Awesome

SpinQuant

This repository contains the code of SpinQuant introduced in our work: "SpinQuant: LLM Quantization with Learned Rotations"

In this work, we found that

1. Rotation is a principle way to remove outliers in the LLMs and assist quantization;
2. Not all rotation helps equally and random rotations produce a large variance in quantized models;
3. Learning rotation with Cayley optimization greatly enhance the final performance.

As a result, SpinQuant narrows the accuracy gap of W4A4KV4 quantization with full precision to merely 2.9 points for the LLaMA-2 7B model on zero-shot reasoning tasks, surpassing LLM-QAT by 19.1 points and SmoothQuant by 25.0 points.

News

• Oct 24, 2024: 🚀 SpinQuant supported the live demo in Meta Connect [post]. We also made our 4-bit quantized LLaMA SpinQuant model publicly available [post] [Llama-3.2-1B-Instruct-SpinQuant_INT4] [Llama-3.2-3B-Instruct-SpinQuant_INT4].

https://github.com/user-attachments/assets/e7bebfd5-ef13-440f-8066-7bf9205ad309

Citation

If you find our code useful for your research, please consider citing:

@article{liu2024spinquant,
    title={SpinQuant--LLM quantization with learned rotations},
    author={Liu, Zechun and Zhao, Changsheng and Fedorov, Igor and Soran, Bilge and Choudhary, Dhruv and Krishnamoorthi, Raghuraman and Chandra, Vikas and Tian, Yuandong and Blankevoort, Tijmen},
    journal={arXiv preprint arXiv:2405.16406},
    year={2024}
}

Run

1. Requirements:

• python 3.9, pytorch >= 2.0
• install pytorch with cuda from https://pytorch.org/get-started/locally/, it is prerequisite for fast-hadamard-transform package.
• pip install -r requirement.txt
• git clone https://github.com/Dao-AILab/fast-hadamard-transform.git
• cd fast-hadamard-transform
• pip install .

2. Steps to run:

For the scripts here, set output_rotation_path output_dir logging_dir optimized_rotation_path to your own locations. For gated repo such as meta-llama, you can set your HF token to access_token.

Step 1: Optimize Rotation Matrix

• For LLaMA-2 7B/13B and LLaMA-3 8B models:
  bash 10_optimize_rotation.sh $model_name $w_bit $a_bit $kv_bit
  e.g., bash scripts/10_optimize_rotation.sh meta-llama/Llama-2-7b 4 4 4 for 4-bit weight 4-bit activation and 4-bit kv-cache on Llama-2-7b model.
• For LLaMA-2 70B and LLaMA-3 70B models:
  bash 11_optimize_rotation_fsdp.sh $model_name $w_bit $a_bit $kv_bit
  e.g., bash scripts/11_optimize_rotation_fsdp.sh meta-llama/Llama-2-70b 4 4 4 for 4-bit weight 4-bit activation and 4-bit kv-cache on Llama-2-70b model.

Step 2: Run PTQ evaluation with optimized rotation

After obtaining the optimized_rotation, put the rotation matrix into optimized_rotation_path for evaluation.

• bash scripts/2_eval_ptq.sh $model_name $w_bit $a_bit $kv_bit

3. Export to ExecuTorch

We also support exporting the quantized model to ExecuTorch, which allows us to utilize the quantization kernels and achieve real-time speedup. For more information on kernel implementation details, please see ExecuTorch, and ExecuTorch with SpinQuant. We currently support 4-bit weight (set group-size to 256 for 8B model and to 32 for smaller model) and 8-bit dynamic activation quantization.

To obtain ExecuTorch-compatible quantized models, you can use the following scripts:

• bash scripts/31_optimize_rotation_executorch.sh $model_name
• bash scripts/32_eval_ptq_executorch.sh $model_name

We also provide an example colab notebook to train and export ExecuTorch compatiable Llama 3.2 models

Note

• If using GPTQ quantization method in Step 2 for quantizing both weight and activations, we optimize the rotation matrices with respect to a network where only activations are quantized.
  e.g. bash 10_optimize_rotation.sh meta-llama/Llama-2-7b 16 4 4 followed by bash 2_eval_ptq.sh meta-llama/Llama-2-7b 4 4 4 with the --optimized_rotation_path pointing to the rotation optimized for W16A4KV4.

Arguments

• --input_model: The model name (or path to the weights)

• --output_rotation_path: The local path we want to store the oprimized rotation matrix

• --per_device_train_batch_size: The batch size for rotation optimization

• --per_device_eval_batch_size: The batch size for PPL evaluation

• --a_bits: The number of bits for activation quantization

• --w_bits: The number of bits for weight quantization

• --v_bits: The number of bits for value quantization

• --k_bits: The number of bits for key quantization

• --w_clip: Whether using the grid search to find best weight clipping range

• --w_rtn: Whether we want to use round-to-nearest quantization. If not having --w_rtn, we are using GPTQ quantization.

• --w_groupsize: The group size for group-wise weight quantization.

• --rotate: Whether we want to rotate the model

• --optimized_rotation_path: The checkpoint path of optimized rotation; Use random rotation if path is not given

Quantized Models

You can download the optimized rotation matrices here.

Acknowledgement

The results reported in the paper is run with the internal LLaMA codebase in Meta. We reproduced our experiments with HuggingFace codebase and released code here, which partially based on HuggingFace transformers, QuaRot, QuIP# and Optimization-on-Stiefel-Manifold-via-Cayley-Transform. SpinQuant is available in LLMC, an Efficient LLM Compression Toolkit.

Contact

Zechun Liu, Reality Labs, Meta Inc (zechunliu at meta dot com)

Changsheng Zhao, Reality Labs, Meta Inc (cszhao at meta dot com)

Relevant Projects

MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases [Paper] [Code]

LLM-QAT: Data-Free Quantization Aware Training for Large Language Models [Paper] [Code]

License

BiT is CC-BY-NC 4.0 licensed as of now.