fp16 weights, declared duration 90/45 -> 30

app.py

@@ -29,7 +29,7 @@ torch.backends.cudnn.allow_tf32 = True
 log = logging.getLogger()
 
 device = 'cuda' if torch.cuda.is_available() else 'cpu'
-dtype = torch.float32
+dtype = torch.float16
 
 # Every weight comes from our own namespace; the CLIP encoder and the 44kHz vocoder are repointed
 # inside the vendored `mmaudio` package (features_utils.py / ext/autoencoder.py).
@@ -119,7 +119,7 @@ def _audio_for_video(video, prompt: str, negative_prompt: str, seed: int, num_st
     return video_info, audios.float().cpu()[0]
 
 
-@spaces.GPU(duration=90)
+@spaces.GPU(duration=30)
 @torch.inference_mode()
 def video_to_audio(video: gr.Video, prompt: str, negative_prompt: str, seed: int, num_steps: int,
                    cfg_strength: float, duration: float):
@@ -132,7 +132,7 @@ def video_to_audio(video: gr.Video, prompt: str, negative_prompt: str, seed: int
     return video_save_path
 
 
-@spaces.GPU(duration=90)
+@spaces.GPU(duration=30)
 @torch.inference_mode()
 def video_to_track(video: gr.Video, prompt: str, negative_prompt: str, seed: int, num_steps: int,
                    cfg_strength: float, duration: float):
@@ -151,7 +151,7 @@ def video_to_track(video: gr.Video, prompt: str, negative_prompt: str, seed: int
     return audio_save_path
 
 
-@spaces.GPU(duration=90)
+@spaces.GPU(duration=30)
 @torch.inference_mode()
 def image_to_audio(image: gr.Image, prompt: str, negative_prompt: str, seed: int, num_steps: int,
                    cfg_strength: float, duration: float):
@@ -181,7 +181,7 @@ def image_to_audio(image: gr.Image, prompt: str, negative_prompt: str, seed: int
     return video_save_path
 
 
-@spaces.GPU(duration=45)
+@spaces.GPU(duration=30)
 @torch.inference_mode()
 def text_to_audio(prompt: str, negative_prompt: str, seed: int, num_steps: int, cfg_strength: float,
                   duration: float):