v3.3.1

中文版

新特性

1. Agent训练部署模块引入agent template，包括hermes, glm4_0414, llama4等10余种agent template，支持agent数据集兼容不同模型的训练切换，文档参考这里。
2. GRPO训练支持调用外部vLLM server，训练与部署显存分配更灵活，训练脚本参考这里。

新模型

1. OpenGVLab/InternVL3-1B系列
2. moonshotai/Kimi-VL-A3B-Instruct系列
3. ZhipuAI/GLM-4-9B-0414, ZhipuAI/GLM-Z1-9B-0414系列

English Version

New Features

1. The Agent training and deployment module introduces agent templates, including more than 10 types such as hermes, glm4_0414, and llama4. These templates support switching between different models for agent dataset compatibility during training. For documentation, refer to here.
2. GRPO training now supports calling an external vLLM server, allowing for more flexible allocation of GPU memory during training and deployment. For the training script, refer to here.

New Models

1. OpenGVLab/InternVL3-1B series
2. moonshotai/Kimi-VL-A3B-Instruct series
3. ZhipuAI/GLM-4-9B-0414, ZhipuAI/GLM-Z1-9B-0414 series

What's Changed

• Fix sampling and rft by @tastelikefeet in #3847
• Fix incorrect retry count check in LazyLLMDataset.getitem by @IamLihua in #3845
• support internvl3 by @hjh0119 in #3842
• fix grpo filter overlong by @hjh0119 in #3844
• dapo-bug by @Evilxya in #3846
• support agent packing by @Jintao-Huang in #3853
• Fix internvl2.5/3 deepspeed packing by @Jintao-Huang in #3855
• fix multimodal target_modules by @Jintao-Huang in #3856
• Fix multimodal target modules by @Jintao-Huang in #3858
• Update FAQ by @slin000111 in #3841
• fix grpo completion length equal zero by @hjh0119 in #3857
• support val_dataset_shuffle by @Jintao-Huang in #3860
• Update swift docker by @Jintao-Huang in #3866
• fix citest & minimax link by @Jintao-Huang in #3868
• fix grpo save checkpoint by @hjh0119 in #3865
• support glm4-z1 by @hjh0119 in #3862
• add paper link by @tastelikefeet in #3886
• refactor mm target_regex (compat peft/vllm) by @Jintao-Huang in #3879
• Support kimi-vl by @Jintao-Huang in #3884
• Fix glm4 z1 by @Jintao-Huang in #3889
• fix bugs by @Jintao-Huang in #3893
• fix typealias to be compatible with Python 3.9 by @hjh0119 in #3895
• Fix ui by @tastelikefeet in #3903
• Fix fp16 bf16 by @Jintao-Huang in #3909
• add rm center_rewards_coefficient argument by @hjh0119 in #3917
• revert swift_from_pretrained by @Jintao-Huang in #3914
• fix grpo doc by @hjh0119 in #3920
• update qwen2_5_omni by @Jintao-Huang in #3908
• Support qwen3 by @Jintao-Huang in #3945
• Decouple vLLM engine and GRPOTrainer. by @hjh0119 in #3911
• Refactor Agent Template by @Jintao-Huang in #3918
• update docs by @Jintao-Huang in #3961
• fix bugs by @Jintao-Huang in #3962
• Support hermes loss_scale by @Jintao-Huang in #3963
• fix parse tools by @Jintao-Huang in #3975
• Update unsloth compatibility by @tastelikefeet in #3970
• Fix qwen2.5-omni use_audio_in_video by @Jintao-Huang in #3987
• Fix web-ui by @tastelikefeet in #3997
• fix get_toolcall & fix ci by @Jintao-Huang in #3999
• fix bugs by @Jintao-Huang in #4001
• fix seq_cls by @Jintao-Huang in #4002

New Contributors

• @IamLihua made their first contribution in #3845
• @Evilxya made their first contribution in #3846

Full Changelog: v3.3.0...v3.3.1