Quantizating Llama-2-7b model using MatMulNBits quantizer

Note

For information on accessing Quark ONNX examples, refer to Accessing ONNX Examples. This example and the relevant files are available at onnx/weights_only_quantization/int4_matmul_nbits/llama2

This example describes how to quantize an Llama-2-7b model using the ONNX MatMulNBits quantizer of Quark including an option for HQQ algorithm.

Pip requirements

Install the necessary python packages:

python -m pip install -r requirements.txt

Prepare model

Download the HF Llama-2-7b-hf checkpoint. The Llama2 models checkpoint can be accessed by submitting a permission request to Meta. For additional details, see the Llama2 page on Huggingface. Upon obtaining permission, download the checkpoint to the Llama-2-7b-hf folder.

Export onnx model from Llama-2-7b-hf torch model:

mkdir models && optimum-cli export onnx --model ./Llama-2-7b-hf --task text-generation-with-past ./models/

Convert the float model to float16 due to reduce memory overhead:

mkdir fp16_models
cp -r models/*.json fp16_models
cp models/tokenizer.model fp16_models
python -m quark.onnx.tools.convert_fp32_to_fp16 --input models/model.onnx --output fp16_models/model.onnx --disable_shape_infer --save_as_external_data --all_tensors_to_one_file

Quantization

The quantizer takes the float16 model and produces a matmul_4bits quantized model:

mkdir matmul_4bits_quantized_models
cp -r fp16_models/*.json matmul_4bits_quantized_models
cp fp16_models/tokenizer.model matmul_4bits_quantized_models
python quantize_model.py --input_model_path fp16_models/model.onnx \
                         --output_model_path matmul_4bits_quantized_models/model.onnx \
                         --config MATMUL_NBITS

This command will generate a quantized model under the matmul_4bits_quantized_models folder, using the 4bits quantization configuration for MatMul.

Evaluation

Test the PPL of the float model on wikitext2.raw:

python onnx_validate.py --model_name_or_path models/ --per_gpu_eval_batch_size 1 --block_size 2048 --do_onnx_eval --no_cuda

Test the PPL of the matmul 4bits quantized model:

python onnx_validate.py --model_name_or_path matmul_4bits_quantized_models/ --per_gpu_eval_batch_size 1 --block_size 2048 --do_onnx_eval --no_cuda

	Float Model	MatMul 4bits Quantized Model
Model Size	26.0 G	3.5 G
PPL	5.6267	5.9228