v3.2.2

中文版

新特性

1. Megatron-SWIFT发布。支持TP、PP、SP、CP等并行技术对Qwen系、Llama系、Deepseek-R1蒸馏系等100+模型进行预训练和微调。支持streaming数据集和序列packing功能支持超大数据集并提升训练效率。更多内容参考Megatron-SWIFT训练文档。
2. 支持多轮GRPO训练以适配例如Deep Search等多轮agent工具调用场景，示例代码参考这里。
   • 支持mini-batch，降低训练时的显存消耗。参考GRPO训练文档。
3. 支持iic/gme-Qwen2-VL-2B-Instruct等多模态模型的Embedding训练。具体参考embedding模型训练文档。
4. 支持大模型和多模态大模型的多标签分类和回归任务的训练到部署。示例脚本参考这里。
5. 支持在训练过程中使用EvalScope对模型进行评测，及时了解模型的训练效果。示例脚本参考评测文档。
6. 书写外置plugin，以支持多模态模型LoRA训练LLM的同时，全参数训练ViT，并采用不同的学习率。避免ViT部分merge-lora造成的精度误差。示例脚本参考这里。

新模型

1. iic/gme-Qwen2-VL-2B-Instruct系列
2. Qwen/Qwen2.5-VL-32B-Instruct
3. LLM-Research/gemma-3-4b-it系列
4. deepseek-ai/DeepSeek-V3-0324
5. mistralai/Mistral-Small-3.1-24B-Instruct-2503系列

English Version

New Features

1. Release of Megatron-SWIFT: Megatron-SWIFT has been released, supporting various parallel technologies such as TP (Tensor Parallelism), PP (Pipeline Parallelism), SP (Sequence Parallelism), and CP (Context Parallelism) for pre-training and fine-tuning over 100 models, including the Qwen series, Llama series, and Deepseek-R1 distillation series. It also supports streaming datasets and sequence packing, enabling the handling of ultra-large datasets while improving training efficiency. For more details, refer to the Megatron-SWIFT Training Documentation.
2. Support for Multi-turn GRPO Training: Supports multi-turn GRPO training to adapt to scenarios such as multi-turn agent tool calls in Deep Search. Example code can be found here.
   • Supports mini-batch training to reduce GPU memory consumption during training. Refer to the GRPO Training Documentation.
3. Embedding Training for Multimodal Models: Supports embedding training for multimodal models such as iic/gme-Qwen2-VL-2B-Instruct. For more information, refer to the Embedding Model Training Documentation.
4. Multi-label Classification and Regression Tasks for Large Models and Multimodal Large Models: Supports end-to-end training and deployment for multi-label classification and regression tasks for large models and multimodal large models. Example scripts can be found here.
5. Model Evaluation with EvalScope During Training: Supports model evaluation using EvalScope during training to monitor training performance in real time. Example scripts can be found in the Evaluation Documentation.
6. Custom External Plugin for LoRA + ViT Training: Provides an external plugin to support LoRA training for LLMs (Large Language Models) while performing full-parameter training for ViTs (Vision Transformers) with different learning rates. This avoids precision errors caused by merging LoRA into the ViT portion. Example code can be found here.

New Models

1. iic/gme-Qwen2-VL-2B-Instruct series
2. Qwen/Qwen2.5-VL-32B-Instruct
3. LLM-Research/gemma-3-4b-it series
4. deepseek-ai/DeepSeek-V3-0324
5. mistralai/Mistral-Small-3.1-24B-Instruct-2503 series

What's Changed

• update code doc by @hjh0119 in #3498
• fix readme by @Jintao-Huang in #3499
• feat: swanlab config add ms-swift by @Zeyi-Lin in #3500
• Support GME models by @tastelikefeet in #3513
• fix docs by @tastelikefeet in #3514
• Fix docs links by @tastelikefeet in #3516
• fix vllm memory leak by @hjh0119 in #3515
• [Docs] Easy .[all] install from git by @xihuai18 in #3518
• Fix bugs by @tastelikefeet in #3520
• support megatron by @Jintao-Huang in #2885
• fix megatron by @Jintao-Huang in #3527
• support gemma3 by @hjh0119 in #3492
• fix megatron pipeline parallel by @Jintao-Huang in #3529
• fix megatron tie_weight by @Jintao-Huang in #3530
• support megatron llama by @Jintao-Huang in #3532
• Support megatron llama3.1 3.2 by @Jintao-Huang in #3537
• 更新LlavaHfTemplate以适配transformers版本大于4.47时对LLaVA和LLaVA-Next模型处理图像token逻辑的修改 by @zsxm1998 in #3521
• refactor llava-hf by @Jintao-Huang in #3538
• fix docs by @Jintao-Huang in #3539
• refactor get_megatron_model_meta by @Jintao-Huang in #3542
• Gather infonce loss and support hard negative samples by @tastelikefeet in #3548
• fix docs by @tastelikefeet in #3553
• fix unsloth by @tastelikefeet in #3554
• fix grpo mllm split modules by @hjh0119 in #3552
• grpo embedding layer lora by @hjh0119 in #3531
• update arguments by @Jintao-Huang in #3556
• update doc by @hjh0119 in #3557
• Support all models' embedding and mask fake negative by @tastelikefeet in #3563
• skip grpo first wake up by @hjh0119 in #3562
• move grpovllmengine import by @hjh0119 in #3568
• fix bugs & support dataset_name by @Jintao-Huang in #3565
• fix wrap by @tastelikefeet in #3572
• Feature: add train-eval loop by @Yunnglin in #3569
• compat vllm>=0.8 by @Jintao-Huang in #3574
• [grpo] Fix Incorrect Placement of Data in eval_queue During async_generate by @hjh0119 in #3573
• Fix lmdeploy 0.7.3 by @tastelikefeet in #3584
• support vit full llm lora by @Jintao-Huang in #3575
• support Mistral3.1-2503 by @hjh0119 in #3588
• Support megatron packing by @Jintao-Huang in #3595
• [megatron] support streaming by @Jintao-Huang in #3609
• fix rft by @lxline in #3602
• [template] refactor replace media tokens by @Jintao-Huang in #3614
• fix top_logprobs by @Jintao-Huang in #3616
• Fix bugs by @Jintao-Huang in #3619
• Support multi turn grpo by @tastelikefeet in #3615
• fix grpo npu context by @hjh0119 in #3597
• support regression multi-label by @Jintao-Huang in #3621
• Support peft 0.15 by @tastelikefeet in #3623
• update grpo warning by @hjh0119 in #3598
• fix grpo rm zero3 by @hjh0119 in #3626
• GRPO mini batch by @hjh0119 in #3205
• fix grpo warning with pt backend by @hjh0119 in #3629
• compat transformers 4.50 by @Jintao-Huang in #3625
• support train_sampler_random by @Jintao-Huang in #3631
• fix grpo multi turn by @tastelikefeet in #3632
• update docs by @Jintao-Huang in #3633
• Support deepseek v3 0324 by @Jintao-Huang in #3637
• fix grpo cosine reward by @hjh0119 in #3638
• fix grpo lora split module by @hjh0119 in #3635
• fix reward model by @Jintao-Huang in #3641
• support qwen2_5_vl_32b by @Jintao-Huang in #3642
• fix grpo warning by @hjh0119 in #3630
• grpo reset prefix cache by @...

v3.2.1

中文版

新特性

1. GRPO支持vLLM的tensor parallel模式。例子参考这里。
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。
3. GRPO支持code ORM。最佳实践参考这里。

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.
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.
3. GRPO supports code ORM. Best practices can be found here.

New Models

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

What's Changed

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

New Contributors

• @tcye made their first contribution in #3467
• @Marquis03 made their first contribution in #3479
• @BC-A made their first contribution in #3460

Full Changelog: v3.2.0...v3.2.1

v3.2.0

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。多模态GRPO实验记录参考这里。
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。
4. 支持moun优化器，参考这里。

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. Records of multi-modal GRPO experiments can be found here.
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.
4. Supports the muon optimizer. For more information, refer to here.

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

New Contributors

• @xavier-h-10 made their first contribution in #3297
• @youyc22 made their first contribution in #3358
• @3manifold made their first contribution in #3357

Full Changelog: v3.1.1...v3.2.0

v3.1.1

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。
2. 支持Embeding模型训练，参考这里。
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。

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.
2. Support for Embedding model training. Refer to this script.
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.

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

New Contributors

• @MrToy made their first contribution in #3101
• @co63oc made their first contribution in #3111
• @Zeyi-Lin made their first contribution in #3158

Full Changelog: v3.1.0...v3.1.1

v3.1.0

中文版

新特性

1. 支持swift sample命令进行数据采样，参考这里。
2. 支持强化微调训练，目前已支持拒绝采样微调，参考这里。
3. Grounding任务自定义数据格式重构，参考这里。
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.
2. Supports reinforcement fine-tuning training, with current support for rejection sampling fine-tuning; refer to here.
   3Grounding task custom data format restructuring; refer to here.
3. 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 #2939
• Fix vllm docs link & fix web-ui by @Jintao-Huang in #2970
• Fix sample by @tastelikefeet in #2971
• support merge-lora & quant by @Jintao-Huang in #2973
• support create_checkpoint_symlink by @Jintao-Huang in #2975
• Sampling and RFT by @tastelikefeet in #2977
• support auto dataset mapping by @Jintao-Huang in #2976
• support qwen2_5 long by @Jintao-Huang in #2982
• sys_prompt from file by @lxline in #2980
• support bytedance-research/UI-TARS-2B-SFT series by @Jintao-Huang in #2987
• support Qwen/Qwen2.5-VL-3B-Instruct series model by @Jintao-Huang in #2996
• fix qwen2_5-vl by @Jintao-Huang in #2998
• support Qwen/Qwen2.5-VL-72B-Instruct by @Jintao-Huang in #2999
• refactor grounding by @Jintao-Huang in #3000
• compatible with trl v0.13 by @hjh0119 in #2992
• update R1 dataset by @Jintao-Huang in #3005
• fix qwen2.5-vl grounding (refactor) by @Jintao-Huang in #2979
• fix deploy by @Jintao-Huang in #3007
• support infer metric: acc/rouge or bleu by @Jintao-Huang in #3008
• support deepseek janus pro by @Jintao-Huang in #3009
• update readme by @Jintao-Huang in #3011
• fix parse_dict by @Jintao-Huang in #3012
• update docs by @Jintao-Huang in #3015
• Fix readme & update docs by @Jintao-Huang in #3018
• fix push to hub by @tastelikefeet in #3024
• Fix bugs by @Jintao-Huang in #3025
• fix bugs by @Jintao-Huang in #3026
• Fix qwen tool template to official format by @Leoyzen in #2988
• fix message merging strategy when multi-turn tool calling. by @Leoyzen in #2986

New Contributors

• @Leoyzen made their first contribution in #2939

Full Changelog: v3.0.3...v3.1.0

v3.0.3

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

Full Changelog: v3.0.2...v3.0.3

v3.0.2

中文版

新特性

1. 支持使用swift app开启可视化推理创空间，参考这里
2. 支持大模型的RM和PPO训练，参考这里
3. 支持SequenceClassification模型（含BERT）的BNB/GPTQ量化，参考这里
4. 支持reward model的推理、部署和BNB/GPTQ量化

新模型

1. ZhipuAI/cogagent-9b-20241220
2. Reward Models: Shanghai_AI_Laboratory/internlm2-1_8b-reward系列, Qwen/Qwen2-Math-RM-72B系列, AI-ModelScope/Skywork-Reward-Llama-3.1-8B系列, AI-ModelScope/GRM_Llama3.1_8B_rewardmodel-ft系列
3. AIDC-AI/Ovis1.6-Gemma2-27B, AIDC-AI/Ovis1.6-Llama3.2-3B
4. PowerInfer/SmallThinker-3B-Preview

新数据集

1. PowerInfer/LONGCOT-Refine-500K, PowerInfer/QWQ-LONGCOT-500K

English Version

New Features

1. Support for using swift app to launch a visual inference creative space, see here
2. Support for RM and PPO training of large models, see here
3. Support for BNB/GPTQ quantization of SequenceClassification models (including BERT), see here
4. Support for inference, deployment, and BNB/GPTQ quantization of reward models

New Models

1. ZhipuAI/cogagent-9b-20241220
2. Reward Models: Shanghai_AI_Laboratory/internlm2-1_8b-reward series, Qwen/Qwen2-Math-RM-72B series, AI-ModelScope/Skywork-Reward-Llama-3.1-8B series, AI-ModelScope/GRM_Llama3.1_8B_rewardmodel-ft series
3. AIDC-AI/Ovis1.6-Gemma2-27B, AIDC-AI/Ovis1.6-Llama3.2-3B
4. PowerInfer/SmallThinker-3B-Preview

New Datasets

1. PowerInfer/LONGCOT-Refine-500K, PowerInfer/QWQ-LONGCOT-500K

What's Changed

• Fix app-ui dropdown by @tastelikefeet in #2787
• fix multi-lora by @Jintao-Huang in #2790
• fix stream infer by @Jintao-Huang in #2793
• fix some web-ui bugs by @tastelikefeet in #2794
• support swift app by @Jintao-Huang in #2792
• fix pt batch infer by @Jintao-Huang in #2800
• fix world_size by @Jintao-Huang in #2801
• update base_model deploy example by @Jintao-Huang in #2803
• fix glm4v by @Jintao-Huang in #2806
• fix swift deploy log error (repeat log) by @Jintao-Huang in #2808
• support ZhipuAI/cogagent-9b-20241220 by @Jintao-Huang in #2810
• fix citest by @Jintao-Huang in #2812
• fix enable_cache by @Jintao-Huang in #2813
• update docs (specific model arguments) by @Jintao-Huang in #2822
• add 'right' option for 'truncation_strategy' by @zsxm1998 in #2754
• Fix glm4v suffix by @Jintao-Huang in #2829
• Update padding side by @Jintao-Huang in #2832
• Update base_to_chat shell by @Jintao-Huang in #2833
• Fix bugs by @Jintao-Huang in #2838
• Fix some bugs by @tastelikefeet in #2848
• support reward_model by @Jintao-Huang in #2849
• Move optimizer to create_optimizer by @tastelikefeet in #2851
• fix post_init by @Jintao-Huang in #2855
• fix cache_name_file by @Jintao-Huang in #2856
• fix telechat template by @Jintao-Huang in #2857
• Update more models by @Jintao-Huang in #2852
• Support quant bert reward by @Jintao-Huang in #2859
• fix jsonl writer by @Jintao-Huang in #2860
• support reward model train by @Jintao-Huang in #2862
• fix vllm video by @Jintao-Huang in #2864
• support mps by @Jintao-Huang in #2866
• Update agent demo by @Jintao-Huang in #2867
• fix bugs by @Jintao-Huang in #2869
• Support ppo by @Jintao-Huang in #2783
• update citest by @Jintao-Huang in #2873
• fix dataset cache bugs by @Jintao-Huang in #2876

New Contributors

• @zsxm1998 made their first contribution in #2754

Full Changelog: v3.0.1...v3.0.2

v3.0.1

中文版

新特性

1. 支持SequenceClassification模型的训练、推理和部署。可以查看以下例子：qwen2.5，bert。
2. LlamaPro支持多模态模型. 例如：qwen2vl、internvl2.5、llama3-vision等。

新模型

1. Qwen/QVQ-72B-Preview
2. iic/DocOwl2
3. OpenGVLab/InternVL2-Pretrain-Models, OpenGVLab/InternVL2_5-4B-AWQ系列, OpenGVLab/InternVL2_5-1B-MPO系列
4. deepseek-ai/DeepSeek-V3系列
5. answerdotai/ModernBERT-base系列
6. AI-ModelScope/paligemma2-3b-pt-224系列, AI-ModelScope/paligemma2-3b-ft-docci-448系列
7. AI-ModelScope/Skywork-o1-Open-Llama-3.1-8B

English Version

New Features:

1. Support for training, inference, and deployment of SequenceClassification models. You can check the following examples: qwen2.5, bert.
2. LlamaPro supports multimodal models, such as qwen2vl, internvl2.5, and llama3-vision.

New Models:

1. Qwen/QVQ-72B-Preview
2. iic/DocOwl2
3. OpenGVLab/InternVL2-Pretrain-Models, OpenGVLab/InternVL2_5-4B-AWQ series, OpenGVLab/InternVL2_5-1B-MPO series
4. deepseek-ai/DeepSeek-V3 series
5. answerdotai/ModernBERT-base series
6. AI-ModelScope/paligemma2-3b-pt-224 series, AI-ModelScope/paligemma2-3b-ft-docci-448 series
7. AI-ModelScope/Skywork-o1-Open-Llama-3.1-8B

What's Changed

• Fix mplug owl2, molmo by @Jintao-Huang in #2724
• fix batch_infer pad_token & florence by @Jintao-Huang in #2725
• Support qwen agent format by @tastelikefeet in #2722
• Support more internvl2.5 awq/mpo & internvl2 pretrain model by @Jintao-Huang in #2726
• support iic/DocOwl2 by @Jintao-Huang in #2728
• update examples by @Jintao-Huang in #2730
• remove files by @Jintao-Huang in #2732
• support paligemma2 by @Jintao-Huang in #2735
• fix windows by @Jintao-Huang in #2733
• support multi-modal llamapro by @tastelikefeet in #2738
• support AI-ModelScope/Skywork-o1-Open-Llama-3.1-8B by @Jintao-Huang in #2739
• Fix windows encoding gbk by @Jintao-Huang in #2741
• fix docs multimodal by @Jintao-Huang in #2742
• support SequenceClassification & update QVQ-72B-Preview by @Jintao-Huang in #2747
• fix web-ui by @Jintao-Huang in #2758
• fix bugs by @Jintao-Huang in #2761
• fix shell by @Jintao-Huang in #2764
• fix app-ui by @tastelikefeet in #2765
• support modern_bert & support bert deploy by @Jintao-Huang in #2767
• fix alpaca by @Jintao-Huang in #2771
• support txt by @Jintao-Huang in #2772
• fix telechat2 template by @Jintao-Huang in #2775
• Fix deepspeed by @Jintao-Huang in #2778
• fix qwen2vl by @Jintao-Huang in #2779
• Fix app ui by @tastelikefeet in #2780
• support deepseek-v3 by @Jintao-Huang in #2781
• Fix app-ui by @tastelikefeet in #2784

Full Changelog: v3.0.0...v3.0.1

v3.0.0

中文版

架构修改与新特性：

具体可以查看这里: https://swift.readthedocs.io/zh-cn/latest/Instruction/ReleaseNote3.0.html

新模型：

1. OpenGVLab/InternVL2_5-1B等系列模型
2. LLM-Research/Llama-3.3-70B-Instruct
3. BAAI/Emu3-Gen
4. deepseek-ai/DeepSeek-V2.5-1210, deepseek-ai/deepseek-vl2等系列模型
5. Shanghai_AI_Laboratory/internlm-xcomposer2d5-ol-7b
6. InfiniAI/Megrez-3b-Instruct, InfiniAI/Megrez-3B-Omni
7. TeleAI/TeleChat2-3B等系列模型

English Version

Architecture Modifications and New Features:

For more details, please visit: https://swift.readthedocs.io/en/latest/Instruction/ReleaseNote3.0.html

New Models:

1. OpenGVLab/InternVL2_5-1B series models
2. LLM-Research/Llama-3.3-70B-Instruct
3. BAAI/Emu3-Gen
4. deepseek-ai/DeepSeek-V2.5-1210, deepseek-ai/deepseek-vl2 series models
5. Shanghai_AI_Laboratory/internlm-xcomposer2d5-ol-7b
6. InfiniAI/Megrez-3b-Instruct, InfiniAI/Megrez-3B-Omni
7. TeleAI/TeleChat2-3B series models

What's Changed

• Refactor All Codes and bump version to 3.0 by @tastelikefeet in #2030
• fix doc by @tastelikefeet in #2545
• fix manifest by @tastelikefeet in #2546
• add doc 2.x by @tastelikefeet in #2548
• fix ui by @tastelikefeet in #2549
• fix infer by @tastelikefeet in #2550
• Refactor mllm by @Jintao-Huang in #2543
• fix ui by @tastelikefeet in #2552
• Fix ui by @tastelikefeet in #2556
• Update ddp infer doc by @Jintao-Huang in #2557
• fix docs by @Jintao-Huang in #2558
• Fix docs by @Jintao-Huang in #2561
• fix log by @tastelikefeet in #2564
• Fix the command line parameter doc by @Jintao-Huang in #2565
• fix context by @Jintao-Huang in #2568
• Documents Updates by @yrk111222 in #2574
• Revert "Documents Updates" by @Jintao-Huang in #2576
• fix hub param by @tastelikefeet in #2572
• Fix bugs by @Jintao-Huang in #2573
• Support internvl2.5 by @Jintao-Huang in #2575
• update english docs by @Jintao-Huang in #2577
• fix en docs by @Jintao-Huang in #2580
• fix docs & add custom example by @Jintao-Huang in #2581
• fix custom example by @Jintao-Huang in #2582
• support llama3.3 by @Jintao-Huang in #2584
• update acc_strategy & fix citest by @Jintao-Huang in #2583
• Support peft0.14 by @tastelikefeet in #2587
• update infer/deploy examples by @Jintao-Huang in #2588
• add image images mapping by @Jintao-Huang in #2594
• update llm sft notebook by @Jintao-Huang in #2599
• fix notebook by @Jintao-Huang in #2600
• Fix streaming by @Jintao-Huang in #2601
• Emu3 gen train by @mi804 in #2602
• compat mllm notebook by @Jintao-Huang in #2604
• Temporarily remove torchacc. by @Jintao-Huang in #2606
• update docs by @Jintao-Huang in #2607
• train and infer scripts for emu3_gen by @mi804 in #2610
• Uodate Document by @yrk111222 in #2615
• update memory usage of emu3-gen by @mi804 in #2611
• move prepare_model by @Jintao-Huang in #2614
• Update mllm notebook by @Jintao-Huang in #2617
• Support all-embedding / all-norm by @Jintao-Huang in #2619
• fix lmdeploy==0.5.* by @Jintao-Huang in #2621
• Support deepseek-ai/DeepSeek-V2.5-1210 by @Jintao-Huang in #2624
• fix use_reentrant gradient_checkpointing by @Jintao-Huang in #2625
• support reward model by @Jintao-Huang in #2628
• fix add_default_tag by @Jintao-Huang in #2631
• fix dataset by @Jintao-Huang in #2636
• fix bugs & update openbuddy models & update docs by @Jintao-Huang in #2638
• fix app-ui by @tastelikefeet in #2641
• Fix post encode by @Jintao-Huang in #2643
• fix bugs by @Jintao-Huang in #2645
• update truncation_strategy by @Jintao-Huang in #2647
• fix swift/Infinity-Instruct by @Jintao-Huang in #2651
• Support LoRA-GA by @lxline in #2650
• support deepseek_vl2 by @Jintao-Huang in #2654
• fix swift/SlimOrca by @Jintao-Huang in #2656
• fix swift/SlimOrca by @Jintao-Huang in #2657
• support Shanghai_AI_Laboratory/internlm-xcomposer2d5-ol-7b:audio by @Jintao-Huang in #2658
• support Shanghai_AI_Laboratory/internlm-xcomposer2d5-ol-7b:base by @Jintao-Huang in #2660
• fix hub by @tastelikefeet in #2661
• fix liger by @tastelikefeet in #2666
• support megrez by @Jintao-Huang in #2667
• fix unsloth resume training by @tastelikefeet in #2668
• fix dataset by @Jintao-Huang in #2670
• Fix bugs by @tastelikefeet in #2671
• fix deepseek_vl2 by @Jintao-Huang in #2675
• support adapters by @Jintao-Huang in #2633
• Support megrez omni by @Jintao-Huang in #2674
• fix docs by @Jintao-Huang in #2679
• fix megrez_omni by @Jintao-Huang in #2680
• fix infer by @Jintao-Huang in #2681
• Fix bugs by @Jintao-Huang in #2687
• Update readme by @Jintao-Huang in #2579
• update wechat by @Jintao-Huang in #2694
• fix readme by @Jintao-Huang in #2696
• Fix web-ui by @tastelikefeet in #2693
• Fix readme by @Jintao-Huang in #2697
• Update banner by @Jintao-Huang in #2699
• fix use_reentrant by @Jintao-Huang in #2700
• update examples by @Jintao-Huang in #2703
• fix eval strategy by @Jintao-Huang in #2707
• Update FAQ by @slin000111 in #2706
• qwen to Qwen by @Jintao-Huang in #2708
• fix timeout & web-ui by @Jintao-Huang in #2709
• Fix multi lora by @tastelikefeet in #2711
• support Qwen/QVQ-72B-Preview by @Jintao-Huang in #2712
• update examples by @Jintao-Huang in #2714
• fix deploy request_config by @Jintao-Huang in #2718
• fix examples by @Jintao-Huang in #2719
• fix gptq group_size by @Jintao-Huang in #2720
• Better error messages by @Jintao-Huang in #2721

New Contributors

• @yrk111222 made their first contribution in #2574
• @lxline made their first contribution in #2650

Full Changelog: v2.6.1...v3.0.0

v2.6.1

New Models:

1. Marco-o1
2. mPLUG-Owl3-7B-241101
3. QwQ-32B-Preview
4. glm-edge, glm-edge-v

New Datasets:

1. OpenO1-SFT

What's Changed

• support part tuner replace_key False by @tastelikefeet in #2438
• bump ms version by @tastelikefeet in #2449
• remove useless code by @tastelikefeet in #2453
• fix qwen2-vl position_ids by @Jintao-Huang in #2461
• fix peft is_multimodal by @Jintao-Huang in #2462
• fix qwen2vl pt infer by @Jintao-Huang in #2463
• [TorchAcc] Update padding strategy when using persistent cache by @eedalong in #2464
• fix kto by @Jintao-Huang in #2478
• Update Common QA by @slin000111 in #2475
• fix awq quant device_map by @Jintao-Huang in #2488
• Fix preprocess num proc by @Jintao-Huang in #2492
• Support marco o1 by @Jintao-Huang in #2496
• fix eval_dataset no by @Jintao-Huang in #2497
• support batch flattening collator by @eedalong in #2499
• fix latex-ocr by @Jintao-Huang in #2510
• support mPLUG-Owl3 241101 by @LukeForeverYoung in #2515
• support qwq by @Jintao-Huang in #2520
• support glm-edge & glm-edge-v by @Jintao-Huang in #2526

New Contributors

• @eedalong made their first contribution in #2464

Full Changelog: v2.6.0...v2.6.1