Ms-swift

2. Web-UI 支持音频和视频。
3. 支持使用 OpenAI API 部署音频和视频模型。
4. 采用新的多模态训练框架。
5. 支持视频模型的推理加速（lmdeploy 和 internvl2 系列）。

新模型：
1. idefics3-8b-llama3
2. llava-hf 72b、110b、llama3-llava
3. deepseek-coder-v2、deepseek-coder-lite-v2、deepseek-v2

What's Changed
* support video deploy by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1632
* fix infer deploy by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1665
* Fix bugs by tastelikefeet in https://github.com/modelscope/ms-swift/pull/1664
* Add generation config support for swift eval by Yunnglin in https://github.com/modelscope/ms-swift/pull/1659
* Add a bunch of small features by tastelikefeet in https://github.com/modelscope/ms-swift/pull/1669
* Support lmdeploy&vllm support video by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1668
* [TorchAcc] fix: add mark_step when GA is greater than 1 by baoleai in https://github.com/modelscope/ms-swift/pull/1677
* Support audio and video in web-ui by tastelikefeet in https://github.com/modelscope/ms-swift/pull/1681
* Fix lmdeploy deploy seed by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1678
* add paper link by tastelikefeet in https://github.com/modelscope/ms-swift/pull/1684
* set default writer_batch_size argument in mllm rlhf by hjh0119 in https://github.com/modelscope/ms-swift/pull/1689
* New multimodal framework by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1594
* fix video component by tastelikefeet in https://github.com/modelscope/ms-swift/pull/1692
* refactor internvl2 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1625
* fix model dtype in rlhf by hjh0119 in https://github.com/modelscope/ms-swift/pull/1696
* Fix minicpmv 2.6 & video/plain-text bug by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1698
* Fix internvl2 device map by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1701
* Support deepseek-v2 base by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1708
* [TorchAcc] update data bucketing strategy by baoleai in https://github.com/modelscope/ms-swift/pull/1706
* [TorchAcc] consolidate FSDP ckpt after saving instead of during loading by baoleai in https://github.com/modelscope/ms-swift/pull/1709
* fix deploy stream by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1713
* update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1714
* Fix mllm rlhf by hjh0119 in https://github.com/modelscope/ms-swift/pull/1707
* support idefics3 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1720
* Fix minicpm-v-v2.6 device_map='auto' & agent deploy by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1733
* Fix kto by hjh0119 in https://github.com/modelscope/ms-swift/pull/1737
* support llava-hf 72b, 110b, llama3-llava-hf by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/1741
* Fix glm4v rlhf by hjh0119 in https://github.com/modelscope/ms-swift/pull/1745

New Contributors
* Yunnglin made their first contribution in https://github.com/modelscope/ms-swift/pull/1659

**Full Changelog**: https://github.com/modelscope/ms-swift/compare/v2.3.0...v2.3.1

中文版

新特性
1. GRPO支持vLLM的tensor parallel模式。例子参考[这里](https://github.com/modelscope/ms-swift/tree/main/examples/train/grpo)。
2. GRPO支持co-locate和optimizer和model的offload，支持分批次导入权重和合并LoRA，节约显存资源，使72B模型的训练可以在四张A100上运行。例子参考[这里](https://github.com/modelscope/ms-swift/blob/main/examples/train/grpo/train_72b_4gpu.sh)。
3. GRPO支持code ORM。最佳实践参考[这里](https://swift.readthedocs.io/zh-cn/latest/BestPractices/GRPO%E4%BB%A3%E7%A0%81%E8%AE%AD%E7%BB%83.html)。

新模型
1. Qwen/QwQ-32B系列
2. inclusionAI/Ling-lite系列

New Features

1. GRPO supports the tensor parallel mode of vLLM. Examples can be found [here](https://github.com/modelscope/ms-swift/tree/main/examples/train/grpo).
2. GRPO supports co-locating offloading for both the optimizer and the model, allows for batch weight loading and LoRA merging, saving GPU memory resources, which enables training of a 72B model on four A100 GPUs. Examples can be found [here](https://github.com/modelscope/ms-swift/blob/main/examples/train/grpo/train_72b_4gpu.sh).
3. GRPO supports code ORM. Best practices can be found [here](https://swift.readthedocs.io/zh-cn/latest/BestPractices/GRPO代码训练.html).

New Models

1. Qwen/QwQ-32B series
2. inclusionAI/Ling-lite series

What's Changed
* Support vllm LLMEngine by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3370
* update publish workflows by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3374
* support ling by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3379
* Support mp mode and hybrid mode of GRPO by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3381
* fix name by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3382
* fix web-ui infer by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3384
* fix bugs by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3385
* fix bugs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3386
* support Qwen/QwQ-32B by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3388
* support qwq-awq by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3391
* support lmdeploy qwen2_5_vl by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3394
* update infer_save by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3400
* update requirements by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3403
* fix ollama export by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3406
* Fix grpo engine by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3412
* fix infer_stream by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3413
* FIx some comments, add dlc script by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3419
* add comments and docs by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3424
* fix issue 1663 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3417
* Support GRPO model and optimizer offload, and split loading model by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3427
* update wechat by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3430
* Fix vllm random by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3437
* fix seed by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3438
* fix_base_deploy by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3442
* fix GRPO device mismatch by hjh0119 in https://github.com/modelscope/ms-swift/pull/3440
* compat vllm==0.5.1 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3444
* fix grpo multimodal doc by mi804 in https://github.com/modelscope/ms-swift/pull/3449
* support grpo code orm by hjh0119 in https://github.com/modelscope/ms-swift/pull/3431
* fix GRPO seed by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3458
* fix grpo multi nodes by hjh0119 in https://github.com/modelscope/ms-swift/pull/3462
* Fix tensor parallel hang by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3464
* fix grpo trainer zero3 always gather parameters by tcye in https://github.com/modelscope/ms-swift/pull/3467
* fix grpo temperature inconsistency by hjh0119 in https://github.com/modelscope/ms-swift/pull/3468
* fix grad_norm nan by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3465
* fix grad_norm by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3469
* update minimax by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3471
* Support 72b script with 4 gpus by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3472
* refactor packing by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3457
* Fix some docs by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3475
* fix grpo ddp hang by hjh0119 in https://github.com/modelscope/ms-swift/pull/3476
* fix moe quant by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3478
* Delete duplicate parameters in train_72b_4gpu.sh by Marquis03 in https://github.com/modelscope/ms-swift/pull/3479
* fix image by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3480
* fix infer gptq internvl2 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3481
* Resume sample by BC-A in https://github.com/modelscope/ms-swift/pull/3460
* fix qwen2_vl flash_attn deepspeed by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3484
* Fix seed of tp=1 by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3486
* fix use_cache by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3487
* Fix qwen2 5 vl grounding by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3491
* fix ovis2 device_map by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3496
* fix template.decode by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3497

New Contributors
* tcye made their first contribution in https://github.com/modelscope/ms-swift/pull/3467
* Marquis03 made their first contribution in https://github.com/modelscope/ms-swift/pull/3479
* BC-A made their first contribution in https://github.com/modelscope/ms-swift/pull/3460

**Full Changelog**: https://github.com/modelscope/ms-swift/compare/v3.2.0...v3.2.1

中文版

新特性
1. GRPO支持多vLLM/lmdeploy数据并行采样，支持异步采样，参考[这里](https://github.com/modelscope/ms-swift/tree/main/examples/train/grpo)。多模态GRPO实验记录参考[这里](https://swift.readthedocs.io/zh-cn/latest/BestPractices/GRPO%E5%A4%9A%E6%A8%A1%E6%80%81%E8%AE%AD%E7%BB%83.html)。
2. `swift deploy` infer_backend为pt时支持动态batch；流式推理接口修改（break change）。
3. `swift infer` infer_backend为vllm/lmdeploy支持数据并行。参考[这里](https://github.com/modelscope/ms-swift/blob/main/examples/infer/vllm/ddp.sh)。
4. 支持moun优化器，参考[这里](https://github.com/modelscope/ms-swift/blob/main/examples/train/optimizer/muon.sh)。

新模型
1. moonshotai/Moonlight-16B-A3B-Instruct
2. LLM-Research/Phi-4-mini-instruct, LLM-Research/Phi-4-multimodal-instruct
3. DeepSeek-V3-awq, deepseek-r1-awq
4. Baichuan-M1-14B-Instruct

新数据集
1. 多模态GRPO：
- lmms-lab/multimodal-open-r1-8k-verified
- okwinds/clevr_cogen_a_train



New Features

1. GRPO supports multi-vLLM/lmdeploy data parallel sampling and asynchronous sampling. For more information, refer to [here](https://github.com/modelscope/ms-swift/tree/main/examples/train/grpo). Records of multi-modal GRPO experiments can be found [here](https://swift.readthedocs.io/zh-cn/latest/BestPractices/GRPO多模态训练.html).
2. When `swift deploy` infer_backend is set to pt, it supports dynamic batching; the streaming inference interface has been modified (breaking change).
3. When `swift infer` infer_backend is set to vllm/lmdeploy, it supports data parallelism. Refer to [here](https://github.com/modelscope/ms-swift/blob/main/examples/infer/vllm/ddp.sh).
4. Supports the muon optimizer. For more information, refer to [here](https://github.com/modelscope/ms-swift/blob/main/examples/train/optimizer/muon.sh).

New Models

1. moonshotai/Moonlight-16B-A3B-Instruct
2. LLM-Research/Phi-4-mini-instruct, LLM-Research/Phi-4-multimodal-instruct
3. DeepSeek-V3-awq, deepseek-r1-awq
4. Baichuan-M1-14B-Instruct

New Datasets

1. Multi-modal GRPO:
- lmms-lab/multimodal-open-r1-8k-verified
- okwinds/clevr_cogen_a_train




What's Changed
* fix setup.py by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3198
* support vllm dp by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3201
* update dataset & fix bugs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3203
* Support multiple vllms by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3202
* update distill docs by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3216
* compatible with trl0.16 by hjh0119 in https://github.com/modelscope/ms-swift/pull/3209
* support r1 awq by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3206
* fix grpo old_per_token_logps by hjh0119 in https://github.com/modelscope/ms-swift/pull/3220
* Support the generation of JanusPro models by DaozeZhang in https://github.com/modelscope/ms-swift/pull/3218
* Update the JanusPro-generation by DaozeZhang in https://github.com/modelscope/ms-swift/pull/3221
* fix load args by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3226
* update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3230
* Speed up GRPO by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3229
* fix docs zh by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3231
* fix deepseek_vl2 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3233
* support moonlight by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3232
* support muon optimizer by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3234
* update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3243
* fix grpo npu vllm by hjh0119 in https://github.com/modelscope/ms-swift/pull/3242
* fix grpo single card by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3246
* save val_dataset by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3248
* fix grpo compat transformers==4.47.* by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3252
* grpo_countdown & fix format reward by mi804 in https://github.com/modelscope/ms-swift/pull/3269
* Support the base64 format of generated images for JanusPro by DaozeZhang in https://github.com/modelscope/ms-swift/pull/3265
* Fix typos by co63oc in https://github.com/modelscope/ms-swift/pull/3266
* compat lmdeploy 0.7 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3256
* fix lmdeploy by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3274
* GRPO+LMDeploy 0.7 by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3277
* Support max memory by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3282
* add lmdeploy dp shell by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3284
* Support Baichuan-M1-14B-Instruct by DaozeZhang in https://github.com/modelscope/ms-swift/pull/3271
* fix grpo top_k by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3293
* fix lmdeploy mllm in grpo by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3296
* Update FAQ by slin000111 in https://github.com/modelscope/ms-swift/pull/3289
* fix: error when uploading model to huggingface by xavier-h-10 in https://github.com/modelscope/ms-swift/pull/3297
* add multimodal clevr exp by mi804 in https://github.com/modelscope/ms-swift/pull/3301
* update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3304
* [refactor] patch_vllm by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3306
* GRPO mllm script by hjh0119 in https://github.com/modelscope/ms-swift/pull/3305
* [refactor & feat] support pt dynamic batch by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3278
* Support ZeRO++ by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3315
* Revert pt engine batch infer by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3316
* optimize model_type by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3318
* Fix bugs & Update docs/datasets by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3322
* fix grpo zero3 by hjh0119 in https://github.com/modelscope/ms-swift/pull/3324
* fix grpo zero3 by hjh0119 in https://github.com/modelscope/ms-swift/pull/3326
* compat vllm>=0.5.1 lmdeploy>=0.5.0 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3332
* update external plugins by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3334
* fix generation_config by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3335
* fix check_model error by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3336
* update get_model_tokenizer_with_flash_attn by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3337
* add geoqa grpo experiment by mi804 in https://github.com/modelscope/ms-swift/pull/3344
* fix max_memory by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3347
* support phi4-multimodal by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3350
* fix：fix bugs in cosine reward of GRPO by youyc22 in https://github.com/modelscope/ms-swift/pull/3358
* Remove entry including invalid `ROADMAP` link from English & Chinese documentation by 3manifold in https://github.com/modelscope/ms-swift/pull/3357
* update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3349
* Support the <video> token for Ovis2 models by DaozeZhang in https://github.com/modelscope/ms-swift/pull/3364
* update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3365
* add grpo openr1 multimodal experiment by mi804 in https://github.com/modelscope/ms-swift/pull/3368
* fix swift app format by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3367

New Contributors
* xavier-h-10 made their first contribution in https://github.com/modelscope/ms-swift/pull/3297
* youyc22 made their first contribution in https://github.com/modelscope/ms-swift/pull/3358
* 3manifold made their first contribution in https://github.com/modelscope/ms-swift/pull/3357

**Full Changelog**: https://github.com/modelscope/ms-swift/compare/v3.1.1...v3.2.0

中文版

新特性
1. 支持大模型、多模态模型、Agent、多节点GRPO训练，参考[这里](https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO.html)。
2. 支持Embeding模型训练，参考[这里](https://github.com/modelscope/ms-swift/blob/main/examples/train/embedding/train.sh)。
3. `swift sample`支持MCTS、蒸馏方式数据采样，支持多模态模型采样。
4. 支持自定义数据集评测，参考[这里](https://swift.readthedocs.io/zh-cn/latest/Instruction/%E8%AF%84%E6%B5%8B.html)。

新模型
1. AIDC-AI/Ovis2-2B系列
2. Qwen/Qwen2.5-VL-72B-Instruct-AWQ系列
3. stepfun-ai/GOT-OCR-2.0-hf
4. stepfun-ai/Step-Audio-Chat
5. mistralai/Mistral-Small-24B-Instruct-2501

新数据集
1. GRPO相关
- AI-ModelScope/MATH-lighteval
- LLM-Research/xlam-function-calling-60k
- AI-MO/NuminaMath-TIR
2. R1相关
- liucong/Chinese-DeepSeek-R1-Distill-data-110k-SFT
- modelscope/MathR, modelscope/MathR-32B-Distill

New Features

1. Support for large models, multimodal models, Agents, and multi-node GRPO training. Refer to [this documentation](https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO.html).
2. Support for Embedding model training. Refer to [this script](https://github.com/modelscope/ms-swift/blob/main/examples/train/embedding/train.sh).
3. `swift sample` supports MCTS and distillation data sampling, as well as multimodal model sampling.
4. Support for custom dataset evaluation. Refer to [this documentation](https://swift.readthedocs.io/zh-cn/latest/Instruction/评测.html).

New Models

1. AIDC-AI/Ovis2-2B series
2. Qwen/Qwen2.5-VL-72B-Instruct-AWQ series
3. stepfun-ai/GOT-OCR-2.0-hf
4. stepfun-ai/Step-Audio-Chat
5. mistralai/Mistral-Small-24B-Instruct-2501

New Datasets

1. Related to GRPO
- AI-ModelScope/MATH-lighteval
- LLM-Research/xlam-function-calling-60k
- AI-MO/NuminaMath-TIR
2. Related to R1
- liucong/Chinese-DeepSeek-R1-Distill-data-110k-SFT
- modelscope/MathR, modelscope/MathR-32B-Distill


What's Changed
* Add evalscope native backend by Yunnglin in https://github.com/modelscope/ms-swift/pull/2981
* support mistralai/Mistral-Small-24B-Instruct-2501 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3030
* MCTS Sampler by lxline in https://github.com/modelscope/ms-swift/pull/2967
* fix windows url by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3041
* Support sample multi modal models by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3048
* Support sft embedding model by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3039
* support GRPO by hjh0119 in https://github.com/modelscope/ms-swift/pull/3022
* fix grpo by hjh0119 in https://github.com/modelscope/ms-swift/pull/3050
* fix grpo by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3051
* update docs (fine-tuning) by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3052
* bump version by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3053
* fix grpo model_type by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3057
* update rlhf documents by hjh0119 in https://github.com/modelscope/ms-swift/pull/3055
* add grpo multinode scripts by hjh0119 in https://github.com/modelscope/ms-swift/pull/3059
* Fix orm env by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3065
* Support external plugins by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3066
* update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3070
* fix grpo nan by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3075
* fix grpo metric_for_best_model by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3077
* register MathR by mi804 in https://github.com/modelscope/ms-swift/pull/3078
* fix accuracy reward by hjh0119 in https://github.com/modelscope/ms-swift/pull/3080
* fix SwiftModel by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3071
* Fix grpo vlm (internvl2.5) by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3081
* Refactor orm prm by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3085
* fix competition math by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3086
* support cuda operations to npu by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3087
* fix grpo temperature 0.7->0.9 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3091
* support grpo vllm lora by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3095
* Feat: Eval custom dataset by Yunnglin in https://github.com/modelscope/ms-swift/pull/3093
* cosine and repetition reward for GRPO by hjh0119 in https://github.com/modelscope/ms-swift/pull/3079
* fix get_device by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3097
* Fix/grpo by MrToy in https://github.com/modelscope/ms-swift/pull/3101
* fix unsloth by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3100
* support grpo npu by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3102
* fix grpo zero3 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3104
* support log completions by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3110
* Fix typos by co63oc in https://github.com/modelscope/ms-swift/pull/3111
* update trl version by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3117
* fix eval docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3118
* Support llamapro for grpo by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3119
* fix grpo trainer by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3120
* fix cleanup error by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3121
* Fix typos by co63oc in https://github.com/modelscope/ms-swift/pull/3123
* refactor patcher by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3124
* Support lmdeploy in GRPO by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3126
* support stepfun-ai/Step-Audio-Chat by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3127
* update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3131
* fix grpo pt infer generation_config by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3135
* support_local_path by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3140
* Support swanlab by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3142
* fix grpo sample by MrToy in https://github.com/modelscope/ms-swift/pull/3144
* fix grpo vllm lora by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3134
* fix create_repo by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3147
* fix grpo zero3 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3149
* docs: report_to add swanlab by Zeyi-Lin in https://github.com/modelscope/ms-swift/pull/3158
* Support Ovis2 models by DaozeZhang in https://github.com/modelscope/ms-swift/pull/3163
* support grpo metric_for_best_model by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3155
* Fix ovis2 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3169
* Support Agent GRPO by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3170
* fix max_length error by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3173
* fix streaming by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3176
* Fix/agent grpo by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3172
* Fix lmdeploy branch by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3145
* fix internvl-4b by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3178
* refactor cosine orm by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3179
* fix sampler reaches max_length by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3180
* Fix prm in sampler by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3184
* Support GOT_OCR2_hf by DaozeZhang in https://github.com/modelscope/ms-swift/pull/3182
* Knowledge Distillation sampling by mi804 in https://github.com/modelscope/ms-swift/pull/3185
* compat vllm==0.7.2 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3083
* support r1 dataset by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3191
* Refactor grpo dataset by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3192
* Add links to agent grpo by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3193

New Contributors
* MrToy made their first contribution in https://github.com/modelscope/ms-swift/pull/3101
* co63oc made their first contribution in https://github.com/modelscope/ms-swift/pull/3111
* Zeyi-Lin made their first contribution in https://github.com/modelscope/ms-swift/pull/3158

**Full Changelog**: https://github.com/modelscope/ms-swift/compare/v3.1.0...v3.1.1

中文版

新特性

1. 支持`swift sample`命令进行数据采样，参考[这里](https://swift.readthedocs.io/zh-cn/latest/Instruction/%E9%87%87%E6%A0%B7.html)。
2. 支持强化微调训练，目前已支持拒绝采样微调，参考[这里](https://swift.readthedocs.io/zh-cn/latest/Instruction/%E5%BC%BA%E5%8C%96%E5%BE%AE%E8%B0%83.html)。
3. Grounding任务自定义数据格式重构，参考[这里](https://github.com/modelscope/ms-swift/blob/main/examples/notebook/qwen2_5-vl-grounding/zh.ipynb)。
4. `swift infer`支持输出推理速度和ACC/ROUGE/BLEU指标。

新模型

1. Qwen/Qwen2.5-VL-3B-Instruct系列
2. Qwen/Qwen2.5-7B-Instruct-1M系列
3. deepseek-ai/Janus-Pro-1B系列
4. bytedance-research/UI-TARS-2B-SFT系列


新数据集
1. ServiceNow-AI/R1-Distill-SFT
2. bespokelabs/Bespoke-Stratos-17k
3. open-thoughts/OpenThoughts-114k

English Version

New Features

1. Supports the `swift sample` command for data sampling; refer to [here](https://swift.readthedocs.io/en/latest/Instruction/Sample.htmll).
2. Supports reinforcement fine-tuning training, with current support for rejection sampling fine-tuning; refer to [here](https://swift.readthedocs.io/en/latest/Instruction/Reinforced-Fine-tuning.html).
3Grounding task custom data format restructuring; refer to [here](https://github.com/modelscope/ms-swift/blob/main/examples/notebook/qwen2_5-vl-grounding/zh.ipynb).
4. `swift infer` supports outputting inference speed and ACC/ROUGE/BLEU metrics.

New Models

1. Qwen/Qwen2.5-VL-3B-Instruct Series
2. Qwen/Qwen2.5-7B-Instruct-1M Series
3. deepseek-ai/Janus-Pro-1B Series
4. bytedance-research/UI-TARS-2B-SFT Series


New Datasets
1. ServiceNow-AI/R1-Distill-SFT
2. bespokelabs/Bespoke-Stratos-17k
3. open-thoughts/OpenThoughts-114k






What's Changed
* add "enable_prefix_caching" args for vllm engine. by Leoyzen in https://github.com/modelscope/ms-swift/pull/2939
* Fix vllm docs link & fix web-ui by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2970
* Fix sample by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2971
* support merge-lora & quant by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2973
* support create_checkpoint_symlink by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2975
* Sampling and RFT by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2977
* support auto dataset mapping by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2976
* support qwen2_5 long by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2982
* sys_prompt from file by lxline in https://github.com/modelscope/ms-swift/pull/2980
* support bytedance-research/UI-TARS-2B-SFT series by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2987
* support Qwen/Qwen2.5-VL-3B-Instruct series model by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2996
* fix qwen2_5-vl by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2998
* support Qwen/Qwen2.5-VL-72B-Instruct by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2999
* refactor grounding by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3000
* compatible with trl v0.13 by hjh0119 in https://github.com/modelscope/ms-swift/pull/2992
* update R1 dataset by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3005
* fix qwen2.5-vl grounding (refactor) by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2979
* fix deploy by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3007
* support infer metric: acc/rouge or bleu by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3008
* support deepseek janus pro by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3009
* update readme by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3011
* fix parse_dict by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3012
* update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3015
* Fix readme & update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3018
* fix push to hub by tastelikefeet in https://github.com/modelscope/ms-swift/pull/3024
* Fix bugs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3025
* fix bugs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/3026
* Fix qwen tool template to official format by Leoyzen in https://github.com/modelscope/ms-swift/pull/2988
* fix message merging strategy when multi-turn tool calling. by Leoyzen in https://github.com/modelscope/ms-swift/pull/2986

New Contributors
* Leoyzen made their first contribution in https://github.com/modelscope/ms-swift/pull/2939

**Full Changelog**: https://github.com/modelscope/ms-swift/compare/v3.0.3...v3.1.0

中文版

新特性
1. 支持多模态大模型SequenceClassification架构用于多模态分类任务，参考[这里](https://github.com/modelscope/ms-swift/blob/main/examples/train/seq_cls/qwen2_vl/sft.sh)。
2. 支持多模态大模型reward model训练。

新模型
1. Shanghai_AI_Laboratory/internlm3-8b-instruct
2. OpenBMB/MiniCPM-o-2_6
3. deepseek-ai/DeepSeek-R1, deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B系列
4. bytedance-research/Valley-Eagle-7B
5. LLM-Research/phi-4
6. Qwen/Qwen2.5-Math-PRM-7B, Qwen/Qwen2.5-Math-PRM-72B
7. MiniMaxAI/MiniMax-Text-01, MiniMaxAI/MiniMax-VL-01

English Version

New Features

1. Support multi-modal large model SequenceClassification architecture for multi-modal classification tasks, see [here](https://github.com/modelscope/ms-swift/blob/main/examples/train/seq_cls/qwen2_vl/sft.sh).
2. Support training of multi-modal reward model.

New Models

1. Shanghai_AI_Laboratory/internlm3-8b-instruct
2. OpenBMB/MiniCPM-o-2_6
3. deepseek-ai/DeepSeek-R1, deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B series
4. bytedance-research/Valley-Eagle-7B
5. LLM-Research/phi-4
6. Qwen/Qwen2.5-Math-PRM-7B, Qwen/Qwen2.5-Math-PRM-72B
7. MiniMaxAI/MiniMax-Text-01, MiniMaxAI/MiniMax-VL-01




What's Changed
* update qlora shell by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2880
* fix docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2882
* support multi round dpo by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2884
* Support infer n parameter by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2893
* Fix qwen vl eval by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2892
* fix infer engine by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2898
* Add phi4 by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2895
* fix link & bug by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2902
* update video infer examples by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2840
* Sampler by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2905
* Fix a bug when lint code by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2906
* Fix bugs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2907
* update plugin doc by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2908
* fix vllm tp stuck by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2909
* fix replace_video2image by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2913
* Fix read file mode by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2915
* fix inspect init by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2916
* Update rm by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2919
* Add internlm3 dense by HIT-cwh in https://github.com/modelscope/ms-swift/pull/2920
* internlm3 lint pass by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2923
* Fix web ui log by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2924
* Support Valley by lxline in https://github.com/modelscope/ms-swift/pull/2921
* support minicpm-o by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2918
* fix vllm tp block by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2927
* update docs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2929
* Support first prms by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2926
* fix Valley by lxline in https://github.com/modelscope/ms-swift/pull/2931
* Support mllm seq_cls/rm by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2934
* fix bugs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2938
* support deepseek-ai/DeepSeek-R1 by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2940
* Fix quant template by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2942
* Support minimax by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2943
* Fix mllm seq cls by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2945
* support deepseek_r1_distill by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2946
* fix demo_hf by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2951
* fix infer_stream by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2952
* fix citest by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2953
* fix bugs by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2954
* update requirements by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2957
* update web-ui images by tastelikefeet in https://github.com/modelscope/ms-swift/pull/2958
* update quant_mllm shell by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2959
* fix max_length error print by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2960
* fix seq_cls patcher by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2963
* ppo compat transformers>=4.47.* by Jintao-Huang in https://github.com/modelscope/ms-swift/pull/2964


**Full Changelog**: https://github.com/modelscope/ms-swift/compare/v3.0.2...v3.0.3