Update generate.py

generate.py

@@ -1,29 +1,39 @@
 import torch
+from transformers import AutoProcessor, MusicgenForConditionalGeneration
 import soundfile as sf
 import uuid
-import os
 
-MODEL = None
+MODEL_NAME = "facebook/musicgen-small"
 
+print("Loading MusicGen model...")
 
-def load_model():
-    global MODEL
-    if MODEL is None:
-        from audiocraft.models import MusicGen
-        MODEL = MusicGen.get_pretrained("facebook/musicgen-small")
+processor = AutoProcessor.from_pretrained(MODEL_NAME)
+model = MusicgenForConditionalGeneration.from_pretrained(MODEL_NAME)
 
+device = "cuda" if torch.cuda.is_available() else "cpu"
+model.to(device)
 
-def generate_music(prompt, duration=10):
+def generate_music(prompt, duration):
 
-    load_model()
+    inputs = processor(
+        text=[prompt],
+        padding=True,
+        return_tensors="pt"
+    ).to(device)
 
-    MODEL.set_generation_params(duration=duration)
+    audio_values = model.generate(
+        **inputs,
+        max_new_tokens=int(duration * 50)
+    )
 
-    wav = MODEL.generate([prompt])
+    filename = f"/tmp/{uuid.uuid4()}.wav"
 
-    filename = f"{uuid.uuid4().hex}.wav"
-    path = f"outputs/{filename}"
+    sampling_rate = model.config.audio_encoder.sampling_rate
 
-    sf.write(path, wav[0].cpu().numpy().T, 32000)
+    sf.write(
+        filename,
+        audio_values[0, 0].cpu().numpy(),
+        sampling_rate
+    )
 
-    return path
+    return filename