<fix> move to gpu earlier.

app.py

@@ -52,20 +52,30 @@ def init_basemodel():
     current_task = None
 
     # init models
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    weight_dtype = torch.bfloat16
     transformer = HunyuanVideoTransformer3DModel.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V', 
                                                                  subfolder="transformer", 
                                                                  inference_subject_driven=False, 
-                                                                 low_cpu_mem_usage=True)
+                                                                 low_cpu_mem_usage=True).requires_grad_(False).to(device, dtype=weight_dtype)
+    torch.cuda.empty_cache()
+    gc.collect()
     scheduler = diffusers.FlowMatchEulerDiscreteScheduler()
     vae = diffusers.AutoencoderKLHunyuanVideo.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V', 
                                                               subfolder="vae", 
-                                                              low_cpu_mem_usage=True)
+                                                              low_cpu_mem_usage=True).requires_grad_(False).to(device, dtype=weight_dtype)
+    torch.cuda.empty_cache()
+    gc.collect()
     text_encoder = transformers.LlavaForConditionalGeneration.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V', 
                                                                               subfolder="text_encoder", 
-                                                                              low_cpu_mem_usage=True)
+                                                                              low_cpu_mem_usage=True).requires_grad_(False).to(device, dtype=weight_dtype)
+    torch.cuda.empty_cache()
+    gc.collect()
     text_encoder_2 = transformers.CLIPTextModel.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V', 
                                                                 subfolder="text_encoder_2", 
-                                                                low_cpu_mem_usage=True)
+                                                                low_cpu_mem_usage=True).requires_grad_(False).to(device, dtype=weight_dtype)
+    torch.cuda.empty_cache()
+    gc.collect()                                              
     tokenizer = transformers.AutoTokenizer.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V', 
                                                            subfolder="tokenizer")
     tokenizer_2 = transformers.CLIPTokenizer.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V', 
@@ -73,14 +83,6 @@ def init_basemodel():
     image_processor = transformers.CLIPImageProcessor.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V', 
                                                                       subfolder="image_processor")
 
-    device = "cuda" if torch.cuda.is_available() else "cpu"
-    weight_dtype = torch.bfloat16
-
-    transformer.requires_grad_(False)
-    vae.requires_grad_(False).to(device, dtype=weight_dtype)
-    text_encoder.requires_grad_(False).to(device, dtype=weight_dtype)
-    text_encoder_2.requires_grad_(False).to(device, dtype=weight_dtype)
-    transformer.to(device, dtype=weight_dtype)
     vae.enable_tiling()
     vae.enable_slicing()