update

app.py

@@ -1,10 +1,11 @@
 import os
 import torch
 import random
-import spaces
+# import spaces
 import numpy as np
 import gradio as gr
 import soundfile as sf
+from accelerate import Accelerator
 from transformers import T5Tokenizer, T5EncoderModel
 from diffusers import DDIMScheduler
 from src.models.conditioners import MaskDiT
@@ -33,9 +34,12 @@ def load_models(config_name, ckpt_path, vae_path, device):
     unet.load_state_dict(torch.load(ckpt_path)['model'])
     unet.eval()
 
+    accelerator = Accelerator(mixed_precision="fp16")
+    unet = accelerator.prepare(unet)
+
     # Load noise scheduler
     noise_scheduler = DDIMScheduler(**params['diff'])
-    
+
     latents = torch.randn((1, 128, 128), device=device)
     noise = torch.randn_like(latents)
     timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (1,), device=device)
@@ -43,6 +47,7 @@ def load_models(config_name, ckpt_path, vae_path, device):
 
     return autoencoder, unet, tokenizer, text_encoder, noise_scheduler, params
 
+
 MAX_SEED = np.iinfo(np.int32).max
 
 # Model and config paths
@@ -57,6 +62,7 @@ device = 'cuda' if torch.cuda.is_available() else 'cpu'
 autoencoder, unet, tokenizer, text_encoder, noise_scheduler, params = load_models(config_name, ckpt_path, vae_path,
                                                                                   device)
 
+
 @spaces.GPU
 def generate_audio(text, length,
                    guidance_scale, guidance_rescale, ddim_steps, eta,
@@ -102,7 +108,7 @@ css = """
 with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
     with gr.Column(elem_id="col-container"):
         gr.Markdown("""
-        # EzAudio Text-to-Audio Generator
+        # EzAudio: High-quality Text-to-Audio Generator
         Generate audio from text using a diffusion transformer. Adjust advanced settings for more control.
         """)
 
@@ -125,10 +131,10 @@ with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
         with gr.Accordion("Advanced Settings", open=False):
             guidance_scale = gr.Slider(minimum=1.0, maximum=10, step=0.1, value=5.0, label="Guidance Scale")
             guidance_rescale = gr.Slider(minimum=0.0, maximum=1, step=0.05, value=0.75, label="Guidance Rescale")
-            ddim_steps = gr.Slider(minimum=25, maximum=200, step=5, value=100, label="DDIM Steps")
+            ddim_steps = gr.Slider(minimum=25, maximum=200, step=5, value=50, label="DDIM Steps")
             eta = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=1.0, label="Eta")
             seed = gr.Slider(minimum=0, maximum=MAX_SEED, step=1, value=0, label="Seed")
-            randomize_seed = gr.Checkbox(label="Randomize Seed", value=False)
+            randomize_seed = gr.Checkbox(label="Randomize Seed (Disable Seed)", value=True)
 
         # Examples block
         gr.Examples(
@@ -147,4 +153,4 @@ with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
         )
 
 # Launch the Gradio demo
-demo.launch()
+demo.launch()