Fix VAE scale tensor device mismatch: recreate scale list on GPU and move T5 encoder

app.py

@@ -620,12 +620,26 @@ def run_graio_demo(args):
                         wan_a2v.vae.mean = wan_a2v.vae.mean.to(cuda_device)
                     if hasattr(wan_a2v.vae, 'std'):
                         wan_a2v.vae.std = wan_a2v.vae.std.to(cuda_device)
+                    # 重新创建 scale 列表，确保在 GPU 上
+                    if hasattr(wan_a2v.vae, 'mean') and hasattr(wan_a2v.vae, 'std'):
+                        wan_a2v.vae.scale = [wan_a2v.vae.mean, 1.0 / wan_a2v.vae.std]
+                    # 更新 VAE 的设备属性
+                    if hasattr(wan_a2v.vae, 'device'):
+                        wan_a2v.vae.device = cuda_device
                 
                 # 移动 CLIP 模型到 GPU
                 if hasattr(wan_a2v, 'clip') and wan_a2v.clip is not None:
                     if hasattr(wan_a2v.clip, 'model'):
                         wan_a2v.clip.model = wan_a2v.clip.model.to(cuda_device)
                 
+                # 移动 T5 encoder 到 GPU（如果不在 CPU 上）
+                if hasattr(wan_a2v, 'text_encoder') and wan_a2v.text_encoder is not None:
+                    if hasattr(wan_a2v.text_encoder, 'model'):
+                        wan_a2v.text_encoder.model = wan_a2v.text_encoder.model.to(cuda_device)
+                    # 更新 T5 encoder 的设备属性
+                    if hasattr(wan_a2v.text_encoder, 'device'):
+                        wan_a2v.text_encoder.device = cuda_device
+                
                 # 更新设备信息
                 wan_a2v.device = cuda_device