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
2026-02-11 12:36:38 +08:00
parent b558856e1e
commit 9a08e27a19
4 changed files with 180 additions and 38 deletions
--- a/scripts/evaluation/world_model_interaction.py
+++ b/scripts/evaluation/world_model_interaction.py
@@ -579,6 +579,12 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:

    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
--- a/src/unifolm_wma/modules/attention.py
+++ b/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
@@ -143,19 +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))
@@ -267,10 +298,14 @@ 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:, :]
-                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:
@@ -279,12 +314,17 @@ 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:, :]
-                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))
@@ -296,14 +336,20 @@ 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:, :]

-                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)

                k, v = map(_reshape_kv, (k, v))
                k_ip, v_ip = map(_reshape_kv, (k_ip, v_ip))
@@ -328,8 +374,11 @@ 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, :]
-            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)
            k, v = map(_reshape_kv, (k, v))
            if use_cache:
                self._kv_cache = {'k': k, 'v': v}
--- a/unitree_z1_dual_arm_stackbox_v2/case1/output.log
+++ b/unitree_z1_dual_arm_stackbox_v2/case1/output.log
@@ -1,10 +1,10 @@
-2026-02-10 22:35:08.834827: 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 22:35:08.884699: 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 22:35:08.884743: 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 22:35:08.886076: 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 22:35:08.893623: 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 11:59:27.241485: 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`.
+2026-02-11 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
+2026-02-11 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
+2026-02-11 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
+2026-02-11 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 22:35:09.824417: 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
 >>> Loading prepared model from ckpts/unifolm_wma_dual.ckpt.prepared.pt ...
 >>> Prepared model loaded.
@@ -25,9 +25,96 @@ 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
+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 [01:08<05:07, 34.17s/it]
+ 27%|██▋       | 3/11 [01:42<04:33, 34.16s/it]
+ 36%|███▋      | 4/11 [02:16<03:59, 34.18s/it]
+ 45%|████▌     | 5/11 [02:50<03:24, 34.14s/it]
+ 55%|█████▍    | 6/11 [03:24<02:50, 34.10s/it]
+ 64%|██████▎   | 7/11 [03:58<02:16, 34.07s/it]
+ 73%|███████▎  | 8/11 [04:32<01:42, 34.03s/it]
+ 82%|████████▏ | 9/11 [05:06<01:08, 34.02s/it]
+ 91%|█████████ | 10/11 [05:40<00:34, 34.04s/it]
+100%|██████████| 11/11 [06:14<00:00, 34.03s/it]
+100%|██████████| 11/11 [06:14<00:00, 34.07s/it]
+>>> Step 1: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 2: generating actions ...
+>>> Step 2: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 3: generating actions ...
+>>> Step 3: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 4: generating actions ...
+>>> Step 4: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 5: generating actions ...
+>>> Step 5: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 6: generating actions ...
+>>> Step 6: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
+>>> Step 7: generating actions ...
+>>> Step 7: interacting with world model ...
+>>>>>>>>>>>>>>>>>>>>>>>>
--- a/unitree_z1_dual_arm_stackbox_v2/case1/psnr_result.json
+++ b/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": 27.279678834152335
+    "psnr": 28.167025381705358
 }