# VibeVoice 7B - 4-bit Quantized

Optimized for RTX 3060/4060 and similar 12GB VRAM GPUs.

## Specifications
- Quantization: 4-bit (nf4)
- Model size: 6.2 GB
- VRAM usage: ~8 GB
- Quality: Very good (minimal degradation)

## Usage

```python
from vibevoice.modular.modeling_vibevoice_inference import VibeVoiceForConditionalGenerationInference
from vibevoice.processor.vibevoice_processor import VibeVoiceProcessor

model = VibeVoiceForConditionalGenerationInference.from_pretrained(
    "Dannidee/VibeVoice7b-low-vram/4bit",
    device_map='cuda',
    torch_dtype=torch.bfloat16,
)
processor = VibeVoiceProcessor.from_pretrained("Dannidee/VibeVoice7b-low-vram/4bit")

# Generate speech
text = "Speaker 1: Hello! Speaker 2: Hi there!"
inputs = processor(
    text=[text],
    voice_samples=[["voice1.wav", "voice2.wav"]],
    padding=True,
    return_tensors="pt",
)

outputs = model.generate(**inputs)
processor.save_audio(outputs.speech_outputs[0], "output.wav")
```