Update generate.py

generate.py

@@ -1,44 +1,39 @@
 import torch
-from audiocraft.models import MusicGen
-from audiocraft.data.audio import audio_write
-from pydub import AudioSegment
-import uuid, os
+from transformers import AutoProcessor, MusicgenForConditionalGeneration
+import soundfile as sf
+import uuid
 
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+MODEL_NAME = "facebook/musicgen-small"
 
-print("Loading MusicGen...")
-model = MusicGen.get_pretrained("facebook/musicgen-small")
+print("Loading MusicGen model...")
 
-OUTPUT_DIR = "outputs"
-os.makedirs(OUTPUT_DIR, exist_ok=True)
+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):
 
-    model.set_generation_params(duration=duration)
+    inputs = processor(
+        text=[prompt],
+        padding=True,
+        return_tensors="pt"
+    ).to(device)
 
-    wav = model.generate([prompt])
-
-    name = str(uuid.uuid4())
-    wav_path = f"{OUTPUT_DIR}/{name}"
-
-    audio_write(
-        wav_path,
-        wav[0].cpu(),
-        model.sample_rate,
-        strategy="loudness",
-        loudness_compressor=True,
+    audio_values = model.generate(
+        **inputs,
+        max_new_tokens=int(duration * 50)
     )
 
-    # Convert to MP3 (smaller + faster)
-    mp3_path = f"{OUTPUT_DIR}/{name}.mp3"
+    filename = f"/tmp/{uuid.uuid4()}.wav"
 
-    AudioSegment.from_wav(f"{wav_path}.wav").export(
-        mp3_path,
-        format="mp3",
-        bitrate="192k"
-    )
+    sampling_rate = model.config.audio_encoder.sampling_rate
 
-    os.remove(f"{wav_path}.wav")
+    sf.write(
+        filename,
+        audio_values[0, 0].cpu().numpy(),
+        sampling_rate
+    )
 
-    return mp3_path
+    return filename