Fix CUDA OOM on long scripts by splitting input into chunks

long_script_inference.py

@@ -0,0 +1,55 @@
+import torch
+from pathlib import Path
+from vibevoice import load_model  # Replace with actual VibeVoice import
+
+# -----------------------
+# Configuration
+# -----------------------
+MODEL_NAME = "1.5B"  # or your preferred model
+SCRIPT_PATH = "long_script.txt"  # path to your input script
+OUTPUT_PATH = "output_audio.wav"
+CHUNK_SIZE = 500  # number of tokens or characters per chunk (adjust for GPU)
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+# -----------------------
+# Load model
+# -----------------------
+print(f"Loading VibeVoice model ({MODEL_NAME})...")
+model = load_model(MODEL_NAME).to(DEVICE)  # Replace with actual load function
+model.eval()
+
+# -----------------------
+# Load script
+# -----------------------
+with open(SCRIPT_PATH, "r", encoding="utf-8") as f:
+    script_text = f.read()
+
+# -----------------------
+# Generate audio in chunks
+# -----------------------
+def generate_long_script(model, text, chunk_size=CHUNK_SIZE):
+    """Split long script and generate audio chunk by chunk."""
+    # Simple character-based chunking
+    chunks = [text[i:i+chunk_size] for i in range(0, len(text), chunk_size)]
+    audio_outputs = []
+
+    for i, chunk in enumerate(chunks):
+        print(f"Generating chunk {i+1}/{len(chunks)}...")
+        # Replace with your actual model.generate() or inference function
+        audio_chunk = model.generate(chunk)  # Example placeholder
+        audio_outputs.append(audio_chunk.cpu())
+
+    # Concatenate audio tensors
+    final_audio = torch.cat(audio_outputs)
+    return final_audio
+
+# -----------------------
+# Generate and save
+# -----------------------
+final_audio = generate_long_script(model, script_text)
+
+# Convert tensor to WAV file
+import soundfile as sf
+sf.write(OUTPUT_PATH, final_audio.numpy(), 22050)  # adjust sample rate if needed
+
+print(f"Audio generation complete! Saved to {OUTPUT_PATH}")