/norm-tweaking

Post post-training-quantization (PTQ) method for improving LLMs. Unofficial implementation of https://arxiv.org/abs/2309.02784

Primary LanguagePython

Norm Tweaking

Unofficial and in-progress implementation of Norm Tweaking: High-performance Low-bit Quantization of Large Language Models. Norm Tweaking adjusts ("tweaks") the LayerNorm parameters to improve the accuracy of a quantized LLM. Many types of quantization are supported — norm tweaking is applied as a correcting step after quantization.

This repo supports:

  • Calibration Data Generation - Follows LLM-QAT's synthetic data generation strategy.
  • Norm Tweaking Algorithm - Adjusts layer norm parameters to improve model performance.

⚠️ NOTE: This repo is functional but does not yet replicate the results from the paper. PRs welcome! See the this discussion or the TODOs for places to contribute.

Get Started

pip install -e .
# Generate calibration data.
python examples/generate_data.py --model gpt2 --save_dir data/ --device cpu # or cuda:0
# Run the norm tweaking algorithm and save the resulting model.
python examples/tweak_model.py --model gpt2 --data data/normtweaking_calibration_data_gpt2_${TIMESTAMP}.safetensors

A norm-tweaked model will be saved in models/.

Grid search of the two tunable parameters is also supported:

# Install https://github.com/EleutherAI/lm-evaluation-harness from main (not from pip)
python examples/tweak_model.py \
  --model gpt2 \
  --data /content/data/normtweaking_calibration_data_gpt2_1024-Jan4-2024.safetensors \
  --save_dir models \
  --device cuda:0 \
  --grid-search-lr 0 0.0001 0.00001 0.000001 0.0000001 \
  --grid-search-scale 0.5 1 5 10

Grid search calculates wikitext perplexity (lower is better) for each setting and saves the best resulting models.

Evaluating Norm Tweaked Models To evaluate a saved model again or on a different dataset, use lm_eval. 
pip install lm-eval
lm_eval --model hf \
    --model_args pretrained=gpt2 \ # replace gpt2 with the path to your norm tweaked model
    --tasks wikitext \
    --device cuda:0 \
    --batch_size 8

Results

Currently this repo has only been tested on the smallest gpt2. With the correct hyperparameters there is a slight improvement in WikiText perplexity as reported by lm_eval.

model scheme lr scale float ppl quantized ppl tweaked ppl delta recovered (%)
gpt2† 6-bit per-channel RTN 2e-4 5 37.3698 37.9450 37.8323 0.1127 (19.5%)
gpt2 6-bit k-means 1e-5 20 37.3698 38.4925 38.4474 0.0451 (4.0%)
gpt2 4-bit k-means 1e-5 25 37.3698 77.8160 69.5995 8.2165 (20.3%)

† lm_eval reports different results than the gpt2 paper for perplexity. Using a different method that nearly matches the original paper's results yields: float 29.9389, quantized 30.6000, tweaked 30.3681, delta 0.2319 (35.1%).

TODO

  • Add a fast eval (delta mean, delta var like in the paper's graph)
  • Add a thorough eval (perplexity using lm-eval-harness)
  • Generate calibration by taking language into account as described in the paper (how?)
  • Support loading pre-quantized weights
  • Convert common formats to pre-quantized weights (gptq, awq, gguf?)
  • Support non-gpt2 models (compare results with the paper)
  • Try end-to-end tweaking (similar to OBD/SqueezeLLM)
  • Try tweaking the final layer norm (ln_f)
  • Try a larger calibration dataset (see AQLM arXiv:2401.06118)
  • Grid search num iters