PaddleFormers v0.4

PaddleFormers 0.4 is officially released! This release introduces several key features and improvements:

✨ New Features

1. Unified CLI Training Workflow

PaddleFormers provides a streamlined training workflow through a unified CLI interface with YAML-based configuration. Users can launch a complete training pipeline with a single command, significantly reducing boilerplate code and accelerating iteration.
It supports CPT, SFT, and DPO training paradigms, covering both full-parameter fine-tuning and LoRA-based fine-tuning.

2. New model support

Added support for the following models:

• deepseek_v3
• llama

3. FlexCheckpoint Enhancements

FlexCheckpoint introduces initial support for Qwen3-MoE and GLM-4.5-MoE, enabling flexible loading and saving of Hugging Face–compatible model weights across diverse distributed training configurations. Model checkpoints produced during training can be directly deployed to mainstream inference frameworks such as vLLM, SGLang, and FastDeploy, without any additional conversion or adaptation. This enables a seamless end-to-end pipeline from training to inference deployment.