优化写入后新的所有结果

- world model 调用 decode_video=False，跳过 16 帧全量 decode
- 只 decode 最后 1 帧给 CLIP embedding / observation queue
- 存 raw latent，循环结束后统一 batch decode 生成最终视频
- 每轮省 15 次 VAE decode，8 轮共省 120 次
- 跳过中间迭代的 wm tensorboard/mp4 保存
psnr微弱下降

.claude/settings.local.json

@@ -0,0 +1,21 @@
+{
+  "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:*)",
+      "Bash(ls:*)",
+      "Bash(for scenario in unitree_g1_pack_camera unitree_z1_dual_arm_cleanup_pencils unitree_z1_dual_arm_stackbox unitree_z1_dual_arm_stackbox_v2 unitree_z1_stackbox)",
+      "Bash(do for case in case1 case2 case3 case4)",
+      "Bash(done)",
+      "Bash(chmod:*)",
+      "Bash(ln:*)"
+    ]
+  }
+}

.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:
-
 local_settings.py
 db.sqlite3
 
@@ -121,6 +120,7 @@ localTest/
 fig/
 figure/
 *.mp4
+
 Data/ControlVAE.yml
 Data/Misc
 Data/Pretrained
@@ -129,4 +129,6 @@ Experiment/checkpoint
 Experiment/log
 
 *.ckpt
-*.0
+
+*.0
+ckpts/unifolm_wma_dual.ckpt.prepared.pt

case4_run.log

@@ -1,135 +0,0 @@
-nohup: ignoring input
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 07:38:45.572744: 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`.
-2026-02-08 07:38:45.576864: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:38:45.624825: 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
-2026-02-08 07:38:45.624883: 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
-2026-02-08 07:38:45.627150: 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
-2026-02-08 07:38:45.638316: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:38:45.638803: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 07:38:46.426363: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> 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%|          | 0/7 [00:00<?, ?it/s]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
-  attn_output = scaled_dot_product_attention(
->>> 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
-
- 14%|█▍        | 1/7 [01:38<09:52, 98.73s/it]
- 29%|██▊       | 2/7 [03:17<08:14, 98.85s/it]
- 43%|████▎     | 3/7 [04:56<06:35, 98.80s/it]
- 57%|█████▋    | 4/7 [06:35<04:56, 98.94s/it]
- 71%|███████▏  | 5/7 [08:14<03:17, 98.93s/it]
- 86%|████████▌ | 6/7 [09:53<01:38, 98.89s/it]
-100%|██████████| 7/7 [11:31<00:00, 98.81s/it]
-100%|██████████| 7/7 [11:31<00:00, 98.85s/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 ...

ckpts/configuration.json

@@ -1 +0,0 @@
-{"framework": "pytorch", "task": "robotics", "allow_remote": true}

configs/inference/world_model_interaction.yaml

@@ -222,7 +222,7 @@ data:
     test:
       target: unifolm_wma.data.wma_data.WMAData
       params:
-        data_dir: '/mnt/ASC1637/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

docs/architecture_overview.typ

@@ -0,0 +1,122 @@
+== Task Comprehension: Diffusion Model and UnifoLM-WMA
+
+This section provides a comprehensive overview of the UnifoLM-WMA-0 deep learning architecture, serving as a practical foundation for the optimization strategies discussed in subsequent sections.
+
+=== Overall Inference Pipeline
+
+UnifoLM-WMA-0 is Unitree Robotics' open-source World-Model-Action framework. Its core task is to predict future video frame sequences along with the corresponding robot action and state trajectories, given a current observation image and a text instruction. The model operates in an interactive simulation mode: each iteration consumes the previous prediction as input and generates the next segment of video and actions, thereby forming a closed-loop rollout. A single iteration of this pipeline can be decomposed into four sequential stages — condition encoding, VAE encoding, DDIM diffusion sampling, and VAE decoding — each of which is described below.
+
+==== Condition Encoding
+
+The condition encoding stage transforms raw multi-modal inputs into a unified context vector that guides the diffusion denoising process, through three parallel encoding paths. On the image side, the input observation image (320#sym.times 512) is processed by a frozen OpenCLIP ViT-H-14 vision encoder, then compressed through a Resampler — a Perceiver-based cross-attention module (4 layers, 12 heads, dim\_head=64, embed\_dim 1280 #sym.arrow 1024) — into 16 image condition tokens per frame, yielding $16 times T = 256$ image tokens for T=16 frames.
+
+On the text side, the instruction is encoded by a frozen OpenCLIP text encoder (`FrozenOpenCLIPEmbedder`, penultimate layer output) into 77 tokens of dimension 1024, computed once and reused across all DDIM steps. On the state side, the robot proprioceptive state (dim 16) is mapped through a SATokenProjector (Perceiver Attention, 1 layer, 16 heads, dim\_head=64, 16 learnable queries) into 16 tokens of dimension 1024.
+
+These three token sets are concatenated to form the unified context vector: `[agent_state(2) | agent_action(16) | text(77) | image(256)]`, totaling 351 tokens per cross-attention operation.
+
+==== VAE Encoding
+
+The observation images are encoded into a compact latent space through an AutoencoderKL (`autoencoder.py`) — a variational autoencoder regularized by KL divergence. The encoder follows a convolutional architecture with 4-level channel multipliers [1, 2, 4, 4] (base channels ch=128, yielding channel widths [128, 256, 512, 512]), 2 residual blocks per level, and a latent channel count of z\_channels=4. The input RGB frames at resolution 320#sym.times 512 are encoded into latent representations at 1/8 spatial resolution, producing tensors of shape `(B, 4, T, 40, 64)`.
+
+A critical configuration parameter is `perframe_ae=True`, which means the VAE processes each of the T=16 frames independently rather than as a 3D volume. While this per-frame strategy avoids the memory overhead of volumetric convolutions, it introduces a sequential loop of T forward passes through the encoder — a point worth noting for latency optimization. The latent representations are scaled by a fixed factor of `scale_factor=0.18215` before being fed into the diffusion process.
+
+==== DDIM Diffusion Sampling
+
+This is the core time-consuming part of inference. A DDIM (Denoising Diffusion Implicit Models) sampler (`ddim.py`) is employed with a default of 50 denoising steps. The diffusion process is parameterized with v-prediction (`parameterization="v"`), 1000 training timesteps, and a linear beta schedule from `linear_start=0.00085` to `linear_end=0.012`, with zero-SNR terminal rescaling enabled (`rescale_betas_zero_snr=True`) and dynamic rescaling applied at `base_scale=0.7` to stabilize generation quality.
+
+Unlike standard video diffusion models that only predict denoised video latents, UnifoLM-WMA simultaneously produces three outputs per step: a video latent prediction `y` of shape `(B, 4, T, 40, 64)`, an action trajectory prediction `a_y` of shape `(B, T, 16)`, and a state trajectory prediction `s_y` of shape `(B, T, 16)`. The three predictions share the same diffusion timestep but employ heterogeneous noise schedules — the video stream uses the DDPM schedule with v-prediction, while the action and state streams use a `DDIMScheduler` from the `diffusers` library with epsilon-prediction and a `squaredcos_cap_v2` beta schedule. This design allows each modality to adopt its optimal denoising strategy.
+
+The sampler also supports classifier-free guidance with `unconditional_guidance_scale` and guidance rescaling, applied only to the video stream to balance generation quality and diversity.
+
+==== VAE Decoding
+
+After the DDIM sampling loop completes, the denoised video latent tensor $x_0$ of shape `(B, 4, T, 40, 64)` is decoded back to RGB pixel space through the AutoencoderKL decoder. Due to the `perframe_ae=True` configuration, decoding is likewise performed frame-by-frame: each of the T=16 latent frames is individually inverse-scaled by $1 slash "scale_factor"$, passed through the decoder's convolutional transpose layers, and reconstructed to a 320#sym.times 512 RGB frame.
+
+In the interactive simulation mode, the decoded video serves a dual purpose — providing the observation image for the next iteration's condition encoding (only the first `exe_steps` frames are needed) and producing the final output video for visualization and evaluation. The action and state trajectories predicted by the DDIM loop are directly used for robot control without further decoding.
+
+=== WMAModel Backbone: Dual-UNet Collaborative Architecture
+
+The WMAModel (`wma_model.py:326`) is the core neural network invoked at every DDIM step, employing a unique dual-UNet collaborative architecture that jointly predicts video, actions, and states within a single forward pass. This tightly-coupled design enables the action and state predictions to directly leverage the rich spatiotemporal features extracted by the video generation backbone, rather than treating them as independent prediction heads.
+
+==== Video UNet
+
+The primary backbone is a 2D convolution-based UNet with temporal extensions. Its key configuration is summarized in the following table:
+
+#figure(
+  table(
+    columns: (3fr, 5fr),
+    [*Parameter*], [*Value*],
+    [Input / Output channels], [8 (4 latent + 4 conditioning) / 4],
+    [Base model channels], [320],
+    [Channel multipliers], [\[1, 2, 4, 4\] #sym.arrow widths \[320, 640, 1280, 1280\]],
+    [Residual blocks per level], [2],
+    [Attention resolutions], [\[4, 2, 1\] (3 of 4 resolution levels)],
+    [Attention head channels], [64],
+    [Transformer depth], [1 per attention resolution],
+    [Context dimension], [1024],
+    [Temporal length], [16 frames],
+  ),
+  caption: [Video UNet configuration parameters.],
+)
+
+The UNet follows the classic encoder-middle-decoder structure with skip connections. At each attention-enabled resolution level, every ResBlock is followed by two transformer modules: a SpatialTransformer that performs spatial self-attention among all $H times W$ tokens within each frame followed by cross-attention with the 351-token context vector, and a TemporalTransformer that performs self-attention among T=16 time-step tokens at each spatial position (configured with `temporal_selfatt_only=True`, i.e., no cross-attention).
+
+During the forward pass, intermediate feature maps are collected after each Downsample layer and the middle block, reshaped from $(B times T, C, H, W)$ to $(B, T, C, H, W)$, accumulating 10 multi-scale feature maps in `hs_a` — the bridge to the Action/State UNets.
+
+==== Action UNet and State UNet
+
+The Action UNet (`conditional_unet1d.py`) is a 1D convolutional UNet specifically designed for predicting robot action trajectories. Its configuration is as follows:
+
+#figure(
+  table(
+    columns: (3fr, 5fr),
+    [*Parameter*], [*Value*],
+    [Input dimension], [16 (agent\_action\_dim)],
+    [Down channel widths], [\[256, 512, 1024, 2048\]],
+    [Kernel size], [5],
+    [GroupNorm groups], [8],
+    [Diffusion step embedding dim], [128],
+    [Horizon], [16],
+    [Action projection dim], [32],
+  ),
+  caption: [Action UNet (ConditionalUnet1D) configuration parameters.],
+)
+
+The Action UNet receives the 10 `hs_a` feature maps from the Video UNet as visual conditioning. The conditioning pipeline involves three stages: (1) SpatialSoftmax compresses each 2D feature map into keypoint coordinates $(B times T, C, 2)$; (2) the compressed features are concatenated with the diffusion timestep embedding and observation encoding (ResNet-18 `MultiImageObsEncoder`), then injected via FiLM modulation to produce per-channel scale/bias for the 1D convolution blocks; (3) `ActionLatentImageCrossAttention` enables action tokens to cross-attend to the Video UNet's spatiotemporal features, allowing visually-grounded action planning.
+
+The input action tensor $(B, T, 16)$ is projected to act\_proj\_dim=32, processed through the 1D UNet, then projected back to $(B, T, 16)$.
+
+The State UNet is an identical `ConditionalUnet1D` instance with the same hyperparameters, operating on the state tensor `x_state` $(B, T, 16)$ instead of the action tensor.
+
+A critical optimization observation: the Action and State UNets are computationally independent — sharing read-only inputs with no data dependencies. The original code executes them sequentially, leaving significant room for CUDA stream parallelization.
+
+=== Multi-Level Design of Attention Mechanisms
+
+The attention mechanisms in UnifoLM-WMA constitute the core computational bottleneck of inference. Their design encompasses four distinct levels, each serving a different purpose in the model's spatiotemporal reasoning, and understanding their structure is essential for identifying optimization opportunities.
+
+The first level is *spatial self-attention* within the SpatialTransformer. For a latent frame at resolution $H times W$, the token count is $H times W$ (e.g., $40 times 64 = 2560$ at the highest resolution). Implemented via xformers `memory_efficient_attention`, reducing peak memory from $O(N^2)$ to $O(N)$. Q/K/V use bias-free linear layers with head count = channel\_dim / num\_head\_channels (e.g., 1280/64 = 20 heads).
+
+The second level is *multi-source cross-attention*, the most distinctive design in UnifoLM-WMA. The unified context vector is split into four semantic sources, each with dedicated K/V projection layers:
+
+#figure(
+  table(
+    columns: (2fr, 1fr, 3fr, 2fr),
+    [*Source*], [*Tokens*], [*K/V Projections*], [*Scale*],
+    [Text], [77], [`to_k` / `to_v` (shared base)], [1.0],
+    [Image], [16#sym.times T], [`to_k_ip` / `to_v_ip`], [`image_cross_attention_scale`],
+    [Agent state], [2], [`to_k_as` / `to_v_as`], [`agent_state_cross_attention_scale`],
+    [Agent action], [16], [`to_k_aa` / `to_v_aa`], [`agent_action_cross_attention_scale`],
+  ),
+  caption: [Multi-source cross-attention configuration.],
+)
+
+The Query vector Q is always derived from the video latent features via `to_q`. For each of the four sources, independent attention scores are computed — $"softmax"(Q dot K_i^T \/ sqrt(d)) dot V_i$ — producing four separate attention outputs. These outputs are then combined via weighted summation:
+
+$ "out" = "out"_"text" + alpha_"img" dot "out"_"ip" + alpha_"state" dot "out"_"as" + alpha_"action" dot "out"_"aa" $
+
+In the current configuration, `cross_attention_scale_learnable=False` (fixed scales). This decoupled design adds 8 extra linear layers versus standard single-source cross-attention, creating opportunities for KV fusion optimization.
+
+The third level is *temporal self-attention* within the TemporalTransformer. The input $(B, C, T, H, W)$ is reshaped to $(B times H times W, C, T)$, so each spatial position becomes an independent batch element and T=16 time steps form the token sequence. Supports relative position encoding via a `RelativePosition` module and optional causal masks; current configuration uses bidirectional temporal attention.
+
+The fourth level is *action-latent-image cross-attention* in the `ActionLatentImageCrossAttention` module. Action tokens $(B, "action_dim", "act_proj_dim")$ as Query cross-attend to Video UNet features reshaped to $(B, T times H times W, C)$ as Key/Value. A `BasicTransformerBlock` (depth=1) performs action self-attention then cross-attention to video features, with zero-initialized `proj_out` and residual connection. This mechanism is the key bridge enabling the action head to access the visual world model's internal representations.
+

env.sh

@@ -1,21 +0,0 @@
-# Note: This script should be sourced, not executed
-# Usage: source env.sh
-#
-# If you need render group permissions, run this first:
-#   newgrp render
-# Then source this script:
-#   source env.sh
-
-# Initialize conda
-source /mnt/ASC1637/miniconda3/etc/profile.d/conda.sh
-
-# Activate conda environment
-conda activate unifolm-wma-o
-
-# Set HuggingFace cache directories
-export HF_HOME=/mnt/ASC1637/hf_home
-export HUGGINGFACE_HUB_CACHE=/mnt/ASC1637/hf_home/hub
-
-echo "Environment configured successfully"
-echo "Conda environment: unifolm-wma-o"
-echo "HF_HOME: $HF_HOME"

run.log

@@ -1,150 +0,0 @@
-nohup: ignoring input
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 08:15:49.934949: 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`.
-2026-02-08 08:15:49.937974: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:15:49.969069: 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
-2026-02-08 08:15:49.969100: 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
-2026-02-08 08:15:49.970909: 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
-2026-02-08 08:15:49.979005: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:15:49.979255: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 08:15:50.597743: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> 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%|          | 0/12 [00:00<?, ?it/s]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
-  attn_output = scaled_dot_product_attention(
->>> 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
-
-  8%|▊         | 1/12 [01:37<17:51, 97.37s/it]
- 17%|█▋        | 2/12 [03:14<16:13, 97.31s/it]
- 25%|██▌       | 3/12 [04:51<14:35, 97.26s/it]
- 33%|███▎      | 4/12 [06:29<12:58, 97.25s/it]
- 42%|████▏     | 5/12 [08:06<11:20, 97.24s/it]
- 50%|█████     | 6/12 [09:43<09:43, 97.24s/it]
- 58%|█████▊    | 7/12 [11:20<08:06, 97.27s/it]
- 67%|██████▋   | 8/12 [12:58<06:29, 97.36s/it]
- 75%|███████▌  | 9/12 [14:36<04:52, 97.49s/it]
- 83%|████████▎ | 10/12 [16:13<03:15, 97.52s/it]
- 92%|█████████▏| 11/12 [17:51<01:37, 97.47s/it]
-100%|██████████| 12/12 [19:28<00:00, 97.35s/it]
-100%|██████████| 12/12 [19:28<00:00, 97.35s/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 ...
->>>>>>>>>>>>>>>>>>>>>>>>
->>> 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 ...

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_20260211_173422.log

@@ -0,0 +1,37 @@
+2026-02-11 17:34:29.188470: 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`.
+2026-02-11 17:34:29.238296: 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
+2026-02-11 17:34:29.238342: 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
+2026-02-11 17:34:29.239649: 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
+2026-02-11 17:34:29.247152: 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.
+2026-02-11 17:34:30.172640: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
+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
+>>> 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
+

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

@@ -1,5 +1,4 @@
 import argparse, os, glob
-from contextlib import nullcontext
 import pandas as pd
 import random
 import torch
@@ -10,6 +9,9 @@ import logging
 import einops
 import warnings
 import imageio
+import atexit
+import multiprocessing as mp
+from concurrent.futures import ThreadPoolExecutor
 
 from pytorch_lightning import seed_everything
 from omegaconf import OmegaConf
@@ -17,8 +19,12 @@ 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
@@ -39,68 +45,6 @@ def get_device_from_parameters(module: nn.Module) -> torch.device:
     return next(iter(module.parameters())).device
 
 
-def apply_precision_settings(model: nn.Module, args: argparse.Namespace) -> nn.Module:
-    """Apply precision settings to model components based on command-line arguments.
-
-    Args:
-        model (nn.Module): The model to apply precision settings to.
-        args (argparse.Namespace): Parsed command-line arguments containing precision settings.
-
-    Returns:
-        nn.Module: Model with precision settings applied.
-    """
-    print(f">>> Applying precision settings:")
-    print(f"    - Diffusion dtype: {args.diffusion_dtype}")
-    print(f"    - Projector mode: {args.projector_mode}")
-    print(f"    - Encoder mode: {args.encoder_mode}")
-    print(f"    - VAE dtype: {args.vae_dtype}")
-
-    # 1. Set Diffusion backbone precision
-    if args.diffusion_dtype == "bf16":
-        # Convert diffusion model weights to bf16
-        model.model.to(torch.bfloat16)
-        model.diffusion_autocast_dtype = torch.bfloat16
-        print("    ✓ Diffusion model weights converted to bfloat16")
-    else:
-        model.diffusion_autocast_dtype = None
-        print("    ✓ Diffusion model using fp32")
-
-    # 2. Set Projector precision
-    if args.projector_mode == "bf16_full":
-        model.state_projector.to(torch.bfloat16)
-        model.action_projector.to(torch.bfloat16)
-        model.projector_autocast_dtype = None
-        print("    ✓ Projectors converted to bfloat16")
-    elif args.projector_mode == "autocast":
-        model.projector_autocast_dtype = torch.bfloat16
-        print("    ✓ Projectors will use autocast (weights fp32, compute bf16)")
-    else:
-        model.projector_autocast_dtype = None
-        # fp32 mode: do nothing, keep original precision
-
-    # 3. Set Encoder precision
-    if args.encoder_mode == "bf16_full":
-        model.embedder.to(torch.bfloat16)
-        model.image_proj_model.to(torch.bfloat16)
-        model.encoder_autocast_dtype = None
-        print("    ✓ Encoders converted to bfloat16")
-    elif args.encoder_mode == "autocast":
-        model.encoder_autocast_dtype = torch.bfloat16
-        print("    ✓ Encoders will use autocast (weights fp32, compute bf16)")
-    else:
-        model.encoder_autocast_dtype = None
-        # fp32 mode: do nothing, keep original precision
-
-    # 4. Set VAE precision
-    if args.vae_dtype == "bf16":
-        model.first_stage_model.to(torch.bfloat16)
-        print("    ✓ VAE converted to bfloat16")
-    else:
-        print("    ✓ VAE kept in fp32 for best quality")
-
-    return model
-
-
 def write_video(video_path: str, stacked_frames: list, fps: int) -> None:
     """Save a list of frames to a video file.
 
@@ -213,6 +157,107 @@ 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 _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]
+
+
+def _video_writer_process(q: mp.Queue, filename: str, fps: int):
+    frames = []
+    while True:
+        item = q.get()
+        if item is None:
+            break
+        frames.append(_video_tensor_to_frames(item))
+    if frames:
+        grid = np.concatenate(frames, axis=0)
+        grid = torch.from_numpy(grid[:, :, :, ::-1].copy())  # BGR → RGB
+        torchvision.io.write_video(filename, grid, fps=fps,
+                                   video_codec='h264', options={'crf': '10'})
+
+
 def get_init_frame_path(data_dir: str, sample: dict) -> str:
     """Construct the init_frame path from directory and sample metadata.
 
@@ -325,11 +370,6 @@ def get_latent_z(model, videos: Tensor) -> Tensor:
     """
     b, c, t, h, w = videos.shape
     x = rearrange(videos, 'b c t h w -> (b t) c h w')
-
-    # Auto-detect VAE dtype and convert input
-    vae_dtype = next(model.first_stage_model.parameters()).dtype
-    x = x.to(dtype=vae_dtype)
-
     z = model.encode_first_stage(x)
     z = rearrange(z, '(b t) c h w -> b c t h w', b=b, t=t)
     return z
@@ -395,7 +435,8 @@ 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).
 
@@ -418,10 +459,13 @@ 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.
     """
@@ -431,20 +475,9 @@ def image_guided_synthesis_sim_mode(
 
     fs = torch.tensor([fs] * batch_size, dtype=torch.long, device=model.device)
 
-    # Auto-detect model dtype and convert inputs accordingly
-    model_dtype = next(model.embedder.parameters()).dtype
-
     img = observation['observation.images.top'].permute(0, 2, 1, 3, 4)
-    cond_img = rearrange(img, 'b o c h w -> (b o) c h w')[-1:].to(dtype=model_dtype)
-
-    # Encoder autocast: weights stay fp32, compute in bf16
-    enc_ac_dtype = getattr(model, 'encoder_autocast_dtype', None)
-    if enc_ac_dtype is not None and model.device.type == 'cuda':
-        enc_ctx = torch.autocast('cuda', dtype=enc_ac_dtype)
-    else:
-        enc_ctx = nullcontext()
-
-    with enc_ctx:
+    cond_img = rearrange(img, 'b o c h w -> (b o) c h w')[-1:]
+    with torch.cuda.amp.autocast(dtype=torch.float16):
         cond_img_emb = model.embedder(cond_img)
         cond_img_emb = model.image_proj_model(cond_img_emb)
 
@@ -460,21 +493,11 @@ def image_guided_synthesis_sim_mode(
         prompts = [""] * batch_size
     cond_ins_emb = model.get_learned_conditioning(prompts)
 
-    # Auto-detect projector dtype and convert inputs
-    projector_dtype = next(model.state_projector.parameters()).dtype
-
-    # Projector autocast: weights stay fp32, compute in bf16
-    proj_ac_dtype = getattr(model, 'projector_autocast_dtype', None)
-    if proj_ac_dtype is not None and model.device.type == 'cuda':
-        proj_ctx = torch.autocast('cuda', dtype=proj_ac_dtype)
-    else:
-        proj_ctx = nullcontext()
-
-    with proj_ctx:
-        cond_state_emb = model.state_projector(observation['observation.state'].to(dtype=projector_dtype))
+    with torch.cuda.amp.autocast(dtype=torch.float16):
+        cond_state_emb = model.state_projector(observation['observation.state'])
         cond_state_emb = cond_state_emb + model.agent_state_pos_emb
 
-        cond_action_emb = model.action_projector(observation['action'].to(dtype=projector_dtype))
+        cond_action_emb = model.action_projector(observation['action'])
         cond_action_emb = cond_action_emb + model.agent_action_pos_emb
 
     if not sim_mode:
@@ -497,17 +520,10 @@ def image_guided_synthesis_sim_mode(
     kwargs.update({"unconditional_conditioning_img_nonetext": None})
     cond_mask = None
     cond_z0 = None
-
-    # Setup autocast context for diffusion sampling
-    autocast_dtype = getattr(model, 'diffusion_autocast_dtype', None)
-    if autocast_dtype is not None and model.device.type == 'cuda':
-        autocast_ctx = torch.autocast('cuda', dtype=autocast_dtype)
-    else:
-        autocast_ctx = nullcontext()
-
+    batch_variants = None
+    samples = None
     if ddim_sampler is not None:
-        with autocast_ctx:
-            samples, actions, states, intermedia = ddim_sampler.sample(
+        samples, actions, states, intermedia = ddim_sampler.sample(
             S=ddim_steps,
             conditioning=cond,
             batch_size=batch_size,
@@ -524,11 +540,12 @@ def image_guided_synthesis_sim_mode(
             guidance_rescale=guidance_rescale,
             **kwargs)
 
-        # Reconstruct from latent to pixel space
-        batch_images = model.decode_first_stage(samples)
-        batch_variants = batch_images
+        if decode_video:
+            # Reconstruct from latent to pixel space
+            batch_images = model.decode_first_stage(samples)
+            batch_variants = batch_images
 
-    return batch_variants, actions, states
+    return batch_variants, actions, states, samples
 
 
 def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
@@ -553,37 +570,67 @@ 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 ...')
 
-    # Apply precision settings before moving to GPU
-    model = apply_precision_settings(model, args)
+    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
 
-    # Export precision-converted checkpoint if requested
-    if args.export_precision_ckpt:
-        export_path = args.export_precision_ckpt
-        os.makedirs(os.path.dirname(export_path) or '.', exist_ok=True)
-        torch.save({"state_dict": model.state_dict()}, export_path)
-        print(f">>> Precision-converted checkpoint saved to: {export_path}")
-        return
+        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 ...')
 
-    # Build unnomalizer
+        # 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).")
+
+    # ---- FP16: convert diffusion backbone + conditioning modules ----
+    model.model.to(torch.float16)
+    model.model.diffusion_model.dtype = torch.float16
+    print(">>> Diffusion backbone (model.model) converted to FP16.")
+
+    # Projectors / MLP → FP16
+    model.image_proj_model.half()
+    model.state_projector.half()
+    model.action_projector.half()
+    print(">>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.")
+
+    # Text/image encoders → FP16
+    model.cond_stage_model.half()
+    model.embedder.half()
+    print(">>> Encoders (cond_stage_model, embedder) converted to FP16.")
+
+    # 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")
+
     # Run over data
     assert (args.height % 16 == 0) and (
         args.width % 16
@@ -628,8 +675,13 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
             # For saving environmental changes in world-model
             sample_save_dir = f'{video_save_dir}/wm/{fs}'
             os.makedirs(sample_save_dir, exist_ok=True)
-            # For collecting interaction videos
-            wm_video = []
+            # Writer process for incremental video saving
+            sample_full_video_file = f"{video_save_dir}/../{sample['videoid']}_full_fs{fs}.mp4"
+            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()
             # Initialize observation queues
             cond_obs_queues = {
                 "observation.images.top":
@@ -685,7 +737,7 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
 
                 # Use world-model in policy to generate action
                 print(f'>>> Step {itr}: generating actions ...')
-                pred_videos_0, pred_actions, _ = image_guided_synthesis_sim_mode(
+                pred_videos_0, pred_actions, _, _ = image_guided_synthesis_sim_mode(
                     model,
                     sample['instruction'],
                     observation,
@@ -698,7 +750,8 @@ 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])):
@@ -726,7 +779,7 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
 
                 # Interaction with the world-model
                 print(f'>>> Step {itr}: interacting with world model ...')
-                pred_videos_1, _, pred_states = image_guided_synthesis_sim_mode(
+                pred_videos_1, _, pred_states, wm_samples = image_guided_synthesis_sim_mode(
                     model,
                     "",
                     observation,
@@ -739,12 +792,16 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
                     fs=model_input_fs,
                     text_input=False,
                     timestep_spacing=args.timestep_spacing,
-                    guidance_rescale=args.guidance_rescale)
+                    guidance_rescale=args.guidance_rescale,
+                    decode_video=False)
+
+                # Decode only the last frame for CLIP embedding in next iteration
+                last_frame_pixel = model.decode_first_stage(wm_samples[:, :, -1:, :, :])
 
                 for idx in range(args.exe_steps):
                     observation = {
                         'observation.images.top':
-                        pred_videos_1[0][:, idx:idx + 1].permute(1, 0, 2, 3),
+                        last_frame_pixel[0, :, 0:1].permute(1, 0, 2, 3),
                         'observation.state':
                         torch.zeros_like(pred_states[0][idx:idx + 1]) if
                         args.zero_pred_state else pred_states[0][idx:idx + 1],
@@ -755,42 +812,26 @@ 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
-                sample_tag = f"{args.dataset}-vid{sample['videoid']}-dm-fs-{fs}/itr-{itr}"
-                log_to_tensorboard(writer,
-                                   pred_videos_0,
-                                   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,
-                                   pred_videos_1,
-                                   sample_tag,
-                                   fps=args.save_fps)
-
-                # Save the imagen videos for decision-making
-                sample_video_file = f'{video_save_dir}/dm/{fs}/itr-{itr}.mp4'
-                save_results(pred_videos_0.cpu(),
-                             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(),
-                             sample_video_file,
-                             fps=args.save_fps)
+                # Save the imagen videos for decision-making (async)
+                if pred_videos_0 is not None:
+                    sample_tag = f"{args.dataset}-vid{sample['videoid']}-dm-fs-{fs}/itr-{itr}"
+                    log_to_tensorboard_async(writer,
+                                             pred_videos_0,
+                                             sample_tag,
+                                             fps=args.save_fps)
 
                 print('>' * 24)
-                # Collect the result of world-model interactions
-                wm_video.append(pred_videos_1[:, :, :args.exe_steps].cpu())
+                # Decode segment and send to writer process
+                seg_video = model.decode_first_stage(
+                    wm_samples[:, :, :args.exe_steps]).detach().cpu()
+                write_q.put(seg_video)
 
-            full_video = torch.cat(wm_video, dim=2)
-            sample_tag = f"{args.dataset}-vid{sample['videoid']}-wd-fs-{fs}/full"
-            log_to_tensorboard(writer,
-                               full_video,
-                               sample_tag,
-                               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)
+            # Stop writer process
+            write_q.put(None)
+            writer_proc.join()
+
+    # Wait for all async I/O to complete
+    _flush_io()
 
 
 def get_parser():
@@ -905,39 +946,15 @@ 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=True,
+        help="Speed mode: policy pass only predicts actions, skip policy video decode/log/save.")
     parser.add_argument("--save_fps",
                         type=int,
                         default=8,
                         help="fps for the saving video")
-    parser.add_argument(
-        "--diffusion_dtype",
-        type=str,
-        choices=["fp32", "bf16"],
-        default="bf16",
-        help="Diffusion backbone precision (fp32/bf16)")
-    parser.add_argument(
-        "--projector_mode",
-        type=str,
-        choices=["fp32", "autocast", "bf16_full"],
-        default="bf16_full",
-        help="Projector precision mode (fp32/autocast/bf16_full)")
-    parser.add_argument(
-        "--encoder_mode",
-        type=str,
-        choices=["fp32", "autocast", "bf16_full"],
-        default="bf16_full",
-        help="Encoder precision mode (fp32/autocast/bf16_full)")
-    parser.add_argument(
-        "--vae_dtype",
-        type=str,
-        choices=["fp32", "bf16"],
-        default="fp32",
-        help="VAE precision (fp32/bf16, most affects image quality)")
-    parser.add_argument(
-        "--export_precision_ckpt",
-        type=str,
-        default=None,
-        help="Export precision-converted checkpoint to this path, then exit.")
     return parser
 
 

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/ddpms.py

@@ -988,7 +988,7 @@ class LatentDiffusion(DDPM):
 
     def instantiate_cond_stage(self, config: OmegaConf) -> None:
         """
-        Build the conditioning stage model.
+        Build the conditioning stage model. Frozen models are converted to FP16.
 
         Args:
             config: OmegaConf config describing the conditioning model to instantiate.
@@ -1000,6 +1000,7 @@ class LatentDiffusion(DDPM):
             self.cond_stage_model.train = disabled_train
             for param in self.cond_stage_model.parameters():
                 param.requires_grad = False
+            self.cond_stage_model.half()
         else:
             model = instantiate_from_config(config)
             self.cond_stage_model = model
@@ -1014,17 +1015,18 @@ class LatentDiffusion(DDPM):
         Returns:
             Conditioning embedding as a tensor (shape depends on cond model).
         """
-        if self.cond_stage_forward is None:
-            if hasattr(self.cond_stage_model, 'encode') and callable(
-                    self.cond_stage_model.encode):
-                c = self.cond_stage_model.encode(c)
-                if isinstance(c, DiagonalGaussianDistribution):
-                    c = c.mode()
+        with torch.cuda.amp.autocast(dtype=torch.float16):
+            if self.cond_stage_forward is None:
+                if hasattr(self.cond_stage_model, 'encode') and callable(
+                        self.cond_stage_model.encode):
+                    c = self.cond_stage_model.encode(c)
+                    if isinstance(c, DiagonalGaussianDistribution):
+                        c = c.mode()
+                else:
+                    c = self.cond_stage_model(c)
             else:
-                c = self.cond_stage_model(c)
-        else:
-            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
-            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
+                assert hasattr(self.cond_stage_model, self.cond_stage_forward)
+                c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
         return c
 
     def get_first_stage_encoding(
@@ -1105,10 +1107,6 @@ class LatentDiffusion(DDPM):
         else:
             reshape_back = False
 
-        # Align input dtype with VAE weights (e.g. fp32 samples → bf16 VAE)
-        vae_dtype = next(self.first_stage_model.parameters()).dtype
-        z = z.to(dtype=vae_dtype)
-
         if not self.perframe_ae:
             z = 1. / self.scale_factor * z
             results = self.first_stage_model.decode(z, **kwargs)
@@ -1961,6 +1959,7 @@ class LatentVisualDiffusion(LatentDiffusion):
             self.image_proj_model.train = disabled_train
             for param in self.image_proj_model.parameters():
                 param.requires_grad = False
+            self.image_proj_model.half()
 
     def _init_embedder(self, config: OmegaConf, freeze: bool = True) -> None:
         """
@@ -1976,6 +1975,7 @@ class LatentVisualDiffusion(LatentDiffusion):
             self.embedder.train = disabled_train
             for param in self.embedder.parameters():
                 param.requires_grad = False
+            self.embedder.half()
 
     def init_normalizers(self, normalize_config: OmegaConf,
                          dataset_stats: Mapping[str, Any]) -> None:
@@ -2179,8 +2179,9 @@ class LatentVisualDiffusion(LatentDiffusion):
             (random_num < 3 * self.uncond_prob).float(), "n -> n 1 1 1")
 
         cond_img = input_mask * img
-        cond_img_emb = self.embedder(cond_img)
-        cond_img_emb = self.image_proj_model(cond_img_emb)
+        with torch.cuda.amp.autocast(dtype=torch.float16):
+            cond_img_emb = self.embedder(cond_img)
+            cond_img_emb = self.image_proj_model(cond_img_emb)
 
         if self.model.conditioning_key == 'hybrid':
             if self.interp_mode:
@@ -2195,11 +2196,12 @@ class LatentVisualDiffusion(LatentDiffusion):
                                       repeat=z.shape[2])
             cond["c_concat"] = [img_cat_cond]
 
-        cond_action = self.action_projector(action)
-        cond_action_emb = self.agent_action_pos_emb + cond_action
-        # Get conditioning states
-        cond_state = self.state_projector(obs_state)
-        cond_state_emb = self.agent_state_pos_emb + cond_state
+        with torch.cuda.amp.autocast(dtype=torch.float16):
+            cond_action = self.action_projector(action)
+            cond_action_emb = self.agent_action_pos_emb + cond_action
+            # Get conditioning states
+            cond_state = self.state_projector(obs_state)
+            cond_state_emb = self.agent_state_pos_emb + cond_state
 
         if self.decision_making_only:
             is_sim_mode = False
@@ -2461,6 +2463,17 @@ class DiffusionWrapper(pl.LightningModule):
         Returns:
             Output from the inner diffusion model (tensor or tuple, depending on the model).
         """
+        with torch.cuda.amp.autocast(dtype=torch.float16):
+            return self._forward_impl(x, x_action, x_state, t,
+                                      c_concat, c_crossattn, c_crossattn_action,
+                                      c_adm, s, mask, **kwargs)
+
+    def _forward_impl(
+        self,
+        x, x_action, x_state, t,
+        c_concat=None, c_crossattn=None, c_crossattn_action=None,
+        c_adm=None, s=None, mask=None, **kwargs,
+    ):
         if self.conditioning_key is None:
             out = self.diffusion_model(x, t)
         elif self.conditioning_key == 'concat':

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)
-            global_cond = rearrange(global_cond,
-                                    '(b t) d -> b 1 (t d)',
-                                    b=B,
-                                    t=self.n_obs_steps)
-            global_cond = repeat(global_cond,
-                                 'b c d -> b (repeat c) d',
-                                 repeat=T)
+            _gc_key = (cond['image'].data_ptr(), cond['agent_pos'].data_ptr())
+            if self._global_cond_cache_enabled and _gc_key in self._global_cond_cache:
+                global_cond = self._global_cond_cache[_gc_key]
+            else:
+                global_cond = self.obs_encoder(cond)
+                global_cond = rearrange(global_cond,
+                                        '(b t) d -> b 1 (t d)',
+                                        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)
-        self.register_buffer('ddim_alphas', ddim_alphas)
-        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        # Ensure tensors are on correct device for efficient indexing
+        self.register_buffer('ddim_sigmas', to_torch(torch.as_tensor(ddim_sigmas)))
+        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):
-            index = total_steps - i - 1
-            ts = torch.full((b, ), step, device=device, dtype=torch.long)
+        ts = torch.empty((b, ), device=device, dtype=torch.long)
+        enable_cross_attn_kv_cache(self.model)
+        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)
-            if mask is not None:
-                assert x0 is not None
-                if clean_cond:
-                    img_orig = x0
-                else:
-                    img_orig = self.model.q_sample(x0, ts)
-                img = img_orig * mask + (1. - mask) * img
+                # Use mask to blend noised original latent (img_orig) & new sampled latent (img)
+                if mask is not None:
+                    assert x0 is not None
+                    if clean_cond:
+                        img_orig = x0
+                    else:
+                        img_orig = self.model.q_sample(x0, ts)
+                    img = img_orig * mask + (1. - mask) * img
 
-            outs = self.p_sample_ddim(
-                img,
-                action,
-                state,
-                cond,
-                ts,
-                index=index,
-                use_original_steps=ddim_use_original_steps,
-                quantize_denoised=quantize_denoised,
-                temperature=temperature,
-                noise_dropout=noise_dropout,
-                score_corrector=score_corrector,
-                corrector_kwargs=corrector_kwargs,
-                unconditional_guidance_scale=unconditional_guidance_scale,
-                unconditional_conditioning=unconditional_conditioning,
-                mask=mask,
-                x0=x0,
-                fs=fs,
-                guidance_rescale=guidance_rescale,
-                **kwargs)
+                outs = self.p_sample_ddim(
+                    img,
+                    action,
+                    state,
+                    cond,
+                    ts,
+                    index=index,
+                    use_original_steps=ddim_use_original_steps,
+                    quantize_denoised=quantize_denoised,
+                    temperature=temperature,
+                    noise_dropout=noise_dropout,
+                    score_corrector=score_corrector,
+                    corrector_kwargs=corrector_kwargs,
+                    unconditional_guidance_scale=unconditional_guidance_scale,
+                    unconditional_conditioning=unconditional_conditioning,
+                    mask=mask,
+                    x0=x0,
+                    fs=fs,
+                    guidance_rescale=guidance_rescale,
+                    **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(
-                model_output_action,
-                step,
-                action,
-                generator=None,
-            ).prev_sample
-            state = dp_ddim_scheduler_state.step(
-                model_output_state,
-                step,
-                state,
-                generator=None,
-            ).prev_sample
+                action = dp_ddim_scheduler_action.step(
+                    model_output_action,
+                    step,
+                    action,
+                    generator=None,
+                ).prev_sample
+                state = dp_ddim_scheduler_state.step(
+                    model_output_state,
+                    step,
+                    state,
+                    generator=None,
+                ).prev_sample
 
-            if callback: callback(i)
-            if img_callback: img_callback(pred_x0, i)
+                if callback: callback(i)
+                if img_callback: img_callback(pred_x0, i)
 
-            if index % log_every_t == 0 or index == total_steps - 1:
-                intermediates['x_inter'].append(img)
-                intermediates['pred_x0'].append(pred_x0)
-                intermediates['x_inter_action'].append(action)
-                intermediates['x_inter_state'].append(state)
+                if index % log_every_t == 0 or index == total_steps - 1:
+                    intermediates['x_inter'].append(img)
+                    intermediates['pred_x0'].append(pred_x0)
+                    intermediates['x_inter_action'].append(action)
+                    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:
-            size = (b, 1, 1, 1, 1)
-        else:
-            size = (b, 1, 1, 1)
-
-        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)
+        # Use 0-d tensors directly (already on device); broadcasting handles shape
+        a_t = alphas[index]
+        a_prev = alphas_prev[index]
+        sigma_t = sigmas[index]
+        sqrt_one_minus_at = sqrt_one_minus_alphas[index]
 
         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,
-                                 self.ddim_scale_arr[index],
-                                 device=device)
-            prev_scale_t = torch.full(size,
-                                      self.ddim_scale_arr_prev[index],
-                                      device=device)
+            scale_t = self.ddim_scale_arr[index]
+            prev_scale_t = self.ddim_scale_arr_prev[index]
             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
@@ -125,7 +150,7 @@ class CrossAttention(nn.Module):
         context = default(context, x)
 
         if self.image_cross_attention and not spatial_self_attn:
-            # assert 1 > 2, ">>> ERROR: should setup xformers and use efficient_forward ..."
+            assert 1 > 2, ">>> ERROR: should setup xformers and use efficient_forward ..."
             context_agent_state = context[:, :self.agent_state_context_len, :]
             context_agent_action = context[:,
                                            self.agent_state_context_len:self.
@@ -140,19 +165,28 @@ class CrossAttention(nn.Module):
                                     self.agent_action_context_len +
                                     self.text_context_len:, :]
 
-            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_aa = self.to_k_aa(context_agent_action)
-            v_aa = self.to_v_aa(context_agent_action)
+            if self._kv_fused:
+                k, v = self.to_kv(context_ins).chunk(2, dim=-1)
+                k_ip, v_ip = self.to_kv_ip(context_image).chunk(2, dim=-1)
+                k_as, v_as = self.to_kv_as(context_agent_state).chunk(2, dim=-1)
+                k_aa, v_aa = self.to_kv_aa(context_agent_action).chunk(2, dim=-1)
+            else:
+                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_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)
-            v = self.to_v(context)
+            if self._kv_fused:
+                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)
-                v = self.to_v(context)
-                k_ip = self.to_k_ip(context_image)
-                v_ip = self.to_v_ip(context_image)
+                if self._kv_fused:
+                    k, v = self.to_kv(context).chunk(2, dim=-1)
+                    k_ip, v_ip = self.to_kv_ip(context_image).chunk(2, dim=-1)
+                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)
-                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)
+                if self._kv_fused:
+                    k, v = self.to_kv(context_ins).chunk(2, dim=-1)
+                    k_ip, v_ip = self.to_kv_ip(context_image).chunk(2, dim=-1)
+                    k_as, v_as = self.to_kv_as(context_agent_state).chunk(2, dim=-1)
+                else:
+                    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)
-                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)
+                if self._kv_fused:
+                    k, v = self.to_kv(context_ins).chunk(2, dim=-1)
+                    k_ip, v_ip = self.to_kv_ip(context_image).chunk(2, dim=-1)
+                    k_as, v_as = self.to_kv_as(context_agent_state).chunk(2, dim=-1)
+                    k_aa, v_aa = self.to_kv_aa(context_agent_action).chunk(2, dim=-1)
+                else:
+                    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_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],
-                                                      q.shape[1],
-                                                      k_aa.shape[1],
-                                                      block_size=16).to(k_aa.device)
+                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))
+                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)
-            v = self.to_v(context)
-
-        b, _, _ = q.shape
-        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()
+            if self._kv_fused:
+                k, v = self.to_kv(context).chunk(2, dim=-1)
+            else:
+                k = self.to_k(context)
+                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
-        col_indices = torch.arange(l2)
+        start_positions = ((torch.arange(b, device=target_device) % block_size) + 1) * num_token
+        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 +685,21 @@ 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 = {}
+        # 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 forward(self,
                 x: Tensor,
                 x_action: Tensor,
@@ -720,58 +735,64 @@ class WMAModel(nn.Module):
                                    repeat_only=False).type(x.dtype)
         emb = self.time_embed(t_emb)
 
-        bt, l_context, _ = context.shape
-        if self.base_model_gen_only:
-            assert l_context == 77 + self.n_obs_steps * 16, ">>> ERROR Context dim 1 ..."  ## NOTE HANDCODE
+        _ctx_key = context.data_ptr()
+        if self._ctx_cache_enabled and _ctx_key in self._ctx_cache:
+            context = self._ctx_cache[_ctx_key]
         else:
-            if l_context == self.n_obs_steps + 77 + t * 16:
-                context_agent_state = context[:, :self.n_obs_steps]
-                context_text = context[:, self.n_obs_steps:self.n_obs_steps +
-                                       77, :]
-                context_img = context[:, self.n_obs_steps + 77:, :]
-                context_agent_state = context_agent_state.repeat_interleave(
-                    repeats=t, dim=0)
-                context_text = context_text.repeat_interleave(repeats=t, dim=0)
-                context_img = rearrange(context_img,
-                                        'b (t l) c -> (b t) l c',
-                                        t=t)
-                context = torch.cat(
-                    [context_agent_state, context_text, context_img], dim=1)
-            elif l_context == self.n_obs_steps + 16 + 77 + t * 16:
-                context_agent_state = context[:, :self.n_obs_steps]
-                context_agent_action = context[:, self.
-                                               n_obs_steps:self.n_obs_steps +
-                                               16, :]
-                context_agent_action = rearrange(
-                    context_agent_action.unsqueeze(2), 'b t l d -> (b t) l d')
-                context_agent_action = self.action_token_projector(
-                    context_agent_action)
-                context_agent_action = rearrange(context_agent_action,
-                                                 '(b o) l d -> b o l d',
-                                                 o=t)
-                context_agent_action = rearrange(context_agent_action,
-                                                 'b o (t l) d -> b o t 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')
+            bt, l_context, _ = context.shape
+            if self.base_model_gen_only:
+                assert l_context == 77 + self.n_obs_steps * 16, ">>> ERROR Context dim 1 ..."  ## NOTE HANDCODE
+            else:
+                if l_context == self.n_obs_steps + 77 + t * 16:
+                    context_agent_state = context[:, :self.n_obs_steps]
+                    context_text = context[:, self.n_obs_steps:self.n_obs_steps +
+                                           77, :]
+                    context_img = context[:, self.n_obs_steps + 77:, :]
+                    context_agent_state = context_agent_state.repeat_interleave(
+                        repeats=t, dim=0)
+                    context_text = context_text.repeat_interleave(repeats=t, dim=0)
+                    context_img = rearrange(context_img,
+                                            'b (t l) c -> (b t) l c',
+                                            t=t)
+                    context = torch.cat(
+                        [context_agent_state, context_text, context_img], dim=1)
+                elif l_context == self.n_obs_steps + 16 + 77 + t * 16:
+                    context_agent_state = context[:, :self.n_obs_steps]
+                    context_agent_action = context[:, self.
+                                                   n_obs_steps:self.n_obs_steps +
+                                                   16, :]
+                    context_agent_action = rearrange(
+                        context_agent_action.unsqueeze(2), 'b t l d -> (b t) l d')
+                    context_agent_action = self.action_token_projector(
+                        context_agent_action)
+                    context_agent_action = rearrange(context_agent_action,
+                                                     '(b o) l d -> b o l d',
+                                                     o=t)
+                    context_agent_action = rearrange(context_agent_action,
+                                                     'b o (t l) d -> b o t 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 +
-                                       16:self.n_obs_steps + 16 + 77, :]
-                context_text = context_text.repeat_interleave(repeats=t, dim=0)
+                    context_text = context[:, self.n_obs_steps +
+                                           16:self.n_obs_steps + 16 + 77, :]
+                    context_text = context_text.repeat_interleave(repeats=t, dim=0)
 
-                context_img = context[:, self.n_obs_steps + 16 + 77:, :]
-                context_img = rearrange(context_img,
-                                        'b (t l) c -> (b t) l c',
-                                        t=t)
-                context_agent_state = context_agent_state.repeat_interleave(
-                    repeats=t, dim=0)
-                context = torch.cat([
-                    context_agent_state, context_agent_action, context_text,
-                    context_img
-                ],
-                                    dim=1)
+                    context_img = context[:, self.n_obs_steps + 16 + 77:, :]
+                    context_img = rearrange(context_img,
+                                            'b (t l) c -> (b t) l c',
+                                            t=t)
+                    context_agent_state = context_agent_state.repeat_interleave(
+                        repeats=t, dim=0)
+                    context = torch.cat([
+                        context_agent_state, context_agent_action, context_text,
+                        context_img
+                    ],
+                                        dim=1)
+            if self._ctx_cache_enabled:
+                self._ctx_cache[_ctx_key] = context
 
         emb = emb.repeat_interleave(repeats=t, dim=0)
 
@@ -832,17 +853,45 @@ class WMAModel(nn.Module):
 
         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
-            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)
+            torch.cuda.current_stream().wait_stream(s_stream)
         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 = {}

unitree_g1_pack_camera/case1/output.log

@@ -1,32 +1,16 @@
-2026-02-08 05:20:49.828675: 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`.
-2026-02-08 05:20:49.831563: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 05:20:49.861366: 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
-2026-02-08 05:20:49.861402: 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
-2026-02-08 05:20:49.862974: 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
-2026-02-08 05:20:49.870402: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 05:20:49.870647: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 05:20:50.486843: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+2026-02-19 18:55:32.160020: 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`.
+2026-02-19 18:55:32.207538: 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
+2026-02-19 18:55:32.207581: 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
+2026-02-19 18:55:32.208613: 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
+2026-02-19 18:55:32.215249: 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.
+2026-02-19 18:55:33.121466: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
 Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -44,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/11 [00:00<?, ?it/s]
   9%|▉         | 1/11 [00:23<03:56, 23.63s/it]
->>> 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
  18%|█▊        | 2/11 [00:46<03:29, 23.24s/it]
  27%|██▋       | 3/11 [01:09<03:05, 23.25s/it]
  36%|███▋      | 4/11 [01:33<02:43, 23.31s/it]
@@ -139,6 +69,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 6: generating actions ...
 >>> Step 6: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 7: generating actions ...
->>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_g1_pack_camera/case1/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_g1_pack_camera/case1/unitree_g1_pack_camera_case1.mp4",
-    "pred_video": "unitree_g1_pack_camera/case1/output/inference/unitree_g1_pack_camera_case1_amd.mp4",
-    "psnr": 16.415668383379177
+    "pred_video": "unitree_g1_pack_camera/case1/output/inference/0_full_fs6.mp4",
+    "psnr": 32.34126103448495
 }

unitree_g1_pack_camera/case1/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_g1_pack_camera"
         --n_iter 11 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_g1_pack_camera/case2/output.log

@@ -1,32 +1,16 @@
-2026-02-08 05:06:45.806187: 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`.
-2026-02-08 05:06:45.809295: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 05:06:45.840950: 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
-2026-02-08 05:06:45.840981: 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
-2026-02-08 05:06:45.842814: 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
-2026-02-08 05:06:45.851049: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 05:06:45.851316: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 05:06:47.225477: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+2026-02-19 19:00:05.944067: 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`.
+2026-02-19 19:00:05.991354: 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
+2026-02-19 19:00:05.991392: 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
+2026-02-19 19:00:05.992414: 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
+2026-02-19 19:00:05.999050: 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.
+2026-02-19 19:00:06.916175: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -44,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/11 [00:00<?, ?it/s]
   9%|▉         | 1/11 [00:24<04:00, 24.04s/it]
->>> 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
  18%|█▊        | 2/11 [00:47<03:31, 23.55s/it]
  27%|██▋       | 3/11 [01:10<03:07, 23.43s/it]
  36%|███▋      | 4/11 [01:33<02:43, 23.42s/it]
@@ -139,6 +69,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 6: generating actions ...
 >>> Step 6: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 7: generating actions ...
->>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_g1_pack_camera/case2/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_g1_pack_camera/case2/unitree_g1_pack_camera_case2.mp4",
-    "pred_video": "unitree_g1_pack_camera/case2/output/inference/unitree_g1_pack_camera_case2_amd.mp4",
-    "psnr": 19.515250190529375
+    "pred_video": "unitree_g1_pack_camera/case2/output/inference/50_full_fs6.mp4",
+    "psnr": 37.49178506869336
 }

unitree_g1_pack_camera/case2/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_g1_pack_camera"
         --n_iter 11 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_g1_pack_camera/case3/output.log

@@ -1,32 +1,16 @@
-2026-02-08 05:08:32.803904: 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`.
-2026-02-08 05:08:32.807010: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 05:08:32.837936: 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
-2026-02-08 05:08:32.837978: 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
-2026-02-08 05:08:32.839785: 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
-2026-02-08 05:08:32.847835: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 05:08:32.848223: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 05:08:34.120114: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+2026-02-19 19:04:41.036634: 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`.
+2026-02-19 19:04:41.084414: 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
+2026-02-19 19:04:41.084452: 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
+2026-02-19 19:04:41.085481: 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
+2026-02-19 19:04:41.092287: 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.
+2026-02-19 19:04:42.000614: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -44,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/11 [00:00<?, ?it/s]
   9%|▉         | 1/11 [00:24<04:01, 24.19s/it]
->>> 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
  18%|█▊        | 2/11 [00:47<03:32, 23.64s/it]
  27%|██▋       | 3/11 [01:10<03:08, 23.50s/it]
  36%|███▋      | 4/11 [01:34<02:44, 23.47s/it]
@@ -139,6 +69,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 6: generating actions ...
 >>> Step 6: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 7: generating actions ...
->>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_g1_pack_camera/case3/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_g1_pack_camera/case3/unitree_g1_pack_camera_case3.mp4",
-    "pred_video": "unitree_g1_pack_camera/case3/output/inference/unitree_g1_pack_camera_case3_amd.mp4",
-    "psnr": 19.429578160315536
+    "pred_video": "unitree_g1_pack_camera/case3/output/inference/100_full_fs6.mp4",
+    "psnr": 29.88155122131729
 }

unitree_g1_pack_camera/case3/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_g1_pack_camera"
         --n_iter 11 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_g1_pack_camera/case4/output.log

@@ -1,32 +1,16 @@
-2026-02-08 05:29:19.728303: 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`.
-2026-02-08 05:29:19.731620: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 05:29:19.761276: 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
-2026-02-08 05:29:19.761301: 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
-2026-02-08 05:29:19.762880: 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
-2026-02-08 05:29:19.770578: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 05:29:19.771072: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 05:29:21.043661: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+2026-02-19 19:09:16.122268: 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`.
+2026-02-19 19:09:16.170290: 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
+2026-02-19 19:09:16.170331: 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
+2026-02-19 19:09:16.171349: 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
+2026-02-19 19:09:16.177993: 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.
+2026-02-19 19:09:17.087425: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
 Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -44,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/11 [00:00<?, ?it/s]
   9%|▉         | 1/11 [00:24<04:01, 24.17s/it]
->>> 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
  18%|█▊        | 2/11 [00:47<03:32, 23.62s/it]
  27%|██▋       | 3/11 [01:10<03:07, 23.49s/it]
  36%|███▋      | 4/11 [01:34<02:44, 23.46s/it]
@@ -139,6 +69,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 6: generating actions ...
 >>> Step 6: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 7: generating actions ...
->>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_g1_pack_camera/case4/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_g1_pack_camera/case4/unitree_g1_pack_camera_case4.mp4",
-    "pred_video": "unitree_g1_pack_camera/case4/output/inference/unitree_g1_pack_camera_case4_amd.mp4",
-    "psnr": 17.80386833747375
+    "pred_video": "unitree_g1_pack_camera/case4/output/inference/200_full_fs6.mp4",
+    "psnr": 35.62512454155058
 }

unitree_g1_pack_camera/case4/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_g1_pack_camera"
         --n_iter 11 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_cleanup_pencils/case1/output.log

@@ -1,41 +1,16 @@
-2026-02-08 12:22:55.885867: 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`.
-2026-02-08 12:22:55.890510: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 12:22:55.938683: 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
-2026-02-08 12:22:55.938759: 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
-2026-02-08 12:22:55.941091: 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
-2026-02-08 12:22:55.952450: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 12:22:55.952933: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 12:22:56.593653: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:149: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
->>> Applying precision settings:
-    - Diffusion dtype: bf16
-    - Projector mode: bf16_full
-    - Encoder mode: bf16_full
-    - VAE dtype: bf16
-    ✓ Diffusion model weights converted to bfloat16
-    ✓ Projectors converted to bfloat16
-    ✓ Encoders converted to bfloat16
-    ✓ VAE converted to bfloat16
+2026-02-19 19:13:51.554194: 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`.
+2026-02-19 19:13:51.601580: 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
+2026-02-19 19:13:51.601622: 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
+2026-02-19 19:13:51.602646: 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
+2026-02-19 19:13:51.609297: 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.
+2026-02-19 19:13:52.517676: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -53,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/8 [00:00<?, ?it/s]
  12%|█▎        | 1/8 [00:24<02:49, 24.16s/it]
->>> 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
  25%|██▌       | 2/8 [00:47<02:22, 23.67s/it]
  38%|███▊      | 3/8 [01:10<01:57, 23.55s/it]
  50%|█████     | 4/8 [01:34<01:34, 23.51s/it]
@@ -139,6 +60,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 4: generating actions ...
 >>> Step 4: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 5: generating actions ...
->>> Step 5: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...
+>>> Step 5: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_z1_dual_arm_cleanup_pencils/case1/psnr_result.json

@@ -1,5 +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/unitree_z1_dual_arm_cleanup_pencils_case1_amd.mp4",
-    "psnr": 19.586376345676264
+    "pred_video": "unitree_z1_dual_arm_cleanup_pencils/case1/output/inference/0_full_fs4.mp4",
+    "psnr": 38.269577028444445
 }

unitree_z1_dual_arm_cleanup_pencils/case1/psnr_result1.json

@@ -1,5 +0,0 @@
-{
-    "gt_video": "/mnt/ASC1637/unifolm-world-model-action/unitree_z1_dual_arm_cleanup_pencils/case1/output/inference/unitree_z1_dual_arm_cleanup_pencils_case1_amd.mp4",
-    "pred_video": "/mnt/ASC1637/unifolm-world-model-action/unitree_z1_dual_arm_cleanup_pencils/case1/output/inference/0_full_fs4.mp4",
-    "psnr": 30.44844270035179
-}

unitree_z1_dual_arm_cleanup_pencils/case1/run_world_model_interaction.sh

@@ -4,7 +4,7 @@ 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_mix_bf16.ckpt \
+        --ckpt_path ckpts/unifolm_wma_dual.ckpt \
         --config configs/inference/world_model_interaction.yaml \
         --savedir "${res_dir}/output" \
         --bs 1 --height 320 --width 512 \
@@ -21,5 +21,5 @@ dataset="unitree_z1_dual_arm_cleanup_pencils"
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
         --perframe_ae \
-        --vae_dtype bf16
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_cleanup_pencils/case2/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 06:59:34.465946: 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`.
-2026-02-08 06:59:34.469367: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 06:59:34.500805: 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
-2026-02-08 06:59:34.500837: 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
-2026-02-08 06:59:34.502917: 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
-2026-02-08 06:59:34.511434: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 06:59:34.511678: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 06:59:35.478194: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+2026-02-19 19:17:16.282875: 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`.
+2026-02-19 19:17:16.330519: 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
+2026-02-19 19:17:16.330561: 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
+2026-02-19 19:17:16.331631: 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
+2026-02-19 19:17:16.338413: 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.
+2026-02-19 19:17:17.250653: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
 Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/8 [00:00<?, ?it/s]
  12%|█▎        | 1/8 [00:24<02:48, 24.06s/it]
->>> 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
  25%|██▌       | 2/8 [00:47<02:21, 23.61s/it]
  38%|███▊      | 3/8 [01:10<01:57, 23.47s/it]
  50%|█████     | 4/8 [01:34<01:33, 23.44s/it]
@@ -132,6 +60,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 4: generating actions ...
 >>> Step 4: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 5: generating actions ...
->>> Step 5: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...
+>>> Step 5: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_z1_dual_arm_cleanup_pencils/case2/psnr_result.json

@@ -1,5 +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/unitree_z1_dual_arm_cleanup_pencils_case2_amd.mp4",
-    "psnr": 20.484298972158296
+    "pred_video": "unitree_z1_dual_arm_cleanup_pencils/case2/output/inference/50_full_fs4.mp4",
+    "psnr": 44.50028075962896
 }

unitree_z1_dual_arm_cleanup_pencils/case2/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_cleanup_pencils"
         --n_iter 8 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_cleanup_pencils/case3/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 07:18:52.629976: 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`.
-2026-02-08 07:18:52.633025: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:18:52.663985: 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
-2026-02-08 07:18:52.664018: 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
-2026-02-08 07:18:52.665837: 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
-2026-02-08 07:18:52.673889: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:18:52.674218: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 07:18:53.298338: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+2026-02-19 19:20:40.444703: 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`.
+2026-02-19 19:20:40.492237: 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
+2026-02-19 19:20:40.492278: 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
+2026-02-19 19:20:40.493360: 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
+2026-02-19 19:20:40.500130: 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.
+2026-02-19 19:20:41.414718: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/8 [00:00<?, ?it/s]
  12%|█▎        | 1/8 [00:24<02:48, 24.06s/it]
->>> 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
  25%|██▌       | 2/8 [00:47<02:21, 23.58s/it]
  38%|███▊      | 3/8 [01:10<01:57, 23.45s/it]
  50%|█████     | 4/8 [01:33<01:33, 23.41s/it]
@@ -132,6 +60,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 4: generating actions ...
 >>> Step 4: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 5: generating actions ...
->>> Step 5: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...
+>>> Step 5: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_z1_dual_arm_cleanup_pencils/case3/psnr_result.json

@@ -1,5 +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/unitree_z1_dual_arm_cleanup_pencils_case3_amd.mp4",
-    "psnr": 21.20205061239349
+    "pred_video": "unitree_z1_dual_arm_cleanup_pencils/case3/output/inference/100_full_fs4.mp4",
+    "psnr": 32.29959078097713
 }

unitree_z1_dual_arm_cleanup_pencils/case3/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_cleanup_pencils"
         --n_iter 8 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_cleanup_pencils/case4/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 07:22:15.333099: 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`.
-2026-02-08 07:22:15.336215: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:22:15.366489: 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
-2026-02-08 07:22:15.366522: 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
-2026-02-08 07:22:15.368294: 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
-2026-02-08 07:22:15.376202: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:22:15.376444: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 07:22:15.995383: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+2026-02-19 19:24:05.230366: 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`.
+2026-02-19 19:24:05.278058: 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
+2026-02-19 19:24:05.278100: 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
+2026-02-19 19:24:05.279133: 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
+2026-02-19 19:24:05.285789: 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.
+2026-02-19 19:24:06.199101: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/8 [00:00<?, ?it/s]
  12%|█▎        | 1/8 [00:24<02:48, 24.06s/it]
->>> 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
  25%|██▌       | 2/8 [00:47<02:21, 23.56s/it]
  38%|███▊      | 3/8 [01:10<01:57, 23.45s/it]
  50%|█████     | 4/8 [01:33<01:33, 23.43s/it]
@@ -132,6 +60,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 4: generating actions ...
 >>> Step 4: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 5: generating actions ...
->>> Step 5: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...
+>>> Step 5: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_z1_dual_arm_cleanup_pencils/case4/psnr_result.json

@@ -1,5 +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/unitree_z1_dual_arm_cleanup_pencils_case4_amd.mp4",
-    "psnr": 21.130122583788612
+    "pred_video": "unitree_z1_dual_arm_cleanup_pencils/case4/output/inference/200_full_fs4.mp4",
+    "psnr": 45.051241961122535
 }

unitree_z1_dual_arm_cleanup_pencils/case4/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_cleanup_pencils"
         --n_iter 8 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox/case1/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 07:24:40.357099: 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`.
-2026-02-08 07:24:40.360365: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:24:40.391744: 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
-2026-02-08 07:24:40.391772: 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
-2026-02-08 07:24:40.393608: 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
-2026-02-08 07:24:40.401837: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:24:40.402077: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 07:24:41.022382: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+2026-02-19 19:27:29.317502: 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`.
+2026-02-19 19:27:29.365030: 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
+2026-02-19 19:27:29.365079: 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
+2026-02-19 19:27:29.366111: 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
+2026-02-19 19:27:29.372733: 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.
+2026-02-19 19:27:30.291220: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
 Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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%|          | 0/7 [00:00<?, ?it/s]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/7 [00:00<?, ?it/s]
  14%|█▍        | 1/7 [00:24<02:24, 24.09s/it]
->>> 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
  29%|██▊       | 2/7 [00:47<01:57, 23.59s/it]
  43%|████▎     | 3/7 [01:10<01:33, 23.46s/it]
  57%|█████▋    | 4/7 [01:33<01:10, 23.42s/it]
@@ -129,6 +57,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 3: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
 >>> Step 4: generating actions ...
->>> Step 4: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 5: generating actions ...
+>>> Step 4: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...

unitree_z1_dual_arm_stackbox/case1/psnr_result.json

@@ -1,5 +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/unitree_z1_dual_arm_stackbox_case1_amd.mp4",
-    "psnr": 21.258130518117493
+    "pred_video": "unitree_z1_dual_arm_stackbox/case1/output/inference/5_full_fs4.mp4",
+    "psnr": 42.717688631296596
 }

unitree_z1_dual_arm_stackbox/case1/run_world_model_interaction.sh

@@ -2,7 +2,7 @@ res_dir="unitree_z1_dual_arm_stackbox/case1"
 dataset="unitree_z1_dual_arm_stackbox"
 
 {
-    time CUDA_VISIBLE_DEVICES=7 python3 scripts/evaluation/world_model_interaction.py \
+    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 \
@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_stackbox"
         --n_iter 7 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox/case2/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 07:25:18.653033: 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`.
-2026-02-08 07:25:18.656060: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:25:18.687077: 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
-2026-02-08 07:25:18.687119: 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
-2026-02-08 07:25:18.688915: 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
-2026-02-08 07:25:18.697008: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:25:18.697255: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 07:25:19.338303: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+2026-02-19 19:30:30.058862: 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`.
+2026-02-19 19:30:30.106200: 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
+2026-02-19 19:30:30.106243: 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
+2026-02-19 19:30:30.107276: 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
+2026-02-19 19:30:30.113917: 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.
+2026-02-19 19:30:31.026240: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
 Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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%|          | 0/7 [00:00<?, ?it/s]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/7 [00:00<?, ?it/s]
  14%|█▍        | 1/7 [00:24<02:24, 24.09s/it]
->>> 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
  29%|██▊       | 2/7 [00:47<01:58, 23.60s/it]
  43%|████▎     | 3/7 [01:10<01:33, 23.48s/it]
  57%|█████▋    | 4/7 [01:34<01:10, 23.43s/it]
@@ -129,6 +57,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 3: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
 >>> Step 4: generating actions ...
->>> Step 4: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 5: generating actions ...
+>>> Step 4: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...

unitree_z1_dual_arm_stackbox/case2/psnr_result.json

@@ -1,5 +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/unitree_z1_dual_arm_stackbox_case2_amd.mp4",
-    "psnr": 23.878153424077645
+    "pred_video": "unitree_z1_dual_arm_stackbox/case2/output/inference/15_full_fs4.mp4",
+    "psnr": 44.90750363879194
 }

unitree_z1_dual_arm_stackbox/case2/run_world_model_interaction.sh

@@ -2,7 +2,7 @@ res_dir="unitree_z1_dual_arm_stackbox/case2"
 dataset="unitree_z1_dual_arm_stackbox"
 
 {
-    time CUDA_VISIBLE_DEVICES=6 python3 scripts/evaluation/world_model_interaction.py \
+    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 \
@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_stackbox"
         --n_iter 7 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox/case3/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 07:35:33.682231: 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`.
-2026-02-08 07:35:33.685275: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:35:33.716682: 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
-2026-02-08 07:35:33.716728: 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
-2026-02-08 07:35:33.718523: 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
-2026-02-08 07:35:33.726756: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:35:33.727105: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 07:35:34.356722: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+2026-02-19 19:33:31.235859: 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`.
+2026-02-19 19:33:31.283866: 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
+2026-02-19 19:33:31.283908: 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
+2026-02-19 19:33:31.284941: 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
+2026-02-19 19:33:31.291610: 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.
+2026-02-19 19:33:32.199716: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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%|          | 0/7 [00:00<?, ?it/s]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/7 [00:00<?, ?it/s]
  14%|█▍        | 1/7 [00:24<02:24, 24.10s/it]
->>> 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
  29%|██▊       | 2/7 [00:47<01:58, 23.62s/it]
  43%|████▎     | 3/7 [01:10<01:34, 23.51s/it]
  57%|█████▋    | 4/7 [01:34<01:10, 23.46s/it]
@@ -129,6 +57,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 3: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
 >>> Step 4: generating actions ...
->>> Step 4: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 5: generating actions ...
+>>> Step 4: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...

unitree_z1_dual_arm_stackbox/case3/psnr_result.json

@@ -1,5 +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/unitree_z1_dual_arm_stackbox_case3_amd.mp4",
-    "psnr": 25.400458754751128
+    "pred_video": "unitree_z1_dual_arm_stackbox/case3/output/inference/25_full_fs4.mp4",
+    "psnr": 39.63695040491171
 }

unitree_z1_dual_arm_stackbox/case3/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_stackbox"
         --n_iter 7 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox/case4/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 07:38:45.572744: 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`.
-2026-02-08 07:38:45.576864: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:38:45.624825: 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
-2026-02-08 07:38:45.624883: 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
-2026-02-08 07:38:45.627150: 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
-2026-02-08 07:38:45.638316: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:38:45.638803: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 07:38:46.426363: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+2026-02-19 19:36:32.251051: 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`.
+2026-02-19 19:36:32.298464: 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
+2026-02-19 19:36:32.298506: 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
+2026-02-19 19:36:32.299538: 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
+2026-02-19 19:36:32.306168: 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.
+2026-02-19 19:36:33.213503: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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%|          | 0/7 [00:00<?, ?it/s]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/7 [00:00<?, ?it/s]
  14%|█▍        | 1/7 [00:24<02:24, 24.05s/it]
->>> 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
  29%|██▊       | 2/7 [00:47<01:57, 23.58s/it]
  43%|████▎     | 3/7 [01:10<01:33, 23.45s/it]
  57%|█████▋    | 4/7 [01:33<01:10, 23.43s/it]
@@ -129,6 +57,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 3: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
 >>> Step 4: generating actions ...
->>> Step 4: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 5: generating actions ...
+>>> Step 4: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...

unitree_z1_dual_arm_stackbox/case4/psnr_result.json

@@ -1,5 +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/unitree_z1_dual_arm_stackbox_case4_amd.mp4",
-    "psnr": 24.098958457373858
+    "pred_video": "unitree_z1_dual_arm_stackbox/case4/output/inference/35_full_fs4.mp4",
+    "psnr": 42.34177660061245
 }

unitree_z1_dual_arm_stackbox/case4/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_stackbox"
         --n_iter 7 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox_v2/case1/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 07:51:23.961486: 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`.
-2026-02-08 07:51:24.200063: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:51:24.522299: 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
-2026-02-08 07:51:24.522350: 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
-2026-02-08 07:51:24.528237: 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
-2026-02-08 07:51:24.579400: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:51:24.579644: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 07:51:25.781311: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+2026-02-19 19:39:32.908698: 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`.
+2026-02-19 19:39:32.956378: 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
+2026-02-19 19:39:32.956417: 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
+2026-02-19 19:39:32.957459: 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
+2026-02-19 19:39:32.964104: 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.
+2026-02-19 19:39:33.875854: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
 Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/11 [00:00<?, ?it/s]
   9%|▉         | 1/11 [00:24<04:01, 24.10s/it]
->>> 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
  18%|█▊        | 2/11 [00:47<03:32, 23.61s/it]
  27%|██▋       | 3/11 [01:10<03:08, 23.50s/it]
  36%|███▋      | 4/11 [01:34<02:44, 23.45s/it]
@@ -141,6 +69,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 6: generating actions ...
 >>> Step 6: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 7: generating actions ...
->>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_z1_dual_arm_stackbox_v2/case1/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_z1_dual_arm_stackbox_v2/case1/unitree_z1_dual_arm_stackbox_v2_case1.mp4",
-    "pred_video": "unitree_z1_dual_arm_stackbox_v2/case1/output/inference/unitree_z1_dual_arm_stackbox_v2_case1_amd.mp4",
-    "psnr": 18.126776535969576
+    "pred_video": "unitree_z1_dual_arm_stackbox_v2/case1/output/inference/5_full_fs4.mp4",
+    "psnr": 26.68301835085306
 }

unitree_z1_dual_arm_stackbox_v2/case1/run_world_model_interaction.sh

@@ -2,7 +2,7 @@ res_dir="unitree_z1_dual_arm_stackbox_v2/case1"
 dataset="unitree_z1_dual_arm_stackbox_v2"
 
 {
-    time CUDA_VISIBLE_DEVICES=7 python3 scripts/evaluation/world_model_interaction.py \
+    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 \
@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_stackbox_v2"
         --n_iter 11 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox_v2/case2/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 07:56:31.144789: 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`.
-2026-02-08 07:56:31.148256: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:56:31.178870: 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
-2026-02-08 07:56:31.178898: 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
-2026-02-08 07:56:31.180683: 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
-2026-02-08 07:56:31.188800: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:56:31.189142: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 07:56:31.810098: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+2026-02-19 19:44:07.724109: 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`.
+2026-02-19 19:44:07.771461: 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
+2026-02-19 19:44:07.771505: 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
+2026-02-19 19:44:07.772537: 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
+2026-02-19 19:44:07.779172: 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.
+2026-02-19 19:44:08.688975: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/11 [00:00<?, ?it/s]
   9%|▉         | 1/11 [00:24<04:00, 24.03s/it]
->>> 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
  18%|█▊        | 2/11 [00:47<03:31, 23.54s/it]
  27%|██▋       | 3/11 [01:10<03:07, 23.42s/it]
  36%|███▋      | 4/11 [01:33<02:43, 23.40s/it]
@@ -141,6 +69,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 6: generating actions ...
 >>> Step 6: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 7: generating actions ...
->>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_z1_dual_arm_stackbox_v2/case2/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_z1_dual_arm_stackbox_v2/case2/unitree_z1_dual_arm_stackbox_v2_case2.mp4",
-    "pred_video": "unitree_z1_dual_arm_stackbox_v2/case2/output/inference/unitree_z1_dual_arm_stackbox_v2_case2_amd.mp4",
-    "psnr": 19.38130614773096
+    "pred_video": "unitree_z1_dual_arm_stackbox_v2/case2/output/inference/15_full_fs4.mp4",
+    "psnr": 27.46347145461597
 }

unitree_z1_dual_arm_stackbox_v2/case2/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_stackbox_v2"
         --n_iter 11 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox_v2/case3/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 07:56:04.467082: 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`.
-2026-02-08 07:56:04.470145: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:56:04.502248: 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
-2026-02-08 07:56:04.502277: 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
-2026-02-08 07:56:04.504088: 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
-2026-02-08 07:56:04.512557: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 07:56:04.512830: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 07:56:05.259641: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+2026-02-19 19:48:42.460586: 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`.
+2026-02-19 19:48:42.508096: 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
+2026-02-19 19:48:42.508140: 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
+2026-02-19 19:48:42.509152: 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
+2026-02-19 19:48:42.515865: 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.
+2026-02-19 19:48:43.425699: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/11 [00:00<?, ?it/s]
   9%|▉         | 1/11 [00:24<04:00, 24.07s/it]
->>> 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
  18%|█▊        | 2/11 [00:47<03:32, 23.62s/it]
  27%|██▋       | 3/11 [01:10<03:08, 23.51s/it]
  36%|███▋      | 4/11 [01:34<02:44, 23.46s/it]
@@ -141,6 +69,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 6: generating actions ...
 >>> Step 6: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 7: generating actions ...
->>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_z1_dual_arm_stackbox_v2/case3/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_z1_dual_arm_stackbox_v2/case3/unitree_z1_dual_arm_stackbox_v2_case3.mp4",
-    "pred_video": "unitree_z1_dual_arm_stackbox_v2/case3/output/inference/unitree_z1_dual_arm_stackbox_v2_case3_amd.mp4",
-    "psnr": 18.74462122425683
+    "pred_video": "unitree_z1_dual_arm_stackbox_v2/case3/output/inference/25_full_fs4.mp4",
+    "psnr": 28.604047286947512
 }

unitree_z1_dual_arm_stackbox_v2/case3/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_stackbox_v2"
         --n_iter 11 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_dual_arm_stackbox_v2/case4/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 08:04:16.104516: 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`.
-2026-02-08 08:04:16.109112: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:04:16.138703: 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
-2026-02-08 08:04:16.138737: 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
-2026-02-08 08:04:16.140302: 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
-2026-02-08 08:04:16.147672: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:04:16.147903: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 08:04:17.363218: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+2026-02-19 19:53:17.574354: 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`.
+2026-02-19 19:53:17.621335: 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
+2026-02-19 19:53:17.621388: 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
+2026-02-19 19:53:17.622415: 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
+2026-02-19 19:53:17.629050: 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.
+2026-02-19 19:53:18.537233: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
 Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/11 [00:00<?, ?it/s]
   9%|▉         | 1/11 [00:24<04:00, 24.09s/it]
->>> 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
  18%|█▊        | 2/11 [00:47<03:32, 23.62s/it]
  27%|██▋       | 3/11 [01:10<03:07, 23.49s/it]
  36%|███▋      | 4/11 [01:34<02:44, 23.47s/it]
@@ -141,6 +69,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 6: generating actions ...
 >>> Step 6: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 7: generating actions ...
->>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_z1_dual_arm_stackbox_v2/case4/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_z1_dual_arm_stackbox_v2/case4/unitree_z1_dual_arm_stackbox_v2_case4.mp4",
-    "pred_video": "unitree_z1_dual_arm_stackbox_v2/case4/output/inference/unitree_z1_dual_arm_stackbox_v2_case4_amd.mp4",
-    "psnr": 19.526448380726254
+    "pred_video": "unitree_z1_dual_arm_stackbox_v2/case4/output/inference/35_full_fs4.mp4",
+    "psnr": 25.578757174083307
 }

unitree_z1_dual_arm_stackbox_v2/case4/run_world_model_interaction.sh

@@ -2,7 +2,7 @@ res_dir="unitree_z1_dual_arm_stackbox_v2/case4"
 dataset="unitree_z1_dual_arm_stackbox_v2"
 
 {
-    time CUDA_VISIBLE_DEVICES=6 python3 scripts/evaluation/world_model_interaction.py \
+    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 \
@@ -20,5 +20,6 @@ dataset="unitree_z1_dual_arm_stackbox_v2"
         --n_iter 11 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_stackbox/case1/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 08:12:47.424053: 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`.
-2026-02-08 08:12:47.427280: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:12:47.458253: 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
-2026-02-08 08:12:47.458288: 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
-2026-02-08 08:12:47.462758: 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
-2026-02-08 08:12:47.518283: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:12:47.518566: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 08:12:48.593011: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+2026-02-19 19:57:52.488339: 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`.
+2026-02-19 19:57:52.536176: 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
+2026-02-19 19:57:52.536222: 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
+2026-02-19 19:57:52.537285: 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
+2026-02-19 19:57:52.544051: 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.
+2026-02-19 19:57:53.469912: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
 Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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%|          | 0/12 [00:00<?, ?it/s]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/12 [00:00<?, ?it/s]
   8%|▊         | 1/12 [00:24<04:24, 24.06s/it]
->>> 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
  17%|█▋        | 2/12 [00:47<03:55, 23.56s/it]
  25%|██▌       | 3/12 [01:10<03:31, 23.46s/it]
  33%|███▎      | 4/12 [01:33<03:07, 23.43s/it]
@@ -144,6 +72,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>>>>>>>>>>>>>>>>>>>>>>>
 >>> Step 7: generating actions ...
 >>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 8: generating actions ...
->>> Step 8: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 8: generating actions ...
+>>> Step 8: interacting with world model ...

unitree_z1_stackbox/case1/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_z1_stackbox/case1/unitree_z1_stackbox_case1.mp4",
-    "pred_video": "unitree_z1_stackbox/case1/output/inference/unitree_z1_stackbox_case1_amd.mp4",
-    "psnr": 19.81391789862606
+    "pred_video": "unitree_z1_stackbox/case1/output/inference/5_full_fs4.mp4",
+    "psnr": 46.05271283048069
 }

unitree_z1_stackbox/case1/run_world_model_interaction.sh

@@ -2,7 +2,7 @@ res_dir="unitree_z1_stackbox/case1"
 dataset="unitree_z1_stackbox"
 
 {
-    time CUDA_VISIBLE_DEVICES=5 python3 scripts/evaluation/world_model_interaction.py \
+    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 \
@@ -20,5 +20,6 @@ dataset="unitree_z1_stackbox"
         --n_iter 12 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_stackbox/case2/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 08:15:49.934949: 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`.
-2026-02-08 08:15:49.937974: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:15:49.969069: 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
-2026-02-08 08:15:49.969100: 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
-2026-02-08 08:15:49.970909: 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
-2026-02-08 08:15:49.979005: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:15:49.979255: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 08:15:50.597743: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+2026-02-19 20:02:50.975402: 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`.
+2026-02-19 20:02:51.023211: 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
+2026-02-19 20:02:51.023253: 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
+2026-02-19 20:02:51.024328: 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
+2026-02-19 20:02:51.031176: 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.
+2026-02-19 20:02:51.947400: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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%|          | 0/12 [00:00<?, ?it/s]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/12 [00:00<?, ?it/s]
   8%|▊         | 1/12 [00:24<04:24, 24.08s/it]
->>> 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
  17%|█▋        | 2/12 [00:47<03:56, 23.62s/it]
  25%|██▌       | 3/12 [01:10<03:31, 23.51s/it]
  33%|███▎      | 4/12 [01:34<03:07, 23.48s/it]
@@ -144,6 +72,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>>>>>>>>>>>>>>>>>>>>>>>
 >>> Step 7: generating actions ...
 >>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 8: generating actions ...
->>> Step 8: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 8: generating actions ...
+>>> Step 8: interacting with world model ...

unitree_z1_stackbox/case2/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_z1_stackbox/case2/unitree_z1_stackbox_case2.mp4",
-    "pred_video": "unitree_z1_stackbox/case2/output/inference/unitree_z1_stackbox_case2_amd.mp4",
-    "psnr": 21.083821459054743
+    "pred_video": "unitree_z1_stackbox/case2/output/inference/15_full_fs4.mp4",
+    "psnr": 43.005233352958804
 }

unitree_z1_stackbox/case2/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_z1_stackbox"
         --n_iter 12 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_stackbox/case3/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 08:16:22.299521: 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`.
-2026-02-08 08:16:22.302545: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:16:22.335354: 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
-2026-02-08 08:16:22.335389: 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
-2026-02-08 08:16:22.337179: 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
-2026-02-08 08:16:22.345296: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:16:22.345548: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 08:16:23.008743: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
-[rank: 0] Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+2026-02-19 20:07:49.410622: 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`.
+2026-02-19 20:07:49.457896: 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
+2026-02-19 20:07:49.457948: 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
+2026-02-19 20:07:49.458967: 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
+2026-02-19 20:07:49.465632: 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.
+2026-02-19 20:07:50.373326: 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.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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%|          | 0/12 [00:00<?, ?it/s]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/12 [00:00<?, ?it/s]
   8%|▊         | 1/12 [00:24<04:25, 24.17s/it]
->>> 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
  17%|█▋        | 2/12 [00:47<03:56, 23.64s/it]
  25%|██▌       | 3/12 [01:10<03:31, 23.53s/it]
  33%|███▎      | 4/12 [01:34<03:07, 23.49s/it]
@@ -144,6 +72,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>>>>>>>>>>>>>>>>>>>>>>>
 >>> Step 7: generating actions ...
 >>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 8: generating actions ...
->>> Step 8: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 8: generating actions ...
+>>> Step 8: interacting with world model ...

unitree_z1_stackbox/case3/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_z1_stackbox/case3/unitree_z1_stackbox_case3.mp4",
-    "pred_video": "unitree_z1_stackbox/case3/output/inference/unitree_z1_stackbox_case3_amd.mp4",
-    "psnr": 21.322784880212172
+    "pred_video": "unitree_z1_stackbox/case3/output/inference/25_full_fs4.mp4",
+    "psnr": 49.489774674892764
 }

unitree_z1_stackbox/case3/run_world_model_interaction.sh

@@ -20,5 +20,6 @@ dataset="unitree_z1_stackbox"
         --n_iter 12 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"

unitree_z1_stackbox/case4/output.log

@@ -1,34 +1,16 @@
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/lightning_fabric/__init__.py:29: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
-  __import__("pkg_resources").declare_namespace(__name__)
-2026-02-08 08:25:54.657305: 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`.
-2026-02-08 08:25:54.660628: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:25:54.691237: 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
-2026-02-08 08:25:54.691275: 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
-2026-02-08 08:25:54.693046: 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
-2026-02-08 08:25:54.701142: I external/local_tsl/tsl/cuda/cudart_stub.cc:31] Could not find cuda drivers on your machine, GPU will not be used.
-2026-02-08 08:25:54.701413: 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 AVX512_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-08 08:25:55.801367: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+2026-02-19 20:12:48.029611: 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`.
+2026-02-19 20:12:48.076914: 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
+2026-02-19 20:12:48.076957: 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
+2026-02-19 20:12:48.077981: 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
+2026-02-19 20:12:48.084620: 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.
+2026-02-19 20:12:49.004753: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
 Global seed set to 123
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/kornia/feature/lightglue.py:44: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
-  @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
-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).
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/open_clip/factory.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  checkpoint = torch.load(checkpoint_path, map_location=map_location)
-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).
-/mnt/ASC1637/unifolm-world-model-action/scripts/evaluation/world_model_interaction.py:86: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
-  state_dict = torch.load(ckpt, map_location="cpu")
->>> model checkpoint loaded.
->>> Load pre-trained model ...
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
+>>> Prepared model loaded.
+>>> Diffusion backbone (model.model) converted to FP16.
+>>> Projectors (image_proj_model, state_projector, action_projector) converted to FP16.
+>>> Encoders (cond_stage_model, embedder) converted to FP16.
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -46,69 +28,15 @@ INFO:root:***** Configing Data *****
 >>> unitree_g1_pack_camera: data stats loaded.
 >>> unitree_g1_pack_camera: normalizer initiated.
 >>> Dataset is successfully loaded ...
+    ✓ KV fused: 66 attention layers
 >>> 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%|          | 0/12 [00:00<?, ?it/s]/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:5501: UserWarning: Attempting to use hipBLASLt on an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at ../aten/src/ATen/Context.cpp:296.)
-  proj = linear(q, w, b)
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Flash attention support on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:225.)
-  attn_output = scaled_dot_product_attention(
-/mnt/ASC1637/miniconda3/envs/unifolm-wma-o/lib/python3.10/site-packages/torch/nn/functional.py:6278: UserWarning: Memory Efficient attention on Navi31 GPU is still experimental. Enable it with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1. (Triggered internally at ../aten/src/ATen/native/transformers/hip/sdp_utils.cpp:269.)
+
   0%|          | 0/12 [00:00<?, ?it/s]
   8%|▊         | 1/12 [00:24<04:24, 24.06s/it]
->>> 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
  17%|█▋        | 2/12 [00:47<03:55, 23.59s/it]
  25%|██▌       | 3/12 [01:10<03:31, 23.49s/it]
  33%|███▎      | 4/12 [01:34<03:07, 23.44s/it]
@@ -144,6 +72,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>>>>>>>>>>>>>>>>>>>>>>>
 >>> Step 7: generating actions ...
 >>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 8: generating actions ...
->>> Step 8: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 8: generating actions ...
+>>> Step 8: interacting with world model ...

unitree_z1_stackbox/case4/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "unitree_z1_stackbox/case4/unitree_z1_stackbox_case4.mp4",
-    "pred_video": "unitree_z1_stackbox/case4/output/inference/unitree_z1_stackbox_case4_amd.mp4",
-    "psnr": 25.32928948331741
+    "pred_video": "unitree_z1_stackbox/case4/output/inference/35_full_fs4.mp4",
+    "psnr": 47.18724378194084
 }

unitree_z1_stackbox/case4/run_world_model_interaction.sh

@@ -2,7 +2,7 @@ res_dir="unitree_z1_stackbox/case4"
 dataset="unitree_z1_stackbox"
 
 {
-    time CUDA_VISIBLE_DEVICES=7 python3 scripts/evaluation/world_model_interaction.py \
+    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 \
@@ -20,5 +20,6 @@ dataset="unitree_z1_stackbox"
         --n_iter 12 \
         --timestep_spacing 'uniform_trailing' \
         --guidance_rescale 0.7 \
-        --perframe_ae
+        --perframe_ae \
+        --fast_policy_no_decode
 } 2>&1 | tee "${res_dir}/output.log"