Note: This repository contains quantization algorithm and the model evaluation code for SpQR method for LLM compression; The efficient inference code will be added soon.

It accompanies the research paper "SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression" .

To run SpQR with falcon make sure that you have torch>=2.0.0 with CUDA support.

Install packages from requirements.txt:

pip install -r requirements.txt

Note: the results reported in the ArXiv paper where obtained using 4.28.dev0 version of transformers, commit id 464d420775.

The script will require downloading and caching locally the relevant tokenizer and the datasets. They will be saved in default Huggingface Datasets directory unless alternative location is provided by env variables. See relevant Datasets documentation section

This repository is expected to work with models of LLaMA, Falcon and OPT families so far.

For quantization with SpQR its is recommended to use the subset of the data model was trained on. I.e. for quantization of LLaMA models we recommend to use the subset of RedPajama and for Falcon quantization - RefinedWeb.Both subsets are stored in data directory:

• data/red_pajama_n=1024.pth
• data/refined_web_n=128.pth

Note These subsets are already processed with the corresponding model tokenizer. Use for different model will lead to unexpected behavior.

For OPT following GPTQ paper we recommend to use c4.

For the sake of convenience one can optionally log the data to Weights and Biases service (wandb). Run pip install wandb for W&B logging. Specify $WANDB_ENTITY, $WANDB_PROJECT, $WANDB_NAME environment variables prior to running experiments. use --wandb argument to enable logging

This code was developed and tested using a single A100 GPU with 80GB GPU RAM. It may successfully run on GPUs with 32GB+ VRAM for perplexity evaluation of up to LLaMA-65B and Falcon-40B models. With --offload activations option, the model perplexity may be evaluated on machines with less VRAM: 24GB+ for Llama 65B and 6GB+ for Llama 7B. The perplexity testing code also requires RAM amount sufficient to hold uncompressed model weights (e.g. ~130GB for Llama65B) and testing datasets. For Language Model Evaluation Harness evaluation one needs to have enough memory to load whole model on one or several devices + activation tensors.

The code requires the LLaMA model to be downloaded in Huggingface format and saved locally. The scripts below assume that $TRANSFORMERS_CACHE variable points to the Huggingface Transformers cache folder.

This script compresses the model and then tests its performance in terms of perplexity using WikiText2, C4, and Penn Treebank datasets.

The command to launch the script should look like this:

export MODEL_PATH=<PATH_TO_MODEL_DIR>
export DATASET=<INSERT DATASET NAME OR PATH TO CUSTOM DATA>

python main.py $MODEL_PATH $DATASET \
    --wbits 4 \
    --groupsize 16 \
    --perchannel \
    --qq_scale_bits 3 \
    --qq_zero_bits 3 \
    --qq_groupsize 16 \
    --outlier_threshold=0.2 \
    --permutation_order act_order \
    --percdamp 1e0 \
    --nsamples 128 

The command above runs near-lossless compression as described in the article. Adjusting the above parameters allows for tighter compression with a slightly greater loss.

Note the launch arguments:

• <PATH_TO_MODEL_DIR> - path to model folder, which contains config.json
• one of [c4, ptb, wikitext2, pajama, refinedweb, none] -- name of dataset to use for compression, or path to an alternative preprocessed and tokenized dataset.
• --wbits 3 -- number of bits for quantized weights representation
• --groupsize 16 -- size of first-order groups for compression
• --qq_groupsize 16 -- size of second-order (quantized) groups for compression
• --qq_scale_bits 3 --qq_zero_bits 3 -- bit sizes for quantizing first order weights' scale and zeros.
• --offload activations -- moves activations to RAM when not used. Reduces VRAM usage while slowing work by ~10%. run python main.py --help for more details on command line arguments, including compression parameters.
• --save --load -- path to save/load quantized model.

To perform zero-shot evaluation, we use Language Model Evaluation Harness framework with slight modifications. This repository contains a copy of LM Evaluation Harness repo from early 2023 in lm-eval-harness folder.

Before running the code make sure that you have all the requirements and dependencies of lm-eval-harness installed. To install them run:

pip install -r lm-evaluation-harness/requirements.txt

The main script launching the evaluation procedure is lmeval.py .

Note. Current version of the script support only LLaMA/Falcon quantization. Therefore, set:

• --model=hf-causal
• --model_args pretrained=$MODEL_PATH where $MODEL_PATH has to be one of the LLaMA models

--quantization_args - list of comma separated arguments for quantizer. For details and options refer to spqr_config.py.

Below is presented an example of benchmark launch.

export MODEL_PATH=<INSERT PATH_TO_MODEL_DIR>
export DATASET=<INSERT DATASET NAME OR PATH TO CUSTOM DATA>

python lmeval.py \
    --model hf-causal \
    --model_args pretrained=$MODEL_PATH,dtype=float16,use_accelerate=True \
    --quantization_args dataset=$DATASET,wbits=4,groupsize=16,perchannel=True,qq_scale_bits=3,qq_zero_bits=3,qq_groupsize=16,percdamp=1.0,outlier_threshold=0.2,simplified_outliers=False,nsamples=128,offload_activations=True \
    --tasks winogrande,piqa,hellaswag,arc_easy,arc_challenge \
    --batch_size 1

Performance and runtime notes:

• For large models (LLaMA-30B, LLaMA-65B) specify max_memory_per_gpu={value}GIB so that there are free 15-20GIB of GPU memory for each GPU to store activations for calibration.
• offload_activations=True slightly reduces peak memory consumption
• Typically LlaMA-30B requires 1-2 A100 GPUs with 80Gb of memory and LlaMA-65B requires 3 A100 with 80Gb each.
• With enough spare GPU memory, one can raise batch size to accelerate evaluation process.

@misc{dettmers2023spqr,
      title={SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression}, 
      author={Tim Dettmers and Ruslan Svirschevski and Vage Egiazarian and Denis Kuznedelev and Elias Frantar and Saleh Ashkboos and Alexander Borzunov and Torsten Hoefler and Dan Alistarh},
      year={2023},
      eprint={2306.03078},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}