Starstruk Logo MarkStarstruk

v3.2.2

SagerNet/sing-boxv3.2.2Dec 8, 2025by XiaoguangHu01

AI Summary

PaddlePaddle 3.2.2 focuses on distributed training optimization, checkpoint loading flexibility (HuggingFace support), and bug fixes for safetensors and BFloat16 precision.

Key Highlights

  • Optimized FlexCheckpoint re-sharding communication
  • Added full interface to paddle.nn.Layer
  • Support for loading HuggingFace format checkpoints
  • Fixed safetensor loading errors
  • Fixed BFloat16 to_tensor precision issues

New Features

  • FlexCheckpoint optimization
  • Layer full interface
  • HuggingFace support
  • Safetensor fix
  • BFloat16 fix

Full Release Notes

# 重要更新
飞桨框架3.2.2版本在分布式并行、算子机制、硬件适配三个方面完成多项优化与升级，进一步提升框架性能与稳定性。

## 1. 分布式训练
- 优化 FlexCheckpoint 的重切分通信流程; 为 paddle.nn.Layer 新增 full 接口, 用于返回完整模型参数; 支持加载 HuggingFace 开源格式的 Checkpoint。([#76249](https://github.com/PaddlePaddle/Paddle/pull/76249), [#76291](https://github.com/PaddlePaddle/Paddle/pull/76291))
- 为 group_sharded_optimizer_stage2 优化器新增 sharded_state_dict 函数。[#76311](https://github.com/PaddlePaddle/Paddle/pull/76311)
- 为 paddle.load 接口修复加载 safetensor 文件 device_id 参数错误及 core_dump 问题。[#76317](https://github.com/PaddlePaddle/Paddle/pull/76317)
- 新增 PipelineDatasetPreprocessor 机制，消除流水线并行策略中可能出现的内存泄漏问题。 [#76260](https://github.com/PaddlePaddle/Paddle/pull/76260)

## 2. 算子机制
- 修复针对 BFloat16 list 场景下的 to_tensor 精度问题。 [#76242](https://github.com/PaddlePaddle/Paddle/pull/76242)

## 3. 硬件适配
- 修改了独立的 XPU 内存监控模块，以确保与最新的内存监控逻辑保持一致。 [#76056](https://github.com/PaddlePaddle/Paddle/pull/76056)

## 4. 贡献者名单
qw86972190, xingmingyyj, zhangbo9674, zhangyuqin1998

# Important Updates

PaddlePaddle Framework version 3.2.2 features multiple optimizations and upgrades across Distributed Parallelism, Operator Mechanism, and Hardware Adaptation to further enhance the framework's performance and stability.

## 1. Distributed Training
- Optimized the communication process for re-sharding in FlexCheckpoint; added the full interface to paddle.nn.Layer for returning complete model parameters; supported loading Checkpoints in the HuggingFace open-source format. ([#76249](https://github.com/PaddlePaddle/Paddle/pull/76249), [#76291](https://github.com/PaddlePaddle/Paddle/pull/76291))
- Added the sharded_state_dict function to the group_sharded_optimizer_stage2 optimizer. [#76311](https://github.com/PaddlePaddle/Paddle/pull/76311)
- Fixed errors regarding the device_id parameter and a core dump issue when loading safetensor files using the paddle.load interface. [#76317](https://github.com/PaddlePaddle/Paddle/pull/76317)
- Introduced the PipelineDatasetPreprocessor mechanism to eliminate potential memory leak issues in the pipeline parallelism strategy. [#76260](https://github.com/PaddlePaddle/Paddle/pull/76260)

## 2. Operator Mechanisms
- Fixed a precision issue in to_tensor for BFloat16 list scenarios. [#76242](https://github.com/PaddlePaddle/Paddle/pull/76242)

## 3. Hardware Adaptation
- Modified the independent XPU memory monitoring module to ensure consistency with the latest memory monitoring logic. [#76056](https://github.com/PaddlePaddle/Paddle/pull/76056)

## 4. List of Contributors
qw86972190, xingmingyyj, zhangbo9674, zhangyuqin1998