back to vanilla, comment out zeroGPU flags

hf_gradio_app.py

@@ -1,5 +1,5 @@
 import os, random, time
-import spaces
+#import spaces
 import uuid
 import tempfile, shutil
 from pydub import AudioSegment
@@ -76,11 +76,11 @@ with torch.inference_mode():
     diffusion_net.requires_grad_(False).eval()
     image_proj.requires_grad_(False).eval()
     audio_proj.requires_grad_(False).eval()
-    #reference_net.enable_xformers_memory_efficient_attention()
-    #diffusion_net.enable_xformers_memory_efficient_attention()
+    reference_net.enable_xformers_memory_efficient_attention()
+    diffusion_net.enable_xformers_memory_efficient_attention()
     noise_scheduler = FlowMatchEulerDiscreteScheduler()
     pipeline = VideoPipeline(vae=vae, reference_net=reference_net, diffusion_net=diffusion_net, scheduler=noise_scheduler, image_proj=image_proj)
-    #pipeline.to(device=device, dtype=weight_dtype)
+    pipeline.to(device=device, dtype=weight_dtype)
 
 def process_audio(file_path, temp_dir):
     # Load the audio file
@@ -99,11 +99,10 @@ def process_audio(file_path, temp_dir):
     print(f"Processed audio saved at: {output_path}")
     return output_path
 
-#@torch.inference_mode()
-@spaces.GPU(duration=240)
+#@spaces.GPU(duration=240)
+@torch.inference_mode()
 def generate(input_video, input_audio, seed, progress=gr.Progress(track_tqdm=True)):
-    pipeline.to(device=device, dtype=weight_dtype)
-
+    
     is_shared_ui = True if "fffiloni/MEMO" in os.environ['SPACE_ID'] else False
     temp_dir = None
     if is_shared_ui: