metadata

license: apache-2.0
language:
  - ar
  - hi
  - id
pipeline_tag: text-generation
tags:
  - multilingual

mGPT-quantized

The concept: 8-bit quantized version of mGPT, a 1.3B param model released by AI-Forever / Sberbank AI in April 2022.

On the GPT scale, it is a similar # of parameters to GPT2-XL, but on 60+ languages.

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

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

AI-Forever also released a 13B-parameter model, and in August 2023 added 1.3B-param models for about 1/3 of the model languages. If your language is Mongolian, for example, use mGPT-1.3B-mongol and not this one.

How was the model created?

Quantization of mGPT 1.3B was done using bitsandbytes library:

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",
    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.
It would be great to load and quantize the 10x larger mGPT-13B, but that would take more resources.