VibeVoice-ASRを量子化してDERはどう変わるか？4bit・8bit・originalを実測比較

データセット : 日本語音声 10分
temperature  : 0
beam_size    : 3
repetition_penalty : 1.5
compute_type : fp16
GPU: RTX5090
評価指標 : DER (collar_sec=0.0, skip_overlap=False)

元モデル (fp16)                 : 16.16 GiB (model-*.safetensors 合計)
8bit量子化後 (vibevoice-asr-8bit): 9.3 GiB
4bit量子化後 (vibevoice-asr-4bit): 5.85 GiB

#!/usr/bin/env python
"""
Quantize VibeVoice ASR model with bitsandbytes and save it.

Usage:
    python quantize_vibevoice.py \
        --model_path microsoft/VibeVoice-ASR \
        --tokenizer_path Qwen/Qwen2.5-1.5B \
        --output_dir ./vibevoice-asr-4bit \
        --quantization 4bit
"""

import argparse
import importlib.metadata
from pathlib import Path

import torch
from transformers import BitsAndBytesConfig

from vibevoice.modular.modeling_vibevoice_asr import (
    VibeVoiceASRForConditionalGeneration,
)
from vibevoice.processor.vibevoice_asr_processor import VibeVoiceASRProcessor


def _parse_dtype(dtype_name: str) -> torch.dtype:
    table = {
        "bf16": torch.bfloat16,
        "fp16": torch.float16,
        "fp32": torch.float32,
    }
    if dtype_name not in table:
        raise ValueError(f"Unsupported dtype: {dtype_name}")
    return table[dtype_name]


def _ensure_bitsandbytes_installed() -> None:
    try:
        version = importlib.metadata.version("bitsandbytes")
    except importlib.metadata.PackageNotFoundError as exc:
        raise RuntimeError(
            "bitsandbytes is not installed.\n"
            "Install it with: uv pip install bitsandbytes"
        ) from exc
    print(f"Detected bitsandbytes=={version}")


def build_bnb_config(
    quantization: str, compute_dtype: torch.dtype
) -> BitsAndBytesConfig:
    if quantization == "4bit":
        return BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_use_double_quant=True,  # さらにメモリを削減
            bnb_4bit_compute_dtype=compute_dtype,
        )
    if quantization == "8bit":
        return BitsAndBytesConfig(load_in_8bit=True)
    raise ValueError(f"Unsupported quantization: {quantization}")


def quantize_and_save(
    model_path: str,
    tokenizer_path: str,
    output_dir: str,
    quantization: str,
    compute_dtype: torch.dtype,
    device_map: str,
) -> None:
    output_path = Path(output_dir)
    output_path.mkdir(parents=True, exist_ok=True)

    # VibeVoice固有: language_model_pretrained_name の指定が必要
    print(f"Loading processor from: {model_path}")
    processor = VibeVoiceASRProcessor.from_pretrained(
        model_path,
        language_model_pretrained_name=tokenizer_path,
    )

    print(f"Loading model in {quantization}...")
    bnb_config = build_bnb_config(
        quantization=quantization, compute_dtype=compute_dtype
    )

    model = VibeVoiceASRForConditionalGeneration.from_pretrained(
        model_path,
        quantization_config=bnb_config,
        device_map=device_map,
        trust_remote_code=True,
    )
    model.eval()

    print(f"Saving quantized model to: {output_dir}")
    model.save_pretrained(output_dir)
    processor.save_pretrained(output_dir)
    print("Done.")

# インストール
pip install bitsandbytes transformers accelerate torch

# 4bit量子化（NF4 + double quantization、compute_dtype=bf16）
python quantize_vibevoice.py \
    --model_path microsoft/VibeVoice-ASR \
    --tokenizer_path Qwen/Qwen2.5-1.5B \
    --output_dir ./vibevoice-asr-4bit \
    --quantization 4bit \
    --compute_dtype bf16

# 8bit量子化
python quantize_vibevoice.py \
    --model_path microsoft/VibeVoice-ASR \
    --tokenizer_path Qwen/Qwen2.5-1.5B \
    --output_dir ./vibevoice-asr-8bit \
    --quantization 8bit

from transformers import BitsAndBytesConfig
from vibevoice.modular.modeling_vibevoice_asr import VibeVoiceASRForConditionalGeneration
from vibevoice.processor.vibevoice_asr_processor import VibeVoiceASRProcessor

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
)

model = VibeVoiceASRForConditionalGeneration.from_pretrained(
    "./vibevoice-asr-4bit",
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)
processor = VibeVoiceASRProcessor.from_pretrained(
    "./vibevoice-asr-4bit",
    language_model_pretrained_name="Qwen/Qwen2.5-1.5B",
)

original : 22,909 MiB (ドラマ系)
4bit     : 24,077 MiB (ドラマ系)
8bit     : 29,597 MiB (ドラマ系)

生成時間の比較（ドラマ系）:
original : 99.47s  (1.0x)
8bit     : 224.39s (2.25x 遅い)
4bit     : 242.48s (2.44x 遅い)