Skip to content

DeepSparkHub 25.09 Release Notes

Choose a tag to compare

@rexlee8776 rexlee8776 released this 26 Sep 02:22
· 77 commits to master since this release

25.09 Release Notes

模型与算法

  • 新增了10个大模型训练微调示例,其中7个为强化学习相关示例,使用了OpenRLHF和verl工具箱
大模型
Qwen2-7B GRPO (verl) Qwen2.5-VL-7B GRPO (verl) Qwen3-8B GRPO (verl)
DeepSeek-LLM-7B PPO (verl) Gemma-2-2B-IT PPO (verl)
Llama-3-8B SFT (OpenRLHF) Mixtral-8x7B-v0.1 SFT (OpenRLHF)
CosyVoice2-0.5B (DeepSpeed) Llama-3-8B (Megatron-LM) Qwen2.5-3B (Colossal-AI)

修复更新

  • 修复了BERT TensorFlow模型在TensorFlow升级后不支持estimator的问题(#ICWA6Y)
  • 修复了ResNet50 TensorFlow模型在TensorFlow升级后,不兼容normalization_layers的问题(#ICWA59)
  • 修复了ResNet50 Pytorch模型使用DALI情况下的steps数异常问题 (#ICV31B)
  • 更新了AlexNet、GoogLeNet等模型所支持的最高SDK版本
  • 同步了Colossal-AI、Deepspeed和Megatron-Deepspeed框架的最新代码

版本关联

DeepSparkHub 25.09版本中使用verl的模型对应IXUCA SDK dev-only版本,其余模型对应SDK 4.3.0。