Model Information

Quantized version of ibm-granite/granite-3.3-2b-instruct using torch.float32 for quantization tuning.

4 bits (INT4)
group size = 128
Symmetrical Quantization
Method WoQ (AutoRound format)

Fast and low memory, 2-3X speedup (slight accuracy drop at W4G128)

Quantization framework: Intel AutoRound v0.4.7

Note: this INT4 version of granite-3.3-2b-instruct has been quantized to run inference through CPU.

Replication Recipe

Step 1 Install Requirements

I suggest to install requirements into a dedicated python-virtualenv or a conda enviroment.

wget https://github.com/intel/auto-round/archive/refs/tags/v0.4.7.tar.gz
tar -xvzf v0.4.7.tar.gz
cd auto-round-0.4.7
pip install -r requirements-cpu.txt --upgrade

Step 2 Build Intel AutoRound wheel from sources

pip install -vvv --no-build-isolation -e .[cpu]

Step 3 Script for Quantization

  from transformers import AutoModelForCausalLM, AutoTokenizer
  model_name = "ibm-granite/granite-3.3-2b-instruct"
  model = AutoModelForCausalLM.from_pretrained(model_name)
  tokenizer = AutoTokenizer.from_pretrained(model_name)
  from auto_round import AutoRound
  bits, group_size, sym, device = 4, 128, True, 'cpu'
  autoround = AutoRound(model, tokenizer, nsamples=128, iters=200, seqlen=512, batch_size=4, bits=bits, group_size=group_size, sym=sym, device=device)
  autoround.quantize()
  output_dir = "./AutoRound/ibm-granite_granite-3.3-2b-instruct-autoround-int4-gs128-sym"
  autoround.save_quantized(output_dir, format='auto_round', inplace=True)