mGPT-quantized

The concept: 8-bit quantized version of mGPT-13B, an LLM released by AI-Forever / Sberbank AI in 2022-2023.

On the GPT scale, it is between the # of parameters for GPT-2 and GPT-3, but comparison is tricky after training on 60+ languages.

My goal is to evaluate this on Hindi and Indonesian tasks, where there are fewer autoregressive language models in this size range.

For English: use a GPT model or LLaMa2-7B

For Arabic: in August 2023 I would recommend the bilingual JAIS model, which is also 13B parameters can be quantized.

In August 2023 AI-Forever added 1.3B-param models for 20+ languages. If your language is Mongolian, for example, it might be better to use mGPT-1.3B-mongol and not this one.

They also have a 1.3B param model for all languages, which I further quantized here: https://huggingface.co/monsoon-nlp/mGPT-quantized

How was the model created?

Quantization of mGPT-13B was done using bitsandbytes library, CoLab Pro with an A100 GPU, and a lot of space on Google Drive.

from transformers import BitsAndBytesConfig, GPT2LMHeadModel

quantization_config = BitsAndBytesConfig(
    load_in_8bit=True,
    bnb_8bit_compute_dtype=torch.bfloat16,
    bnb_8bit_use_double_quant=True,
    bnb_8bit_quant_type="nf4",
)

qmodel = GPT2LMHeadModel.from_pretrained(
    "ai-forever/mGPT-13B",
    load_in_8bit=True,
    torch_dtype=torch.bfloat16,
    quantization_config=quantization_config,
    device_map="auto"
)

qmodel.save_pretrained("model_name")

Future steps

mGPT could be further quantized (4-bit), but model.save_pretrained() currently throws a NotImplementedError error.