fbaldassarri/iGeniusAI_Italia-9B-Instruct-v0.1-autoround-int8-gs64-auto-asym

Model Information

Quantized version of iGeniusAI/Italia-9B-Instruct-v0.1 using torch.float32 for quantization tuning.

• 8 bits (INT8)
• group size = 64
• Asymmetrical Quantization
• Method WoQ (AutoRound format)

Quantization framework: Intel AutoRound v0.4.6

Note: this INT8 version of Italia-9B-Instruct-v0.1 has been quantized using NVIDIA CUDA libraries.

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/v0.4.6.tar.gz
tar -xvzf v0.4.6.tar.gz
cd auto-round-0.4.6
pip install -r requirements.txt --upgrade

Step 2 Build Intel AutoRound wheel from sources

pip install -vvv --no-build-isolation -e .

Step 3 Script for Quantization

  from transformers import AutoModelForCausalLM, AutoTokenizer, GPTNeoXModel
  model_name = "iGeniusAI/Italia-9B-Instruct-v0.1"
  model = GPTNeoXModel.from_pretrained(model_name, trust_remote_code=True)
  tokenizer = AutoTokenizer.from_pretrained(model_name)
  from auto_round import AutoRound
  bits, group_size, sym, device, amp = 8, 64, False, 'auto', False
  autoround = AutoRound(model, tokenizer, nsamples=128, iters=200, seqlen=512, batch_size=4, bits=bits, group_size=group_size, sym=sym, device=device, amp=amp)
  autoround.quantize()
  output_dir = "./AutoRound/iGeniusAI_Italia-9B-Instruct-v0.1-autoround-int8-gs64-auto-asym"
  autoround.save_quantized(output_dir, format='auto_round', inplace=True)

Note: the GPTNeoXSdpaAttention class is deprecated in favor of simply modifying the config._attn_implementationattribute of the GPTNeoXAttention class. So this require transformers<4.48.

License

MIT

Disclaimer

This quantized model comes with no warranty. It has been developed only for research purposes.