KV 融合实现完成。改动总结： 速度微弱提升psnr略微上升

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

.claude/settings.local.json

@@ -4,7 +4,12 @@
       "Bash(conda env list:*)",
       "Bash(mamba env:*)",
       "Bash(micromamba env list:*)",
-      "Bash(echo:*)"
+      "Bash(echo:*)",
+      "Bash(git show:*)",
+      "Bash(nvidia-smi:*)",
+      "Bash(conda activate unifolm-wma)",
+      "Bash(conda info:*)",
+      "Bash(direnv allow:*)"
     ]
   }
 }

.envrc

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

.gitignore

@@ -131,3 +131,4 @@ Experiment/log
 *.ckpt
 
 *.0
+ckpts/unifolm_wma_dual.ckpt.prepared.pt

scripts/evaluation/world_model_interaction.py

@@ -9,6 +9,8 @@ import logging
 import einops
 import warnings
 import imageio
+import atexit
+from concurrent.futures import ThreadPoolExecutor
 
 from pytorch_lightning import seed_everything
 from omegaconf import OmegaConf
@@ -16,8 +18,9 @@ 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
@@ -153,6 +156,81 @@ def save_results(video: Tensor, filename: str, fps: int = 8) -> None:
                                options={'crf': '10'})
 
 
+# ========== Async I/O ==========
+_io_executor: Optional[ThreadPoolExecutor] = None
+_io_futures: List[Any] = []
+
+
+def _get_io_executor() -> ThreadPoolExecutor:
+    global _io_executor
+    if _io_executor is None:
+        _io_executor = ThreadPoolExecutor(max_workers=2)
+    return _io_executor
+
+
+def _flush_io():
+    """Wait for all pending async I/O to finish."""
+    global _io_futures
+    for fut in _io_futures:
+        try:
+            fut.result()
+        except Exception as e:
+            print(f">>> [async I/O] error: {e}")
+    _io_futures.clear()
+
+
+atexit.register(_flush_io)
+
+
+def _save_results_sync(video_cpu: Tensor, filename: str, fps: int) -> None:
+    """Synchronous save on CPU tensor (runs in background thread)."""
+    video = torch.clamp(video_cpu.float(), -1., 1.)
+    n = video.shape[0]
+    video = video.permute(2, 0, 1, 3, 4)
+    frame_grids = [
+        torchvision.utils.make_grid(framesheet, nrow=int(n), padding=0)
+        for framesheet in video
+    ]
+    grid = torch.stack(frame_grids, dim=0)
+    grid = (grid + 1.0) / 2.0
+    grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
+    torchvision.io.write_video(filename,
+                               grid,
+                               fps=fps,
+                               video_codec='h264',
+                               options={'crf': '10'})
+
+
+def save_results_async(video: Tensor, filename: str, fps: int = 8) -> None:
+    """Submit video saving to background thread pool."""
+    video_cpu = video.detach().cpu()
+    fut = _get_io_executor().submit(_save_results_sync, video_cpu, filename, fps)
+    _io_futures.append(fut)
+
+
+def _log_to_tb_sync(writer, video_cpu: Tensor, tag: str, fps: int) -> None:
+    """Synchronous TensorBoard log on CPU tensor (runs in background thread)."""
+    if video_cpu.dim() == 5:
+        n = video_cpu.shape[0]
+        video = video_cpu.permute(2, 0, 1, 3, 4)
+        frame_grids = [
+            torchvision.utils.make_grid(framesheet, nrow=int(n), padding=0)
+            for framesheet in video
+        ]
+        grid = torch.stack(frame_grids, dim=0)
+        grid = (grid + 1.0) / 2.0
+        grid = grid.unsqueeze(dim=0)
+        writer.add_video(tag, grid, fps=fps)
+
+
+def log_to_tensorboard_async(writer, data: Tensor, tag: str, fps: int = 10) -> None:
+    """Submit TensorBoard logging to background thread pool."""
+    if isinstance(data, torch.Tensor) and data.dim() == 5:
+        data_cpu = data.detach().cpu()
+        fut = _get_io_executor().submit(_log_to_tb_sync, writer, data_cpu, tag, fps)
+        _io_futures.append(fut)
+
+
 def get_init_frame_path(data_dir: str, sample: dict) -> str:
     """Construct the init_frame path from directory and sample metadata.
 
@@ -462,26 +540,51 @@ 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)
+
+    prepared_path = args.ckpt_path + ".prepared.pt"
+    if os.path.exists(prepared_path):
+        # ---- Fast path: load the fully-prepared model ----
+        print(f">>> Loading prepared model from {prepared_path} ...")
+        model = torch.load(prepared_path,
+                           map_location=f"cuda:{gpu_no}",
+                           weights_only=False,
+                           mmap=True)
+        model.eval()
+        print(f">>> Prepared model loaded.")
+    else:
+        # ---- Normal path: construct + load checkpoint ----
         config['model']['params']['wma_config']['params'][
             'use_checkpoint'] = False
         model = instantiate_from_config(config.model)
         model.perframe_ae = args.perframe_ae
+
         assert os.path.exists(args.ckpt_path), "Error: checkpoint Not Found!"
         model = load_model_checkpoint(model, args.ckpt_path)
         model.eval()
+        model = model.cuda(gpu_no)
         print(f'>>> Load pre-trained model ...')
 
-    # 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).")
+
+    # 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
@@ -654,16 +757,16 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
                     cond_obs_queues = populate_queues(cond_obs_queues,
                                                       observation)
 
-                # Save the imagen videos for decision-making
+                # Save the imagen videos for decision-making (async)
                 if pred_videos_0 is not None:
                     sample_tag = f"{args.dataset}-vid{sample['videoid']}-dm-fs-{fs}/itr-{itr}"
-                    log_to_tensorboard(writer,
+                    log_to_tensorboard_async(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,
+                log_to_tensorboard_async(writer,
                                          pred_videos_1,
                                          sample_tag,
                                          fps=args.save_fps)
@@ -671,12 +774,12 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
                 # Save the imagen videos for decision-making
                 if pred_videos_0 is not None:
                     sample_video_file = f'{video_save_dir}/dm/{fs}/itr-{itr}.mp4'
-                    save_results(pred_videos_0.cpu(),
+                    save_results_async(pred_videos_0,
                                        sample_video_file,
                                        fps=args.save_fps)
                 # Save videos environment changes via world-model interaction
                 sample_video_file = f'{video_save_dir}/wm/{fs}/itr-{itr}.mp4'
-                save_results(pred_videos_1.cpu(),
+                save_results_async(pred_videos_1,
                                    sample_video_file,
                                    fps=args.save_fps)
 
@@ -686,12 +789,15 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
 
             full_video = torch.cat(wm_video, dim=2)
             sample_tag = f"{args.dataset}-vid{sample['videoid']}-wd-fs-{fs}/full"
-            log_to_tensorboard(writer,
+            log_to_tensorboard_async(writer,
                                      full_video,
                                      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)
+            save_results_async(full_video, sample_full_video_file, fps=args.save_fps)
+
+    # Wait for all async I/O to complete
+    _flush_io()
 
 
 def get_parser():

src/unifolm_wma/modules/attention.py

@@ -100,6 +100,7 @@ class CrossAttention(nn.Module):
         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:
@@ -117,6 +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
@@ -143,6 +165,12 @@ class CrossAttention(nn.Module):
                                     self.agent_action_context_len +
                                     self.text_context_len:, :]
 
+            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)
@@ -154,6 +182,9 @@ class CrossAttention(nn.Module):
         else:
             if not spatial_self_attn:
                 context = context[:, :self.text_context_len, :]
+            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)
 
@@ -267,6 +298,10 @@ class CrossAttention(nn.Module):
         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:, :]
+                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)
@@ -279,6 +314,11 @@ class CrossAttention(nn.Module):
                 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:, :]
+                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)
@@ -296,6 +336,12 @@ class CrossAttention(nn.Module):
                 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:, :]
 
+                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)
@@ -328,6 +374,9 @@ class CrossAttention(nn.Module):
             if not spatial_self_attn:
                 assert 1 > 2, ">>> ERROR: you should never go into here ..."
                 context = context[:, :self.text_context_len, :]
+            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))

src/unifolm_wma/modules/networks/wma_model.py

@@ -688,6 +688,17 @@ class WMAModel(nn.Module):
         # 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,
@@ -842,15 +853,16 @@ 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
+            torch.cuda.current_stream().wait_stream(s_stream)
-            if b > 1:
-                s_y = self.state_unet(x_state, timesteps[:ba], hs_a,
-                                      context_action[:2], **kwargs)
-            else:
-                s_y = self.state_unet(x_state, timesteps, hs_a,
-                                      context_action[:2], **kwargs)
         else:
             a_y = torch.zeros_like(x_action)
             s_y = torch.zeros_like(x_state)

unitree_z1_dual_arm_stackbox_v2/case1/output.log

@@ -1,24 +1,13 @@
-2026-02-10 17:39:22.590654: 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 11:59:27.241485: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
-2026-02-10 17:39:22.640645: 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 11:59:27.291755: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
-2026-02-10 17:39:22.640689: 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 11:59:27.291807: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
-2026-02-10 17:39:22.642010: 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 11:59:27.293169: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
-2026-02-10 17:39:22.649530: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+2026-02-11 11:59:27.300838: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
 To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
-2026-02-10 17:39:23.575804: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
+2026-02-11 11:59:28.228009: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
 Global seed set to 123
-INFO:mainlogger:LatentVisualDiffusion: Running in v-prediction mode
+>>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
-INFO:unifolm_wma.models.diffusion_head.conditional_unet1d:number of parameters: 5.010531e+08
+>>> Prepared model loaded.
-INFO:unifolm_wma.models.diffusion_head.conditional_unet1d:number of parameters: 5.010531e+08
-AE working on z of shape (1, 4, 32, 32) = 4096 dimensions.
-INFO:root:Loaded ViT-H-14 model config.
-DEBUG:urllib3.connectionpool:Starting new HTTPS connection (1): hf-mirror.com:443
-DEBUG:urllib3.connectionpool:https://hf-mirror.com:443 "HEAD /laion/CLIP-ViT-H-14-laion2B-s32B-b79K/resolve/main/open_clip_pytorch_model.bin HTTP/1.1" 302 0
-INFO:root:Loading pretrained ViT-H-14 weights (laion2b_s32b_b79k).
-INFO:root:Loaded ViT-H-14 model config.
-DEBUG:urllib3.connectionpool:https://hf-mirror.com:443 "HEAD /laion/CLIP-ViT-H-14-laion2B-s32B-b79K/resolve/main/open_clip_pytorch_model.bin HTTP/1.1" 302 0
-INFO:root:Loading pretrained ViT-H-14 weights (laion2b_s32b_b79k).
->>> model checkpoint loaded.
->>> Load pre-trained model ...
 INFO:root:***** Configing Data *****
 >>> unitree_z1_stackbox: 1 data samples loaded.
 >>> unitree_z1_stackbox: data stats loaded.
@@ -36,13 +25,16 @@ 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]
+  9%|▉         | 1/11 [00:34<05:40, 34.05s/it]>>> Step 0: generating actions ...
+>>> Step 0: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
 >>> Step 1: generating actions ...
 DEBUG:PIL.Image:Importing BlpImagePlugin
@@ -92,9 +84,7 @@ DEBUG:PIL.Image:Importing WmfImagePlugin
 DEBUG:PIL.Image:Importing WebPImagePlugin
 DEBUG:PIL.Image:Importing WmfImagePlugin
 DEBUG:PIL.Image:Importing XbmImagePlugin
-DEBUG:PIL.Image:Importing XVThumbImagePlugin
+DEBUG:PIL.Image:Importing XpmImagePlugin
-
-  9%|▉         | 1/11 [00:35<05:55, 35.52s/it]
 DEBUG:PIL.Image:Importing XVThumbImagePlugin
 
  18%|█▊        | 2/11 [01:08<05:07, 34.17s/it]
@@ -125,6 +115,6 @@ DEBUG:PIL.Image:Importing XVThumbImagePlugin
 >>> Step 6: generating actions ...
 >>> Step 6: interacting with world model ...
 >>>>>>>>>>>>>>>>>>>>>>>>
->>> Step 7: generating actions ...
+>>> Step 7: generating actions ...
->>> Step 7: interacting with world model ...
+>>> Step 7: interacting with world model ...
->>>>>>>>>>>>>>>>>>>>>>>>
+>>>>>>>>>>>>>>>>>>>>>>>>

unitree_z1_dual_arm_stackbox_v2/case1/psnr_result.json

@@ -1,5 +1,5 @@
 {
     "gt_video": "/home/qhy/unifolm-world-model-action/unitree_z1_dual_arm_stackbox_v2/case1/unitree_z1_dual_arm_stackbox_v2_case1.mp4",
     "pred_video": "/home/qhy/unifolm-world-model-action/unitree_z1_dual_arm_stackbox_v2/case1/output/inference/5_full_fs4.mp4",
-    "psnr": 25.12008483689618
+    "psnr": 28.167025381705358
 }