v3.3.0
中文版
新特性
- 支持DAPO算法,训练文档参考:https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO.html#dapo
- 支持多模态模型的序列packing,包括qwen2-vl、qwen2.5-vl、qwen2.5-omni和internvl2.5系列,训练速度提升100%。训练脚本参考:https://github.com/modelscope/ms-swift/tree/main/examples/train/packing
- 新增SWIFT和Megatron-SWIFT镜像,参考这里:https://swift.readthedocs.io/zh-cn/latest/GetStarted/SWIFT%E5%AE%89%E8%A3%85.html#id3
- 多模态/Omni/Moe量化能力增强,量化脚本参考这里:https://github.com/modelscope/ms-swift/blob/main/examples/export/quantize
新模型
- Qwen/Qwen2.5-Omni-7B
- LLM-Research/Llama-4-Scout-17B-16E-Instruct系列
- cognitivecomputations/DeepSeek-V3-0324-AWQ
English Version
New Features
- Supports the DAPO algorithm; training documentation can be found here: https://swift.readthedocs.io/en/latest/Instruction/GRPO.html#dapo
- Supports sequence packing for multimodal models, including qwen2-vl, qwen2.5-vl, qwen2.5-omni, and the internvl2.5 series, with a 100% increase in training speed. Training scripts can be found here: https://github.com/modelscope/ms-swift/tree/main/examples/train/packing
- Added SWIFT and Megatron-SWIFT mirrors, see details here: https://swift.readthedocs.io/en/latest/GetStarted/SWIFT-installation.html#mirror
- Enhanced quantization capabilities for Multimodal/Omni/Moe models, shell scripts can be found here: https://github.com/modelscope/ms-swift/blob/main/examples/export/quantize
New Models
- Qwen/Qwen2.5-Omni-7B
- LLM-Research/Llama-4-Scout-17B-16E-Instruct series
- cognitivecomputations/DeepSeek-V3-0324-AWQ
What's Changed
- fix shell by @Jintao-Huang in #3675
- support Qwen/Qwen2.5-Omni-7B (sft/dpo/grpo) by @Jintao-Huang in #3613
- fix grpo rank by @hjh0119 in #3687
- Grpo vl72b script by @hjh0119 in #3692
- fix import error by @Jintao-Huang in #3700
- [megatron] fix val_dataset streaming by @Jintao-Huang in #3699
- fix grpo qwen2_5_omni by @Jintao-Huang in #3701
- fix grpo vl by @Jintao-Huang in #3704
- update warning_once by @Jintao-Huang in #3706
- fix grpo template copy by @Jintao-Huang in #3708
- fix adalora by @tastelikefeet in #3714
- fix qwen2_5-omni by @Jintao-Huang in #3716
- Fix grpo dora by @hjh0119 in #3709
- support qwen2_5_vl packing by @Jintao-Huang in #3694
- fix qwen2_5 omni by @Jintao-Huang in #3734
- fix grpo train dataloader by @Jintao-Huang in #3736
- support internvl2.5 packing by @Jintao-Huang in #3735
- Support qwen2 5-vl awq quant & update shell by @Jintao-Huang in #3743
- support moe quant by @Jintao-Huang in #3772
- update liger kernel by @Jintao-Huang in #3775
- support llama4 by @Jintao-Huang in #3777
- support DAPO by @hjh0119 in #3725
- [Gemma] Fixing the ndarray cast warning by @Reichenbachian in #3791
- add swift docker by @Jintao-Huang in #3796
- support streaming shuffle by @Jintao-Huang in #3782
- grpo lmdeploy warn by @hjh0119 in #3800
- fix math accuracy by @hjh0119 in #3795
- fix grounding dataset concat by @Jintao-Huang in #3802
- fix omni max_model_len by @Jintao-Huang in #3803
- fix get_config_attrs by @Jintao-Huang in #3807
- Fix grpo ovis2 by @Jintao-Huang in #3808
- more grpo log by @hjh0119 in #3801
- fix reward_template by @Jintao-Huang in #3813
- [GRPO] fix template copy (async generate) by @Jintao-Huang in #3814
- update docs by @Jintao-Huang in #3815
- optimize zero3 rlhf code by @Jintao-Huang in #3816
- fix grpo zero3 inflight params by @hjh0119 in #3818
- fix grpo log_completions by @Jintao-Huang in #3819
- vLLM 0.8.3 support for GRPO colocate mode by @hjh0119 in < 601C a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="2983344148" data-permission-text="Title is private" data-url="https://github.com/modelscope/ms-swift/issues/3820" data-hovercard-type="pull_request" data-hovercard-url="/modelscope/ms-swift/pull/3820/hovercard" href="https://github.com/modelscope/ms-swift/pull/3820">#3820
- fix web-ui by @Jintao-Huang in #3822
- fix telechat by @hjh0119 in #3825
- fix omni zero3 by @Jintao-Huang in #3826
- feat: grpo async generate thread-safe queue production by @hjh0119 in #3821
- fix grpo async generate by @hjh0119 in #3829
- update docs grpo vllm by @Jintao-Huang in #3831
- support omni vllm by @Jintao-Huang in #3832
- remove sequence_parallel_size by @Jintao-Huang in #3835
- update grpo type annotations by @hjh0119 in #3834
- fix grpo multi turn tp by @hjh0119 in #3837
- [docs] fix seq_parallel by @Jintao-Huang in #3838
New Contributors
- @Reichenbachian made their first contribution in #3791
Full Changelog: v3.2.2...v3.3.0