This project mainly integrates several LLMs quantization methods derived from GPTQ

Include:

Methods	Quantize	PPL Eval	Task Eval	Save
GPTQ	✅	✅	TODO	TODO
PB-LLM	✅	✅	TODO	TODO
BiLLM	✅	✅	TODO	TODO
CLAQ	✅	✅	TODO	TODO
Z-Fold	✅	✅	TODO	TODO
decoupleQ	TODO	TODO	TODO	TODO
Slim	✅	TODO	TODO	TODO
Slim Plus	TODO	TODO	TODO	TODO
QuIP	✅	✅	TODO	TODO

GPTQ for LLaMA families

python llama.py --method gptq --model ${MODEL_DIR} --dataset c4 --wbits 4 --true-sequential --act-order --save

Binarization for LLaMA families

python llama.py --method billm --model ${MODEL_DIR} --dataset c4 --low_quant_method braq --blocksize 128 --salient_metric hessian --save

Z-fold for LLaMA families

python llama.py --method zfold --model ${MODEL_DIR} --dataset c4 --wbits 4 --salient_metric hessian --act-order --use_zfold --save

CLAQ for LLaMA families

# Run single-precision quantization and compute PPL results
python llama.py --method claq --model ${MODEL_DIR} --dataset c4 --wbits 4 --true-sequential --act-order  --save

# Run Adaptive Precision quantization to quantize the model to 2.1 bit and compute PPL results
python llama.py --method claq --model ${MODEL_DIR} --dataset c4 --wbits 2 --true-sequential --act-order --outlierorder 2.1 --save

# Run Outlier Reservation quantization to keep 0.07 bit of full-precison outliers and compute PPL results
python llama.py --method claq --model ${MODEL_DIR} --dataset c4 --wbits 2 --true-sequential --act-order --outlier_col_dynamic --save

# Run Adaptive Precision + Outlier Reservation quantization to quantize the model to 2.12 bit and compute PPL results
python llama.py --method claq --model ${MODEL_DIR} --dataset c4 --wbits 2 --true-sequential --act-order --outlierorder 2.05 --outlier_col_dynamic

PB-LLM for LLaMA families

python llama.py --method pbllm --model ${MODEL_DIR} --dataset c4 --low_quant_method xnor --low_frac 0.5 --high_bit 8 --salient_metric hessian --save

python llama.py --method pbllm --model ${MODEL_DIR} --dataset c4 --low_quant_method xnor --low_frac 0.8 --high_bit 8 --salient_metric hessian --save

python llama.py --method pbllm --model ${MODEL_DIR} --dataset c4 --low_quant_method xnor --low_frac 0.9 --high_bit 8 --salient_metric hessian --save

python llama.py --method pbllm --model ${MODEL_DIR} --dataset c4 --low_quant_method xnor --low_frac 0.95 --high_bit 8 --salient_metric hessian --save

DecoupleQ for LLaMA families

python llama.py --method decoupleq --model ${MODEL_DIR} --dataset c4 --true-sequential --act-order --new-eval \
    --wbits 2 \
    --group-size -1 \
    --max-iter-num 4 \
    --iters-before-round 200 \
    --inner-iters-for-round 5 \
    --blockwise-minimize-epoch 4 \
    --round-fn gptq \
    --blockwise-minimize-lr 1.0e-5 \
    --train-LN \
    --save

QuIP

# Compute full precision (FP16) results
python llama.py --method quip --model ${MODEL_DIR} --dataset c4
# Run a quantization method with baseline processing
python llama.py --method quip --model ${MODEL_DIR} --dataset c4 --wbits 4 --quant gptq --pre_gptqH --save 

SliM

# W2A16G128
python llama.py --method slim --model ${MODEL_DIR} --dataset wikitext2 --wbits 4 --groupsize 128 --tasks piqa,arc_easy,arc_challenge,boolq,hellaswag,winogrande

# W3A16G128
python llama.py --method slim --model ${MODEL_DIR} --dataset wikitext2 --wbits 3 --groupsize 128 --tasks piqa,arc_easy,arc_challenge,boolq,hellaswag,winogrande