README.md

---
license: mit
datasets:
- wikitext
---

[gpt2-xl](https://huggingface.co/openai-community/gpt2-xl) quantized to 4-bit using [AutoGPTQ](https://github.com/AutoGPTQ/AutoGPTQ).

To use, first install AutoGPTQ:

```shell
pip install auto-gptq
```

Then load the model from the hub:
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig

model_name = "smpanaro/gpt2-AutoGPTQ-4bit-128g"
model = AutoGPTQForCausalLM.from_quantized(model_name, use_triton=True)
# Note: despite this model being quantized only using groups and desc_act=False, Triton still seems to be required.
```


|Model|4-Bit Perplexity|16-Bit Perplexity|Delta|
|--|--|--|--|
|[smpanaro/gpt2-AutoGPTQ-4bit-128g](https://huggingface.co/smpanaro/gpt2-AutoGPTQ-4bit-128g)|26.5000|25.1875|1.3125|
|[smpanaro/gpt2-medium-AutoGPTQ-4bit-128g](https://huggingface.co/smpanaro/gpt2-medium-AutoGPTQ-4bit-128g)|19.1719|18.4739|0.698|
|[smpanaro/gpt2-large-AutoGPTQ-4bit-128g](https://huggingface.co/smpanaro/gpt2-large-AutoGPTQ-4bit-128g)|16.6875|16.4541|0.2334|
|smpanaro/gpt2-xl-AutoGPTQ-4bit-128g|14.9297|14.7951|0.1346|
<sub>Wikitext perplexity measured as in the [huggingface docs](https://huggingface.co/docs/transformers/en/perplexity), lower is better</sub>