<feat> lower RAM usage when loading models

app.py

@@ -59,23 +59,27 @@ def init_basemodel():
     transformer = HunyuanVideoTransformer3DModel.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V', 
                                                                  subfolder="transformer", 
                                                                  inference_subject_driven=False, 
-                                                                 low_cpu_mem_usage=True).requires_grad_(False).to(device, dtype=weight_dtype)
+                                                                 low_cpu_mem_usage=True, 
+                                                                 torch_dtype=weight_dtype).requires_grad_(False).to(device)
     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).requires_grad_(False).to(device, dtype=weight_dtype)
+                                                              low_cpu_mem_usage=True, 
+                                                              torch_dtype=weight_dtype).requires_grad_(False).to(device)
     torch.cuda.empty_cache()
     gc.collect()
     text_encoder = transformers.LlavaForConditionalGeneration.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V', 
                                                                               subfolder="text_encoder", 
-                                                                              low_cpu_mem_usage=True).requires_grad_(False).to(device, dtype=weight_dtype)
+                                                                              low_cpu_mem_usage=True, 
+                                                                              torch_dtype=weight_dtype).requires_grad_(False).to(device)
     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).requires_grad_(False).to(device, dtype=weight_dtype)
+                                                                low_cpu_mem_usage=True, 
+                                                                torch_dtype=weight_dtype).requires_grad_(False).to(device)
     torch.cuda.empty_cache()
     gc.collect()                                              
     tokenizer = transformers.AutoTokenizer.from_pretrained('hunyuanvideo-community/HunyuanVideo-I2V',