Step by step example how to dump weights data for PyTorch model with Neural Insights

1. Introduction

2. Preparation

3. Running the quantization

Introduction

In this instruction weight data will be dumped using Neural Insights. PyTorch GPT-J-6B model will be used as an example.

Preparation

Source

First you need to install Intel® Neural Compressor.

# Install Neural Compressor
git clone https://github.com/intel/neural-compressor.git
cd neural-compressor 
pip install -r requirements.txt 
python setup.py install

# Install Neural Insights
pip install -r neural_insights/requirements.txt
python setup.py install neural_insights

Requirements

cd /examples/pytorch/nlp/huggingface_models/language-modeling/quantization/ptq_static/fx
pip install -r requirements.txt

Running the quantization

Before applying quantization, modify some code in run_clm.py file to enable Neural Insights:

1. Set the argument diagnosis to be True in PostTrainingQuantConfig so that Neural Insights will dump weights of quantizable Ops in this model.

conf = PostTrainingQuantConfig(
    accuracy_criterion=accuracy_criterion,
    diagnosis=True,
)

1. Quantize the model with following command:

python run_clm.py \
  --model_name_or_path EleutherAI/gpt-j-6B \
  --dataset_name wikitext\
  --dataset_config_name wikitext-2-raw-v1 \
  --do_train \
  --do_eval \
  --tune \
  --output_dir saved_results

Results would be dumped into nc_workspace directory in similar structure:

├── history.snapshot
├── input_model.pt
├── inspect_saved
│   ├── fp32
│   │   └── inspect_result.pkl
│   └── quan
│       └── inspect_result.pkl
├── model_summary.txt
└── weights_table.csv