offload text_encoder and transformer to CPU

inference_script.py

@@ -468,6 +468,9 @@ def process_video(
         if pipe.transformer.config.use_rotary_positional_embeddings
         else None
     )
+    
+    # offload text encoder to CPU
+    pipe.text_encoder.to("cpu")
 
     # Predict noise
     predicted_noise = pipe.transformer(
@@ -478,6 +481,8 @@ def process_video(
         return_dict=False,
     )[0]
 
+    # offload transformer to CPU
+    pipe.transformer.to("cpu")
     
     latent_generate = pipe.scheduler.get_velocity(
         predicted_noise, latent, timesteps