EETQ

The Easy & Efficient Quantization for Transformers (EETQ) library supports int8 weight-only per-channel quantization for NVIDIA GPUs. It uses high-performance GEMM and GEMV kernels from FasterTransformer and TensorRT-LLM. The attention layer is optimized with FlashAttention2. No calibration dataset is required, and the model doesn’t need to be pre-quantized. Accuracy degradation is negligible owing to the per-channel quantization.

EETQ further supports fine-tuning with PEFT.

Install EETQ from the release page or source code. CUDA 11.4+ is required for EETQ.

pip install --no-cache-dir https://github.com/NetEase-FuXi/EETQ/releases/download/v1.0.0/EETQ-1.0.0+cu121+torch2.1.2-cp310-cp310-linux_x86_64.whl

Quantize a model on-the-fly by defining the quantization data type in EetqConfig.

from transformers import AutoModelForCausalLM, EetqConfig

quantization_config = EetqConfig("int8")
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B",
    torch_dtype="auto",
    device_map="auto",
    quantization_config=quantization_config
)

Save the quantized model with save_pretrained() so it can be reused again with from_pretrained().

quant_path = "/path/to/save/quantized/model"
model.save_pretrained(quant_path)
model = AutoModelForCausalLM.from_pretrained(quant_path, device_map="auto")