(context_image)
+                    v_ip = self.to_v_ip(context_image)
+                    k_as = self.to_k_as(context_agent_state)
+                    v_as = self.to_v_as(context_agent_state)
+                    k_aa = self.to_k_aa(context_agent_action)
+                    v_aa = self.to_v_aa(context_agent_action)
 
-                attn_mask_aa = self._get_attn_mask_aa(x.shape[0],
+                k, v = map(_reshape_kv, (k, v))
-                                                      q.shape[1],
+                k_ip, v_ip = map(_reshape_kv, (k_ip, v_ip))
-                                                      k_aa.shape[1],
+                k_as, v_as = map(_reshape_kv, (k_as, v_as))
-                                                      block_size=16).to(k_aa.device)
+                k_aa, v_aa = map(_reshape_kv, (k_aa, v_aa))
+
+                attn_mask_aa_raw = self._get_attn_mask_aa(x.shape[0],
+                                                          q.shape[1],
+                                                          k_aa.shape[1],
+                                                          block_size=16,
+                                                          device=k_aa.device)
+                attn_mask_aa = attn_mask_aa_raw.unsqueeze(1).repeat(1, h, 1, 1).reshape(
+                    b * h, attn_mask_aa_raw.shape[1], attn_mask_aa_raw.shape[2]).to(q.dtype)
+
+                if use_cache:
+                    self._kv_cache = {
+                        'k': k, 'v': v, 'k_ip': k_ip, 'v_ip': v_ip,
+                        'k_as': k_as, 'v_as': v_as, 'k_aa': k_aa, 'v_aa': v_aa,
+                        'attn_mask_aa': attn_mask_aa,
+                    }
         else:
             if not spatial_self_attn:
                 assert 1 > 2, ">>> ERROR: you should never go into here ..."
                 context = context[:, :self.text_context_len, :]
-            k = self.to_k(context)
+            if self._kv_fused:
-            v = self.to_v(context)
+                k, v = self.to_kv(context).chunk(2, dim=-1)
-
+            else:
-        b, _, _ = q.shape
+                k = self.to_k(context)
-        q = q.unsqueeze(3).reshape(b, q.shape[1], self.heads, self.dim_head).permute(0, 2, 1, 3).reshape(b * self.heads, q.shape[1], self.dim_head).contiguous()
+                v = self.to_v(context)
+            k, v = map(_reshape_kv, (k, v))
+            if use_cache:
+                self._kv_cache = {'k': k, 'v': v}
         if k is not None:
-            k, v = map(
-                lambda t: t.unsqueeze(3).reshape(b, t.shape[
-                    1], self.heads, self.dim_head).permute(0, 2, 1, 3).reshape(
-                        b * self.heads, t.shape[1], self.dim_head).contiguous(),
-                (k, v),
-            )
             out = xformers.ops.memory_efficient_attention(q,
                                                           k,
                                                           v,
                                                           attn_bias=None,
                                                           op=None)
             out = (out.unsqueeze(0).reshape(
-                b, self.heads, out.shape[1],
+                b, h, out.shape[1],
                 self.dim_head).permute(0, 2, 1,
                                        3).reshape(b, out.shape[1],
-                                                  self.heads * self.dim_head))
+                                                  h * self.dim_head))
 
         if k_ip is not None:
-            # For image cross-attention
-            k_ip, v_ip = map(
-                lambda t: t.unsqueeze(3).reshape(b, t.shape[
-                    1], self.heads, self.dim_head).permute(0, 2, 1, 3).reshape(
-                        b * self.heads, t.shape[1], self.dim_head).contiguous(
-                        ),
-                (k_ip, v_ip),
-            )
             out_ip = xformers.ops.memory_efficient_attention(q,
                                                              k_ip,
                                                              v_ip,
                                                              attn_bias=None,
                                                              op=None)
             out_ip = (out_ip.unsqueeze(0).reshape(
-                b, self.heads, out_ip.shape[1],
+                b, h, out_ip.shape[1],
                 self.dim_head).permute(0, 2, 1,
                                        3).reshape(b, out_ip.shape[1],
-                                                  self.heads * self.dim_head))
+                                                  h * self.dim_head))
 
         if k_as is not None:
-            # For agent state cross-attention
-            k_as, v_as = map(
-                lambda t: t.unsqueeze(3).reshape(b, t.shape[
-                    1], self.heads, self.dim_head).permute(0, 2, 1, 3).reshape(
-                        b * self.heads, t.shape[1], self.dim_head).contiguous(
-                        ),
-                (k_as, v_as),
-            )
             out_as = xformers.ops.memory_efficient_attention(q,
                                                              k_as,
                                                              v_as,
                                                              attn_bias=None,
                                                              op=None)
             out_as = (out_as.unsqueeze(0).reshape(
-                b, self.heads, out_as.shape[1],
+                b, h, out_as.shape[1],
                 self.dim_head).permute(0, 2, 1,
                                        3).reshape(b, out_as.shape[1],
-                                                  self.heads * self.dim_head))
+                                                  h * self.dim_head))
+
         if k_aa is not None:
-            # For agent action cross-attention
-            k_aa, v_aa = map(
-                lambda t: t.unsqueeze(3).reshape(b, t.shape[
-                    1], self.heads, self.dim_head).permute(0, 2, 1, 3).reshape(
-                        b * self.heads, t.shape[1], self.dim_head).contiguous(
-                        ),
-                (k_aa, v_aa),
-            )
-
-            attn_mask_aa = attn_mask_aa.unsqueeze(1).repeat(1,self.heads,1,1).reshape(
-                    b * self.heads, attn_mask_aa.shape[1], attn_mask_aa.shape[2])
-            attn_mask_aa = attn_mask_aa.to(q.dtype)
-
             out_aa = xformers.ops.memory_efficient_attention(
                 q, k_aa, v_aa, attn_bias=attn_mask_aa, op=None)
-
             out_aa = (out_aa.unsqueeze(0).reshape(
-                b, self.heads, out_aa.shape[1],
+                b, h, out_aa.shape[1],
                 self.dim_head).permute(0, 2, 1,
                                        3).reshape(b, out_aa.shape[1],
-                                                  self.heads * self.dim_head))
+                                                  h * self.dim_head))
         if exists(mask):
             raise NotImplementedError
 
@@ -386,17 +448,43 @@ class CrossAttention(nn.Module):
 
         return self.to_out(out)
 
-    def _get_attn_mask_aa(self, b, l1, l2, block_size=16):
+    def _get_attn_mask_aa(self, b, l1, l2, block_size=16, device=None):
+        cache_key = (b, l1, l2, block_size)
+        if hasattr(self, '_attn_mask_aa_cache_key') and self._attn_mask_aa_cache_key == cache_key:
+            cached = self._attn_mask_aa_cache
+            if device is not None and cached.device != torch.device(device):
+                cached = cached.to(device)
+                self._attn_mask_aa_cache = cached
+            return cached
+
+        target_device = device if device is not None else 'cpu'
         num_token = l2 // block_size
-        start_positions = ((torch.arange(b) % block_size) + 1) * num_token
+        start_positions = ((torch.arange(b, device=target_device) % block_size) + 1) * num_token
-        col_indices = torch.arange(l2)
+        col_indices = torch.arange(l2, device=target_device)
         mask_2d = col_indices.unsqueeze(0) >= start_positions.unsqueeze(1)
         mask = mask_2d.unsqueeze(1).expand(b, l1, l2)
-        attn_mask = torch.zeros_like(mask, dtype=torch.float)
+        attn_mask = torch.zeros(b, l1, l2, dtype=torch.float, device=target_device)
         attn_mask[mask] = float('-inf')
+
+        self._attn_mask_aa_cache_key = cache_key
+        self._attn_mask_aa_cache = attn_mask
         return attn_mask
 
 
+def enable_cross_attn_kv_cache(module):
+    for m in module.modules():
+        if isinstance(m, CrossAttention):
+            m._kv_cache_enabled = True
+            m._kv_cache = {}
+
+
+def disable_cross_attn_kv_cache(module):
+    for m in module.modules():
+        if isinstance(m, CrossAttention):
+            m._kv_cache_enabled = False
+            m._kv_cache = {}
+
+
 class BasicTransformerBlock(nn.Module):
 
     def __init__(self,

src/unifolm_wma/modules/networks/wma_model.py

@@ -685,6 +685,28 @@ class WMAModel(nn.Module):
         self.action_token_projector = instantiate_from_config(
             stem_process_config)
 
+        # Context precomputation cache
+        self._ctx_cache_enabled = False
+        self._ctx_cache = {}
+        self._trt_backbone = None  # TRT engine for video UNet backbone
+        # Reusable CUDA stream for parallel state_unet / action_unet
+        self._state_stream = torch.cuda.Stream()
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state.pop('_state_stream', None)
+        return state
+
+    def __setstate__(self, state):
+        self.__dict__.update(state)
+        self._state_stream = torch.cuda.Stream()
+
+    def load_trt_backbone(self, engine_path, n_hs_a=9):
+        """Load a TensorRT engine for the video UNet backbone."""
+        from unifolm_wma.trt_utils import TRTBackbone
+        self._trt_backbone = TRTBackbone(engine_path, n_hs_a=n_hs_a)
+        print(f">>> TRT backbone loaded from {engine_path}")
+
     def forward(self,
                 x: Tensor,
                 x_action: Tensor,
@@ -720,58 +742,64 @@ class WMAModel(nn.Module):
                                    repeat_only=False).type(x.dtype)
         emb = self.time_embed(t_emb)
 
-        bt, l_context, _ = context.shape
+        _ctx_key = context.data_ptr()
-        if self.base_model_gen_only:
+        if self._ctx_cache_enabled and _ctx_key in self._ctx_cache:
-            assert l_context == 77 + self.n_obs_steps * 16, ">>> ERROR Context dim 1 ..."  ## NOTE HANDCODE
+            context = self._ctx_cache[_ctx_key]
         else:
-            if l_context == self.n_obs_steps + 77 + t * 16:
+            bt, l_context, _ = context.shape
-                context_agent_state = context[:, :self.n_obs_steps]
+            if self.base_model_gen_only:
-                context_text = context[:, self.n_obs_steps:self.n_obs_steps +
+                assert l_context == 77 + self.n_obs_steps * 16, ">>> ERROR Context dim 1 ..."  ## NOTE HANDCODE
-                                       77, :]
+            else:
-                context_img = context[:, self.n_obs_steps + 77:, :]
+                if l_context == self.n_obs_steps + 77 + t * 16:
-                context_agent_state = context_agent_state.repeat_interleave(
+                    context_agent_state = context[:, :self.n_obs_steps]
-                    repeats=t, dim=0)
+                    context_text = context[:, self.n_obs_steps:self.n_obs_steps +
-                context_text = context_text.repeat_interleave(repeats=t, dim=0)
+                                           77, :]
-                context_img = rearrange(context_img,
+                    context_img = context[:, self.n_obs_steps + 77:, :]
-                                        'b (t l) c -> (b t) l c',
+                    context_agent_state = context_agent_state.repeat_interleave(
-                                        t=t)
+                        repeats=t, dim=0)
-                context = torch.cat(
+                    context_text = context_text.repeat_interleave(repeats=t, dim=0)
-                    [context_agent_state, context_text, context_img], dim=1)
+                    context_img = rearrange(context_img,
-            elif l_context == self.n_obs_steps + 16 + 77 + t * 16:
+                                            'b (t l) c -> (b t) l c',
-                context_agent_state = context[:, :self.n_obs_steps]
+                                            t=t)
-                context_agent_action = context[:, self.
+                    context = torch.cat(
-                                               n_obs_steps:self.n_obs_steps +
+                        [context_agent_state, context_text, context_img], dim=1)
-                                               16, :]
+                elif l_context == self.n_obs_steps + 16 + 77 + t * 16:
-                context_agent_action = rearrange(
+                    context_agent_state = context[:, :self.n_obs_steps]
-                    context_agent_action.unsqueeze(2), 'b t l d -> (b t) l d')
+                    context_agent_action = context[:, self.
-                context_agent_action = self.action_token_projector(
+                                                   n_obs_steps:self.n_obs_steps +
-                    context_agent_action)
+                                                   16, :]
-                context_agent_action = rearrange(context_agent_action,
+                    context_agent_action = rearrange(
-                                                 '(b o) l d -> b o l d',
+                        context_agent_action.unsqueeze(2), 'b t l d -> (b t) l d')
-                                                 o=t)
+                    context_agent_action = self.action_token_projector(
-                context_agent_action = rearrange(context_agent_action,
+                        context_agent_action)
-                                                 'b o (t l) d -> b o t l d',
+                    context_agent_action = rearrange(context_agent_action,
-                                                 t=t)
+                                                     '(b o) l d -> b o l d',
-                context_agent_action = context_agent_action.permute(
+                                                     o=t)
-                    0, 2, 1, 3, 4)
+                    context_agent_action = rearrange(context_agent_action,
-                context_agent_action = rearrange(context_agent_action,
+                                                     'b o (t l) d -> b o t l d',
-                                                 'b t o l d -> (b t) (o l) d')
+                                                     t=t)
+                    context_agent_action = context_agent_action.permute(
+                        0, 2, 1, 3, 4)
+                    context_agent_action = rearrange(context_agent_action,
+                                                     'b t o l d -> (b t) (o l) d')
 
-                context_text = context[:, self.n_obs_steps +
+                    context_text = context[:, self.n_obs_steps +
-                                       16:self.n_obs_steps + 16 + 77, :]
+                                           16:self.n_obs_steps + 16 + 77, :]
-                context_text = context_text.repeat_interleave(repeats=t, dim=0)
+                    context_text = context_text.repeat_interleave(repeats=t, dim=0)
 
-                context_img = context[:, self.n_obs_steps + 16 + 77:, :]
+                    context_img = context[:, self.n_obs_steps + 16 + 77:, :]
-                context_img = rearrange(context_img,
+                    context_img = rearrange(context_img,
-                                        'b (t l) c -> (b t) l c',
+                                            'b (t l) c -> (b t) l c',
-                                        t=t)
+                                            t=t)
-                context_agent_state = context_agent_state.repeat_interleave(
+                    context_agent_state = context_agent_state.repeat_interleave(
-                    repeats=t, dim=0)
+                        repeats=t, dim=0)
-                context = torch.cat([
+                    context = torch.cat([
-                    context_agent_state, context_agent_action, context_text,
+                        context_agent_state, context_agent_action, context_text,
-                    context_img
+                        context_img
-                ],
+                    ],
-                                    dim=1)
+                                        dim=1)
+            if self._ctx_cache_enabled:
+                self._ctx_cache[_ctx_key] = context
 
         emb = emb.repeat_interleave(repeats=t, dim=0)
 
@@ -791,58 +819,92 @@ class WMAModel(nn.Module):
             fs_embed = fs_embed.repeat_interleave(repeats=t, dim=0)
             emb = emb + fs_embed
 
-        h = x.type(self.dtype)
+        if self._trt_backbone is not None:
-        adapter_idx = 0
+            # TRT path: run backbone via TensorRT engine
-        hs = []
+            h_in = x.type(self.dtype).contiguous()
-        hs_a = []
+            y, hs_a = self._trt_backbone(h_in, emb.contiguous(), context.contiguous())
-        for id, module in enumerate(self.input_blocks):
+        else:
-            h = module(h, emb, context=context, batch_size=b)
+            # PyTorch path: original backbone
-            if id == 0 and self.addition_attention:
+            h = x.type(self.dtype)
-                h = self.init_attn(h, emb, context=context, batch_size=b)
+            adapter_idx = 0
-            # plug-in adapter features
+            hs = []
-            if ((id + 1) % 3 == 0) and features_adapter is not None:
+            hs_a = []
-                h = h + features_adapter[adapter_idx]
+            for id, module in enumerate(self.input_blocks):
-                adapter_idx += 1
+                h = module(h, emb, context=context, batch_size=b)
-            if id != 0:
+                if id == 0 and self.addition_attention:
-                if isinstance(module[0], Downsample):
+                    h = self.init_attn(h, emb, context=context, batch_size=b)
+                # plug-in adapter features
+                if ((id + 1) % 3 == 0) and features_adapter is not None:
+                    h = h + features_adapter[adapter_idx]
+                    adapter_idx += 1
+                if id != 0:
+                    if isinstance(module[0], Downsample):
+                        hs_a.append(
+                            rearrange(hs[-1], '(b t) c h w -> b t c h w', t=t))
+                hs.append(h)
+            hs_a.append(rearrange(h, '(b t) c h w -> b t c h w', t=t))
+
+            if features_adapter is not None:
+                assert len(
+                    features_adapter) == adapter_idx, 'Wrong features_adapter'
+            h = self.middle_block(h, emb, context=context, batch_size=b)
+            hs_a.append(rearrange(h, '(b t) c h w -> b t c h w', t=t))
+
+            hs_out = []
+            for module in self.output_blocks:
+                h = torch.cat([h, hs.pop()], dim=1)
+                h = module(h, emb, context=context, batch_size=b)
+                if isinstance(module[-1], Upsample):
                     hs_a.append(
-                        rearrange(hs[-1], '(b t) c h w -> b t c h w', t=t))
+                        rearrange(hs_out[-1], '(b t) c h w -> b t c h w', t=t))
-            hs.append(h)
+                hs_out.append(h)
-        hs_a.append(rearrange(h, '(b t) c h w -> b t c h w', t=t))
+            h = h.type(x.dtype)
+            hs_a.append(rearrange(hs_out[-1], '(b t) c h w -> b t c h w', t=t))
 
-        if features_adapter is not None:
+            y = self.out(h)
-            assert len(
+            y = rearrange(y, '(b t) c h w -> b c t h w', b=b)
-                features_adapter) == adapter_idx, 'Wrong features_adapter'
-        h = self.middle_block(h, emb, context=context, batch_size=b)
-        hs_a.append(rearrange(h, '(b t) c h w -> b t c h w', t=t))
-
-        hs_out = []
-        for module in self.output_blocks:
-            h = torch.cat([h, hs.pop()], dim=1)
-            h = module(h, emb, context=context, batch_size=b)
-            if isinstance(module[-1], Upsample):
-                hs_a.append(
-                    rearrange(hs_out[-1], '(b t) c h w -> b t c h w', t=t))
-            hs_out.append(h)
-        h = h.type(x.dtype)
-        hs_a.append(rearrange(hs_out[-1], '(b t) c h w -> b t c h w', t=t))
-
-        y = self.out(h)
-        y = rearrange(y, '(b t) c h w -> b c t h w', b=b)
 
         if not self.base_model_gen_only:
             ba, _, _ = x_action.shape
+            ts_state = timesteps[:ba] if b > 1 else timesteps
+            # Run action_unet and state_unet in parallel via CUDA streams
+            s_stream = self._state_stream
+            s_stream.wait_stream(torch.cuda.current_stream())
+            with torch.cuda.stream(s_stream):
+                s_y = self.state_unet(x_state, ts_state, hs_a,
+                                      context_action[:2], **kwargs)
             a_y = self.action_unet(x_action, timesteps[:ba], hs_a,
                                    context_action[:2], **kwargs)
-            # Predict state
+            torch.cuda.current_stream().wait_stream(s_stream)
-            if b > 1:
-                s_y = self.state_unet(x_state, timesteps[:ba], hs_a,
-                                      context_action[:2], **kwargs)
-            else:
-                s_y = self.state_unet(x_state, timesteps, hs_a,
-                                      context_action[:2], **kwargs)
         else:
             a_y = torch.zeros_like(x_action)
             s_y = torch.zeros_like(x_state)
 
         return y, a_y, s_y
+
+
+def enable_ctx_cache(model):
+    """Enable context precomputation cache on WMAModel and its action/state UNets."""
+    for m in model.modules():
+        if isinstance(m, WMAModel):
+            m._ctx_cache_enabled = True
+            m._ctx_cache = {}
+    # conditional_unet1d cache
+    from unifolm_wma.models.diffusion_head.conditional_unet1d import ConditionalUnet1D
+    for m in model.modules():
+        if isinstance(m, ConditionalUnet1D):
+            m._global_cond_cache_enabled = True
+            m._global_cond_cache = {}
+
+
+def disable_ctx_cache(model):
+    """Disable and clear context precomputation cache."""
+    for m in model.modules():
+        if isinstance(m, WMAModel):
+            m._ctx_cache_enabled = False
+            m._ctx_cache = {}
+    from unifolm_wma.models.diffusion_head.conditional_unet1d import ConditionalUnet1D
+    for m in model.modules():
+        if isinstance(m, ConditionalUnet1D):
+            m._global_cond_cache_enabled = False
+            m._global_cond_cache = {}

src/unifolm_wma/trt_utils.py

@@ -0,0 +1,151 @@
+"""TensorRT acceleration utilities for the video UNet backbone."""
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from unifolm_wma.modules.networks.wma_model import Downsample, Upsample
+
+
+class VideoBackboneForExport(nn.Module):
+    """Wrapper that isolates the video UNet backbone for ONNX export.
+
+    Takes already-preprocessed inputs (after context/time embedding prep)
+    and returns y + hs_a as a flat tuple.
+    """
+
+    def __init__(self, wma_model):
+        super().__init__()
+        self.input_blocks = wma_model.input_blocks
+        self.middle_block = wma_model.middle_block
+        self.output_blocks = wma_model.output_blocks
+        self.out = wma_model.out
+        self.addition_attention = wma_model.addition_attention
+        if self.addition_attention:
+            self.init_attn = wma_model.init_attn
+        self.dtype = wma_model.dtype
+
+    def forward(self, h, emb, context):
+        t = 16
+        b = 1
+
+        hs = []
+        hs_a = []
+        h = h.type(self.dtype)
+        for id, module in enumerate(self.input_blocks):
+            h = module(h, emb, context=context, batch_size=b)
+            if id == 0 and self.addition_attention:
+                h = self.init_attn(h, emb, context=context, batch_size=b)
+            if id != 0:
+                if isinstance(module[0], Downsample):
+                    hs_a.append(rearrange(hs[-1], '(b t) c h w -> b t c h w', t=t))
+            hs.append(h)
+        hs_a.append(rearrange(h, '(b t) c h w -> b t c h w', t=t))
+
+        h = self.middle_block(h, emb, context=context, batch_size=b)
+        hs_a.append(rearrange(h, '(b t) c h w -> b t c h w', t=t))
+
+        hs_out = []
+        for module in self.output_blocks:
+            h = torch.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context=context, batch_size=b)
+            if isinstance(module[-1], Upsample):
+                hs_a.append(rearrange(hs_out[-1], '(b t) c h w -> b t c h w', t=t))
+            hs_out.append(h)
+        hs_a.append(rearrange(hs_out[-1], '(b t) c h w -> b t c h w', t=t))
+
+        y = self.out(h.type(h.dtype))
+        y = rearrange(y, '(b t) c h w -> b c t h w', b=b)
+        return (y, *hs_a)
+
+
+def export_backbone_onnx(model, save_path, context_len=95):
+    wma = model.model.diffusion_model
+    wrapper = VideoBackboneForExport(wma)
+    wrapper.eval().cuda()
+
+    for m in wrapper.modules():
+        if hasattr(m, 'checkpoint'):
+            m.checkpoint = False
+        if hasattr(m, 'use_checkpoint'):
+            m.use_checkpoint = False
+
+    import xformers.ops
+    _orig_mea = xformers.ops.memory_efficient_attention
+    def _sdpa_replacement(q, k, v, attn_bias=None, op=None, **kw):
+        return torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_bias)
+    xformers.ops.memory_efficient_attention = _sdpa_replacement
+
+    BT = 16
+    emb_dim = wma.model_channels * 4
+    ctx_dim = 1024
+    in_ch = wma.in_channels
+
+    dummy_h = torch.randn(BT, in_ch, 40, 64, device='cuda', dtype=torch.float32)
+    dummy_emb = torch.randn(BT, emb_dim, device='cuda', dtype=torch.float32)
+    dummy_ctx = torch.randn(BT, context_len, ctx_dim, device='cuda', dtype=torch.float32)
+
+    with torch.no_grad():
+        outputs = wrapper(dummy_h, dummy_emb, dummy_ctx)
+        n_outputs = len(outputs)
+        print(f">>> Backbone has {n_outputs} outputs (1 y + {n_outputs-1} hs_a)")
+        for i, o in enumerate(outputs):
+            print(f"    output[{i}]: {o.shape} {o.dtype}")
+
+    output_names = ['y'] + [f'hs_a_{i}' for i in range(n_outputs - 1)]
+
+    torch.onnx.export(
+        wrapper,
+        (dummy_h, dummy_emb, dummy_ctx),
+        save_path,
+        input_names=['h', 'emb', 'context'],
+        output_names=output_names,
+        opset_version=17,
+        do_constant_folding=True,
+    )
+    print(f">>> ONNX exported to {save_path}")
+    xformers.ops.memory_efficient_attention = _orig_mea
+    return n_outputs
+
+
+class TRTBackbone:
+    """TensorRT runtime wrapper for the video UNet backbone."""
+
+    def __init__(self, engine_path, n_hs_a=9):
+        import tensorrt as trt
+
+        self.logger = trt.Logger(trt.Logger.WARNING)
+        with open(engine_path, 'rb') as f:
+            runtime = trt.Runtime(self.logger)
+            self.engine = runtime.deserialize_cuda_engine(f.read())
+        self.context = self.engine.create_execution_context()
+        self.n_hs_a = n_hs_a
+
+        import numpy as np
+        self.output_buffers = {}
+        for i in range(self.engine.num_io_tensors):
+            name = self.engine.get_tensor_name(i)
+            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.OUTPUT:
+                shape = self.engine.get_tensor_shape(name)
+                np_dtype = trt.nptype(self.engine.get_tensor_dtype(name))
+                buf = torch.empty(list(shape), dtype=torch.from_numpy(np.empty(0, dtype=np_dtype)).dtype, device='cuda')
+                self.output_buffers[name] = buf
+                print(f"    TRT output '{name}': {list(shape)} {buf.dtype}")
+
+    def __call__(self, h, emb, context):
+        import tensorrt as trt
+        for name, tensor in [('h', h), ('emb', emb), ('context', context)]:
+            expected_dtype = trt.nptype(self.engine.get_tensor_dtype(name))
+            torch_expected = torch.from_numpy(__import__('numpy').empty(0, dtype=expected_dtype)).dtype
+            if tensor.dtype != torch_expected:
+                tensor = tensor.to(torch_expected)
+            self.context.set_tensor_address(name, tensor.contiguous().data_ptr())
+
+        for name, buf in self.output_buffers.items():
+            self.context.set_tensor_address(name, buf.data_ptr())
+
+        self.context.execute_async_v3(torch.cuda.current_stream().cuda_stream)
+        torch.cuda.synchronize()
+
+        y = self.output_buffers['y']
+        hs_a = [self.output_buffers[f'hs_a_{i}'] for i in range(self.n_hs_a)]
+        return y, hs_a

unitree_g1_pack_camera/case1/output.log

@@ -0,0 +1,179 @@
+2026-02-18 19:01:56.891895: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-02-18 19:01:56.940243: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+2026-02-18 19:01:56.940285: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+2026-02-18 19:01:56.941395: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+2026-02-18 19:01:56.948327: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-02-18 19:01:57.870809: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+Global seed set to 123
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+INFO:root:***** Configing Data *****
+>>> unitree_z1_stackbox: 1 data samples loaded.
+>>> unitree_z1_stackbox: data stats loaded.
+>>> unitree_z1_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox_v2: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: normalizer initiated.
+>>> unitree_z1_dual_arm_cleanup_pencils: 1 data samples loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: data stats loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: normalizer initiated.
+>>> unitree_g1_pack_camera: 1 data samples loaded.
+>>> unitree_g1_pack_camera: data stats loaded.
+>>> unitree_g1_pack_camera: normalizer initiated.
+>>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
+    TRT output 'y': [1, 4, 16, 40, 64] torch.float32
+    TRT output 'hs_a_0': [1, 16, 320, 40, 64] torch.float32
+    TRT output 'hs_a_1': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_2': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_3': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_4': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_5': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_6': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_7': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_8': [1, 16, 320, 40, 64] torch.float32
+>>> TRT backbone loaded from /home/qhy/unifolm-world-model-action/scripts/evaluation/../../trt_engines/video_backbone.engine
+>>> Generate 16 frames under each generation ...
+DEBUG:h5py._conv:Creating converter from 3 to 5
+DEBUG:PIL.PngImagePlugin:STREAM b'IHDR' 16 13
+DEBUG:PIL.PngImagePlugin:STREAM b'pHYs' 41 9
+DEBUG:PIL.PngImagePlugin:STREAM b'IDAT' 62 4096
+
+  0%|          | 0/11 [00:00<?, ?it/s][02/18/2026-19:02:10] [TRT] [W] Using default stream in enqueueV3() may lead to performance issues due to additional calls to cudaStreamSynchronize() by TensorRT to ensure correct synchronization. Please use non-default stream instead.
+
+  9%|▉         | 1/11 [00:17<02:51, 17.15s/it]>>> Step 0: generating actions ...
+>>> Step 0: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 1: generating actions ...
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Image: failed to import FpxImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Image: failed to import MicImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XVThumbImagePlugin
+DEBUG:PIL.Image:Importing XVThumbImagePlugin
+
+ 18%|█▊        | 2/11 [00:33<02:31, 16.87s/it]
+ 27%|██▋       | 3/11 [00:50<02:14, 16.76s/it]
+ 36%|███▋      | 4/11 [01:07<01:57, 16.81s/it]
+ 45%|████▌     | 5/11 [01:24<01:41, 16.85s/it]
+ 55%|█████▍    | 6/11 [01:41<01:24, 16.82s/it]
+ 64%|██████▎   | 7/11 [01:57<01:07, 16.82s/it]
+ 73%|███████▎  | 8/11 [02:14<00:50, 16.83s/it]
+ 82%|████████▏ | 9/11 [02:31<00:33, 16.80s/it]
+ 91%|█████████ | 10/11 [02:48<00:16, 16.81s/it]
+100%|██████████| 11/11 [03:05<00:00, 16.81s/it]
+100%|██████████| 11/11 [03:05<00:00, 16.83s/it]
+>>> Step 1: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 2: generating actions ...
+>>> Step 2: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 3: generating actions ...
+>>> Step 3: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 4: generating actions ...
+>>> Step 4: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...
+>>> Step 5: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 6: generating actions ...
+>>> Step 6: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_g1_pack_camera/case1/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_g1_pack_camera/case1/unitree_g1_pack_camera_case1.mp4",
+    "pred_video": "unitree_g1_pack_camera/case1/output/inference/0_full_fs6.mp4",
+    "psnr": 35.615362167470806
+}

unitree_g1_pack_camera/case1/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_g1_pack_camera/case1"
+dataset="unitree_g1_pack_camera"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_g1_pack_camera/case1/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 6 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 11 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_g1_pack_camera/case1/world_model_interaction_prompts/unitree_g1_pack_camera.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+0,x,x,unitree_g1_pack_camera,mount camera,x,x,x,G1_Dex1,30

unitree_g1_pack_camera/case2/output.log

@@ -0,0 +1,179 @@
+2026-02-18 19:05:45.956647: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-02-18 19:05:46.004149: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+2026-02-18 19:05:46.004193: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+2026-02-18 19:05:46.005265: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+2026-02-18 19:05:46.012074: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-02-18 19:05:46.932966: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+Global seed set to 123
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+INFO:root:***** Configing Data *****
+>>> unitree_z1_stackbox: 1 data samples loaded.
+>>> unitree_z1_stackbox: data stats loaded.
+>>> unitree_z1_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox_v2: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: normalizer initiated.
+>>> unitree_z1_dual_arm_cleanup_pencils: 1 data samples loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: data stats loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: normalizer initiated.
+>>> unitree_g1_pack_camera: 1 data samples loaded.
+>>> unitree_g1_pack_camera: data stats loaded.
+>>> unitree_g1_pack_camera: normalizer initiated.
+>>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
+    TRT output 'y': [1, 4, 16, 40, 64] torch.float32
+    TRT output 'hs_a_0': [1, 16, 320, 40, 64] torch.float32
+    TRT output 'hs_a_1': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_2': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_3': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_4': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_5': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_6': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_7': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_8': [1, 16, 320, 40, 64] torch.float32
+>>> TRT backbone loaded from /home/qhy/unifolm-world-model-action/scripts/evaluation/../../trt_engines/video_backbone.engine
+>>> Generate 16 frames under each generation ...
+DEBUG:h5py._conv:Creating converter from 3 to 5
+DEBUG:PIL.PngImagePlugin:STREAM b'IHDR' 16 13
+DEBUG:PIL.PngImagePlugin:STREAM b'pHYs' 41 9
+DEBUG:PIL.PngImagePlugin:STREAM b'IDAT' 62 4096
+
+  0%|          | 0/11 [00:00<?, ?it/s][02/18/2026-19:05:59] [TRT] [W] Using default stream in enqueueV3() may lead to performance issues due to additional calls to cudaStreamSynchronize() by TensorRT to ensure correct synchronization. Please use non-default stream instead.
+
+  9%|▉         | 1/11 [00:16<02:47, 16.71s/it]>>> Step 0: generating actions ...
+>>> Step 0: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 1: generating actions ...
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Image: failed to import FpxImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Image: failed to import MicImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XVThumbImagePlugin
+DEBUG:PIL.Image:Importing XVThumbImagePlugin
+
+ 18%|█▊        | 2/11 [00:33<02:30, 16.75s/it]
+ 27%|██▋       | 3/11 [00:50<02:15, 16.91s/it]
+ 36%|███▋      | 4/11 [01:07<01:59, 17.02s/it]
+ 45%|████▌     | 5/11 [01:24<01:41, 16.98s/it]
+ 55%|█████▍    | 6/11 [01:41<01:24, 16.94s/it]
+ 64%|██████▎   | 7/11 [01:58<01:07, 16.90s/it]
+ 73%|███████▎  | 8/11 [02:15<00:50, 16.83s/it]
+ 82%|████████▏ | 9/11 [02:31<00:33, 16.80s/it]
+ 91%|█████████ | 10/11 [02:49<00:16, 16.94s/it]
+100%|██████████| 11/11 [03:06<00:00, 16.97s/it]
+100%|██████████| 11/11 [03:06<00:00, 16.91s/it]
+>>> Step 1: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 2: generating actions ...
+>>> Step 2: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 3: generating actions ...
+>>> Step 3: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 4: generating actions ...
+>>> Step 4: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...
+>>> Step 5: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 6: generating actions ...
+>>> Step 6: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_g1_pack_camera/case2/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_g1_pack_camera/case2/unitree_g1_pack_camera_case2.mp4",
+    "pred_video": "unitree_g1_pack_camera/case2/output/inference/50_full_fs6.mp4",
+    "psnr": 34.61979248212279
+}

unitree_g1_pack_camera/case2/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_g1_pack_camera/case2"
+dataset="unitree_g1_pack_camera"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_g1_pack_camera/case2/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 6 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 11 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_g1_pack_camera/case2/world_model_interaction_prompts/unitree_g1_pack_camera.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+50,x,x,unitree_g1_pack_camera,mount camera,x,x,x,G1_Dex1,30

unitree_g1_pack_camera/case3/output.log

@@ -0,0 +1,146 @@
+2026-02-18 19:09:35.113634: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-02-18 19:09:35.161428: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+2026-02-18 19:09:35.161474: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+2026-02-18 19:09:35.162551: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+2026-02-18 19:09:35.169325: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-02-18 19:09:36.089250: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+Global seed set to 123
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+INFO:root:***** Configing Data *****
+>>> unitree_z1_stackbox: 1 data samples loaded.
+>>> unitree_z1_stackbox: data stats loaded.
+>>> unitree_z1_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox_v2: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: normalizer initiated.
+>>> unitree_z1_dual_arm_cleanup_pencils: 1 data samples loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: data stats loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: normalizer initiated.
+>>> unitree_g1_pack_camera: 1 data samples loaded.
+>>> unitree_g1_pack_camera: data stats loaded.
+>>> unitree_g1_pack_camera: normalizer initiated.
+>>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
+    TRT output 'y': [1, 4, 16, 40, 64] torch.float32
+    TRT output 'hs_a_0': [1, 16, 320, 40, 64] torch.float32
+    TRT output 'hs_a_1': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_2': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_3': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_4': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_5': [1, 16, 1280, 5, 8] torch.float32
+    TRT output 'hs_a_6': [1, 16, 1280, 10, 16] torch.float32
+    TRT output 'hs_a_7': [1, 16, 640, 20, 32] torch.float32
+    TRT output 'hs_a_8': [1, 16, 320, 40, 64] torch.float32
+>>> TRT backbone loaded from /home/qhy/unifolm-world-model-action/scripts/evaluation/../../trt_engines/video_backbone.engine
+>>> Generate 16 frames under each generation ...
+DEBUG:h5py._conv:Creating converter from 3 to 5
+DEBUG:PIL.PngImagePlugin:STREAM b'IHDR' 16 13
+DEBUG:PIL.PngImagePlugin:STREAM b'pHYs' 41 9
+DEBUG:PIL.PngImagePlugin:STREAM b'IDAT' 62 4096
+
+  0%|          | 0/11 [00:00<?, ?it/s][02/18/2026-19:09:49] [TRT] [W] Using default stream in enqueueV3() may lead to performance issues due to additional calls to cudaStreamSynchronize() by TensorRT to ensure correct synchronization. Please use non-default stream instead.
+
+  9%|▉         | 1/11 [00:16<02:45, 16.53s/it]>>> Step 0: generating actions ...
+>>> Step 0: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 1: generating actions ...
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Image: failed to import FpxImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Image: failed to import MicImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XVThumbImagePlugin

unitree_g1_pack_camera/case3/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_g1_pack_camera/case3/unitree_g1_pack_camera_case3.mp4",
+    "pred_video": "unitree_g1_pack_camera/case3/output/inference/100_full_fs6.mp4",
+    "psnr": 37.034952654534486
+}

unitree_g1_pack_camera/case3/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_g1_pack_camera/case3"
+dataset="unitree_g1_pack_camera"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_g1_pack_camera/case3/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 6 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 11 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_g1_pack_camera/case3/world_model_interaction_prompts/unitree_g1_pack_camera.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+100,x,x,unitree_g1_pack_camera,mount camera,x,x,x,G1_Dex1,30

unitree_g1_pack_camera/case4/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_g1_pack_camera/case4/unitree_g1_pack_camera_case4.mp4",
+    "pred_video": "unitree_g1_pack_camera/case4/output/inference/200_full_fs6.mp4",
+    "psnr": 31.43390896360405
+}

unitree_g1_pack_camera/case4/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_g1_pack_camera/case4"
+dataset="unitree_g1_pack_camera"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_g1_pack_camera/case4/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 6 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 11 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_g1_pack_camera/case4/world_model_interaction_prompts/unitree_g1_pack_camera.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+200,x,x,unitree_g1_pack_camera,mount camera,x,x,x,G1_Dex1,30

unitree_z1_dual_arm_cleanup_pencils/case1/output.log

@@ -0,0 +1,121 @@
+2026-02-10 15:38:28.973314: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
+2026-02-10 15:38:29.023024: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+2026-02-10 15:38:29.023070: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+2026-02-10 15:38:29.024393: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+2026-02-10 15:38:29.031901: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-02-10 15:38:29.955454: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+Global seed set to 123
+INFO:mainlogger:LatentVisualDiffusion: Running in v-prediction mode
+INFO:unifolm_wma.models.diffusion_head.conditional_unet1d:number of parameters: 5.010531e+08
+INFO:unifolm_wma.models.diffusion_head.conditional_unet1d:number of parameters: 5.010531e+08
+AE working on z of shape (1, 4, 32, 32) = 4096 dimensions.
+INFO:root:Loaded ViT-H-14 model config.
+DEBUG:urllib3.connectionpool:Starting new HTTPS connection (1): hf-mirror.com:443
+DEBUG:urllib3.connectionpool:https://hf-mirror.com:443 "HEAD /laion/CLIP-ViT-H-14-laion2B-s32B-b79K/resolve/main/open_clip_pytorch_model.bin HTTP/1.1" 302 0
+INFO:root:Loading pretrained ViT-H-14 weights (laion2b_s32b_b79k).
+INFO:root:Loaded ViT-H-14 model config.
+DEBUG:urllib3.connectionpool:https://hf-mirror.com:443 "HEAD /laion/CLIP-ViT-H-14-laion2B-s32B-b79K/resolve/main/open_clip_pytorch_model.bin HTTP/1.1" 302 0
+INFO:root:Loading pretrained ViT-H-14 weights (laion2b_s32b_b79k).
+>>> model checkpoint loaded.
+>>> Load pre-trained model ...
+INFO:root:***** Configing Data *****
+>>> unitree_z1_stackbox: 1 data samples loaded.
+>>> unitree_z1_stackbox: data stats loaded.
+>>> unitree_z1_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox: normalizer initiated.
+>>> unitree_z1_dual_arm_stackbox_v2: 1 data samples loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: data stats loaded.
+>>> unitree_z1_dual_arm_stackbox_v2: normalizer initiated.
+>>> unitree_z1_dual_arm_cleanup_pencils: 1 data samples loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: data stats loaded.
+>>> unitree_z1_dual_arm_cleanup_pencils: normalizer initiated.
+>>> unitree_g1_pack_camera: 1 data samples loaded.
+>>> unitree_g1_pack_camera: data stats loaded.
+>>> unitree_g1_pack_camera: normalizer initiated.
+>>> Dataset is successfully loaded ...
+>>> Generate 16 frames under each generation ...
+DEBUG:h5py._conv:Creating converter from 3 to 5
+DEBUG:PIL.PngImagePlugin:STREAM b'IHDR' 16 13
+DEBUG:PIL.PngImagePlugin:STREAM b'pHYs' 41 9
+DEBUG:PIL.PngImagePlugin:STREAM b'IDAT' 62 4096
+
+  0%|          | 0/8 [00:00<?, ?it/s]>>> Step 0: generating actions ...
+>>> Step 0: interacting with world model ...
+DEBUG:PIL.Image:Importing BlpImagePlugin
+DEBUG:PIL.Image:Importing BmpImagePlugin
+DEBUG:PIL.Image:Importing BufrStubImagePlugin
+DEBUG:PIL.Image:Importing CurImagePlugin
+DEBUG:PIL.Image:Importing DcxImagePlugin
+DEBUG:PIL.Image:Importing DdsImagePlugin
+DEBUG:PIL.Image:Importing EpsImagePlugin
+DEBUG:PIL.Image:Importing FitsImagePlugin
+DEBUG:PIL.Image:Importing FitsStubImagePlugin
+DEBUG:PIL.Image:Importing FliImagePlugin
+DEBUG:PIL.Image:Importing FpxImagePlugin
+DEBUG:PIL.Image:Image: failed to import FpxImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing FtexImagePlugin
+DEBUG:PIL.Image:Importing GbrImagePlugin
+DEBUG:PIL.Image:Importing GifImagePlugin
+DEBUG:PIL.Image:Importing GribStubImagePlugin
+DEBUG:PIL.Image:Importing Hdf5StubImagePlugin
+DEBUG:PIL.Image:Importing IcnsImagePlugin
+DEBUG:PIL.Image:Importing IcoImagePlugin
+DEBUG:PIL.Image:Importing ImImagePlugin
+DEBUG:PIL.Image:Importing ImtImagePlugin
+DEBUG:PIL.Image:Importing IptcImagePlugin
+DEBUG:PIL.Image:Importing JpegImagePlugin
+DEBUG:PIL.Image:Importing Jpeg2KImagePlugin
+DEBUG:PIL.Image:Importing McIdasImagePlugin
+DEBUG:PIL.Image:Importing MicImagePlugin
+DEBUG:PIL.Image:Image: failed to import MicImagePlugin: No module named 'olefile'
+DEBUG:PIL.Image:Importing MpegImagePlugin
+DEBUG:PIL.Image:Importing MpoImagePlugin
+DEBUG:PIL.Image:Importing MspImagePlugin
+DEBUG:PIL.Image:Importing PalmImagePlugin
+DEBUG:PIL.Image:Importing PcdImagePlugin
+DEBUG:PIL.Image:Importing PcxImagePlugin
+DEBUG:PIL.Image:Importing PdfImagePlugin
+DEBUG:PIL.Image:Importing PixarImagePlugin
+DEBUG:PIL.Image:Importing PngImagePlugin
+DEBUG:PIL.Image:Importing PpmImagePlugin
+DEBUG:PIL.Image:Importing PsdImagePlugin
+DEBUG:PIL.Image:Importing QoiImagePlugin
+DEBUG:PIL.Image:Importing SgiImagePlugin
+DEBUG:PIL.Image:Importing SpiderImagePlugin
+DEBUG:PIL.Image:Importing SunImagePlugin
+DEBUG:PIL.Image:Importing TgaImagePlugin
+DEBUG:PIL.Image:Importing TiffImagePlugin
+DEBUG:PIL.Image:Importing WebPImagePlugin
+DEBUG:PIL.Image:Importing WmfImagePlugin
+DEBUG:PIL.Image:Importing XbmImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
+DEBUG:PIL.Image:Importing XVThumbImagePlugin
+
+ 12%|█▎        | 1/8 [01:14<08:41, 74.51s/it]
+ 25%|██▌       | 2/8 [02:29<07:28, 74.79s/it]
+ 38%|███▊      | 3/8 [03:44<06:14, 74.81s/it]
+ 50%|█████     | 4/8 [04:59<04:59, 74.78s/it]
+ 62%|██████▎   | 5/8 [06:13<03:44, 74.73s/it]
+ 75%|███████▌  | 6/8 [07:28<02:29, 74.66s/it]
+ 88%|████████▊ | 7/8 [08:42<01:14, 74.56s/it]
+100%|██████████| 8/8 [09:56<00:00, 74.51s/it]
+100%|██████████| 8/8 [09:56<00:00, 74.62s/it]
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 1: generating actions ...
+>>> Step 1: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 2: generating actions ...
+>>> Step 2: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 3: generating actions ...
+>>> Step 3: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 4: generating actions ...
+>>> Step 4: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...
+>>> Step 5: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_z1_dual_arm_cleanup_pencils/case1/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_z1_dual_arm_cleanup_pencils/case1/unitree_z1_dual_arm_cleanup_pencils_case1.mp4",
+    "pred_video": "unitree_z1_dual_arm_cleanup_pencils/case1/output/inference/0_full_fs4.mp4",
+    "psnr": 47.911564449209735
+}

unitree_z1_dual_arm_cleanup_pencils/case1/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_z1_dual_arm_cleanup_pencils/case1"
+dataset="unitree_z1_dual_arm_cleanup_pencils"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_z1_dual_arm_cleanup_pencils/case1/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 4 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 8 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_cleanup_pencils/case1/world_model_interaction_prompts/unitree_z1_dual_arm_cleanup_pencils.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+0,x,x,unitree_z1_dual_arm_cleanup_pencils,clean up eraser and pencils,x,x,x,Z1_Dual_Dex1,30

unitree_z1_dual_arm_cleanup_pencils/case2/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_z1_dual_arm_cleanup_pencils/case2/unitree_z1_dual_arm_cleanup_pencils_case2.mp4",
+    "pred_video": "unitree_z1_dual_arm_cleanup_pencils/case2/output/inference/50_full_fs4.mp4",
+    "psnr": 48.344571927558974
+}

unitree_z1_dual_arm_cleanup_pencils/case2/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_z1_dual_arm_cleanup_pencils/case2"
+dataset="unitree_z1_dual_arm_cleanup_pencils"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_z1_dual_arm_cleanup_pencils/case2/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 4 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 8 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_cleanup_pencils/case2/world_model_interaction_prompts/unitree_z1_dual_arm_cleanup_pencils.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+50,x,x,unitree_z1_dual_arm_cleanup_pencils,clean up eraser and pencils,x,x,x,Z1_Dual_Dex1,30

unitree_z1_dual_arm_cleanup_pencils/case3/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_z1_dual_arm_cleanup_pencils/case3/unitree_z1_dual_arm_cleanup_pencils_case3.mp4",
+    "pred_video": "unitree_z1_dual_arm_cleanup_pencils/case3/output/inference/100_full_fs4.mp4",
+    "psnr": 41.152374490134825
+}

unitree_z1_dual_arm_cleanup_pencils/case3/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_z1_dual_arm_cleanup_pencils/case3"
+dataset="unitree_z1_dual_arm_cleanup_pencils"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_z1_dual_arm_cleanup_pencils/case3/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 4 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 8 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_cleanup_pencils/case3/world_model_interaction_prompts/unitree_z1_dual_arm_cleanup_pencils.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+100,x,x,unitree_z1_dual_arm_cleanup_pencils,clean up eraser and pencils,x,x,x,Z1_Dual_Dex1,30

unitree_z1_dual_arm_cleanup_pencils/case4/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_z1_dual_arm_cleanup_pencils/case4/unitree_z1_dual_arm_cleanup_pencils_case4.mp4",
+    "pred_video": "unitree_z1_dual_arm_cleanup_pencils/case4/output/inference/200_full_fs4.mp4",
+    "psnr": 46.025723557253855
+}

unitree_z1_dual_arm_cleanup_pencils/case4/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_z1_dual_arm_cleanup_pencils/case4"
+dataset="unitree_z1_dual_arm_cleanup_pencils"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_z1_dual_arm_cleanup_pencils/case4/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 4 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 8 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_cleanup_pencils/case4/world_model_interaction_prompts/unitree_z1_dual_arm_cleanup_pencils.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+200,x,x,unitree_z1_dual_arm_cleanup_pencils,clean up eraser and pencils,x,x,x,Z1_Dual_Dex1,30

unitree_z1_dual_arm_stackbox/case1/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_z1_dual_arm_stackbox/case1/unitree_z1_dual_arm_stackbox_case1.mp4",
+    "pred_video": "unitree_z1_dual_arm_stackbox/case1/output/inference/5_full_fs4.mp4",
+    "psnr": 44.3480149502738
+}

unitree_z1_dual_arm_stackbox/case1/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_z1_dual_arm_stackbox/case1"
+dataset="unitree_z1_dual_arm_stackbox"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_z1_dual_arm_stackbox/case1/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 4 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 7 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox/case1/world_model_interaction_prompts/unitree_z1_dual_arm_stackbox.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+5,x,x,unitree_z1_dual_arm_stackbox,"Pick up the red cup on the table.",x,x,x,Unitree Z1 Robot Dual-Arm,30

unitree_z1_dual_arm_stackbox/case2/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_z1_dual_arm_stackbox/case2/unitree_z1_dual_arm_stackbox_case2.mp4",
+    "pred_video": "unitree_z1_dual_arm_stackbox/case2/output/inference/15_full_fs4.mp4",
+    "psnr": 39.867728254007716
+}

unitree_z1_dual_arm_stackbox/case2/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_z1_dual_arm_stackbox/case2"
+dataset="unitree_z1_dual_arm_stackbox"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_z1_dual_arm_stackbox/case2/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 4 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 7 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox/case2/world_model_interaction_prompts/unitree_z1_dual_arm_stackbox.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+15,x,x,unitree_z1_dual_arm_stackbox,"Pick up the red cup on the table.",x,x,x,Unitree Z1 Robot Dual-Arm,30

unitree_z1_dual_arm_stackbox/case3/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_z1_dual_arm_stackbox/case3/unitree_z1_dual_arm_stackbox_case3.mp4",
+    "pred_video": "unitree_z1_dual_arm_stackbox/case3/output/inference/25_full_fs4.mp4",
+    "psnr": 39.19101039445159
+}

unitree_z1_dual_arm_stackbox/case3/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_z1_dual_arm_stackbox/case3"
+dataset="unitree_z1_dual_arm_stackbox"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_z1_dual_arm_stackbox/case3/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 4 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 7 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox/case3/world_model_interaction_prompts/unitree_z1_dual_arm_stackbox.csv

@@ -0,0 +1,2 @@
+videoid,contentUrl,duration,data_dir,instruction,dynamic_confidence,dynamic_wording,dynamic_source_category,embodiment,fps
+25,x,x,unitree_z1_dual_arm_stackbox,"Pick up the red cup on the table.",x,x,x,Unitree Z1 Robot Dual-Arm,30

unitree_z1_dual_arm_stackbox/case4/psnr_result.json

@@ -0,0 +1,5 @@
+{
+    "gt_video": "unitree_z1_dual_arm_stackbox/case4/unitree_z1_dual_arm_stackbox_case4.mp4",
+    "pred_video": "unitree_z1_dual_arm_stackbox/case4/output/inference/35_full_fs4.mp4",
+    "psnr": 40.29563315341769
+}

unitree_z1_dual_arm_stackbox/case4/run_world_model_interaction.sh

@@ -0,0 +1,24 @@
+res_dir="unitree_z1_dual_arm_stackbox/case4"
+dataset="unitree_z1_dual_arm_stackbox"
+
+{
+    time CUDA_VISIBLE_DEVICES=0 python3 scripts/evaluation/world_model_interaction.py \
+        --seed 123 \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
+        --config configs/inference/world_model_interaction.yaml \
+        --savedir "${res_dir}/output" \
+        --bs 1 --height 320 --width 512 \
+        --unconditional_guidance_scale 1.0 \
+        --ddim_steps 50 \
+        --ddim_eta 1.0 \
+        --prompt_dir "unitree_z1_dual_arm_stackbox/case4/world_model_interaction_prompts" \
+        --dataset ${dataset} \
+        --video_length 16 \
+        --frame_stride 4 \
+        --n_action_steps 16 \
+        --exe_steps 16 \
+        --n_iter 7 \
+        --timestep_spacing 'uniform_trailing' \
+        --guidance_rescale 0.7 \
+        --perframe_ae
+} 2>&1 | tee "${res_dir}/output.log"