- state_unet 放到一个独立的 CUDA stream 上执行

- action_unet 在默认 stream 上同时执行
  - 用 wait_stream 确保两者都完成后再返回
两个 1D UNet 输入完全独立，共享的 hs_a 和 context_action 都是只读的。GPU 利用率只有 ~31%，小张量 kernel 不会打满 GPU，两个 stream 可以真正并行。

src/unifolm_wma/modules/networks/wma_model.py

@@ -688,6 +688,8 @@ 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 forward(self,
                 x: Tensor,
@@ -842,15 +844,16 @@ 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
-            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)