PaddleFormers is a Transformer model library built on the PaddlePaddle deep learning framework, delivering both ease of use and high-performance capabilities. It provides a unified model definition interface, modular training components, and comprehensive distributed training strategies specifically designed for large language model development pipelines. This enables developers to train large models efficiently with minimal complexity, making it suitable for diverse scenarios ranging from academic research to industrial applications.

[2025年06月28日] 🎉 PaddleFormers 0.1 is officially released! This initial version supports SFT/DPO training paradigms, configurable distributed training via unified Trainer API, and integrates PEFT, MergeKit, and Quantization APIs for diverse LLM applications.

Implements 4D parallel strategies through unified Trainer API, lowering the barrier to distributed LLM training.

Integrates Packing dataflow and FlashMask operators for SFT/DPO training, eliminating padding waste and boosting throughput.

Features Unified Checkpoint storage tools for LLMs, enabling training resumption and dynamic resource scaling. Additionally implements asynchronous storage (up to 95% faster) and Optimizer State Quantization (78% storage reduction), ensuring industrial training meets both efficiency and stability requirements.

Requires Python 3.8+ and PaddlePaddle 3.1+.

# Install via pip
pip install paddleformers
# Install development version
git clone https://github.com/PaddlePaddle/PaddleFormers.git
cd PaddleFormers
pip install -e .

This example shows how to load Qwen model for text generation with PaddleFormers Auto API:

from paddleformers.transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B-Base")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B-Base", dtype="bfloat16", convert_from_hf=True)
input_features = tokenizer("Give me a short introduction to large language model.", return_tensors="pd")
outputs = model.generate(**input_features, max_new_tokens=128)
print(tokenizer.batch_decode(outputs[0], skip_special_tokens=True))

Getting started with supervised fine-tuning (SFT) using PaddleFormers:

from paddleformers.trl import SFTConfig, SFTTrainer
from datasets import load_dataset
dataset = load_dataset("ZHUI/alpaca_demo", split="train")
training_args = SFTConfig(output_dir="Qwen/Qwen3-0.6B-SFT", device="gpu", model_init_kwargs={"convert_from_hf": True})
trainer = SFTTrainer(
 args=training_args,
 model="Qwen/Qwen3-0.6B-Base",
 train_dataset=dataset,
)
trainer.train()

We welcome all contributions! See CONTRIBUTING.md for guidelines.

This repository's source code is available under the Apache 2.0 License.