---
pipeline_tag: text-generation
license: mit
base_model: openai-community/gpt2-large
library_name: kerasformers
language:
- en
tags:
- keras
- kerasformers
- gpt2
- gpt-2
- text-generation
- pytorch
- jax
- tf
---

# Run GPT-2 with Keras 3: JAX, PyTorch, or TensorFlow

[![GitHub](https://img.shields.io/badge/GitHub-KerasFormers-black?logo=github)](https://github.com/IMvision12/KerasFormers) [![Docs](https://img.shields.io/badge/Docs-GPT--2-blue)](https://imvision12.github.io/KerasFormers/gpt2/)

# kerasformers/gpt2_large

Paper: [Language Models are Unsupervised Multitask Learners (Radford et al., 2019)](https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf)

GPT-2 is OpenAI's decoder-only transformer language model trained on WebText:
learned absolute position embeddings, pre-LayerNorm blocks, `gelu_new`
activations, a tied output head, and a byte-level BPE tokenizer. This is the
**774M** variant, a base completion model (no chat template).

For more details, see the upstream [model card](https://huggingface.co/openai-community/gpt2-large).

Pure-**Keras 3** conversion of [`openai-community/gpt2-large`](https://huggingface.co/openai-community/gpt2-large) for
[kerasformers](https://github.com/IMvision12/KerasFormers). One implementation runs unmodified on
**TensorFlow / Torch / JAX**.

## Quick start

```python
import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from kerasformers.models.gpt2 import GPT2TextGenerate, GPT2Tokenizer

model = GPT2TextGenerate.from_weights("kerasformers/gpt2_large")
tokenizer = GPT2Tokenizer.from_weights("kerasformers/gpt2_large")

inputs = tokenizer("The meaning of life is")
outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0]))
```

All GPT-2 sizes load the same way with `from_weights("kerasformers/<variant>")`:

| Variant | Hub | Params |
|---|---|---|
| `gpt2` | [`kerasformers/gpt2`](https://huggingface.co/kerasformers/gpt2) | 124M |
| `gpt2_medium` | [`kerasformers/gpt2_medium`](https://huggingface.co/kerasformers/gpt2_medium) | 355M |
| `gpt2_large` | [`kerasformers/gpt2_large`](https://huggingface.co/kerasformers/gpt2_large) | 774M |
| `gpt2_xl` | [`kerasformers/gpt2_xl`](https://huggingface.co/kerasformers/gpt2_xl) | 1.5B |

## Tips

- Set `KERAS_BACKEND` **before** importing Keras / kerasformers.
- This is a base completion model: it continues a prompt and is not
  instruction-tuned.
- See the [GPT-2 docs](https://imvision12.github.io/KerasFormers/gpt2/) and [Loading Weights](https://imvision12.github.io/KerasFormers/loading_weights/).
- Upstream safetensors still work via the `hf:` prefix, e.g.
  `GPT2TextGenerate.from_weights("hf:openai-community/gpt2-large")`.

## Special Thanks

A huge thank you to the OpenAI GPT-2 authors for creating and releasing these models.

License: MIT, inherited from the upstream OpenAI GPT-2 release.