Patch release v4.0.2

Full Changelog: v4.0.1...v4.0.2

Patch release v4.0.1

Full Changelog: v4.0.0...v4.0.1

v4.0.0

中文版

新特性

1. 架构优化
   a. 目录结构重构与依赖关系优化，使用模块化设计，提升架构的可扩展性和可定制性。
   b. model_type与template解耦，简化同一 model_type 含多个 template 的模型支持流程。
   c. Megatron-SWIFT 训练循环重写，使用 megatron-core 替代 megatron-lm 依赖。（兼容Ascend NPU）
2. Megatron-SWIFT
   a. 新模型支持：Qwen3.5系列、GLM4.7-Flash、MiniMax-M2.1、OLMoE。
   b. Embedding 任务支持，训练示例：https://github.com/modelscope/ms-swift/tree/main/examples/megatron/embedding
   c. Reranker 任务支持，训练示例：https://github.com/modelscope/ms-swift/tree/main/examples/megatron/reranker
   d. 新增save_total_limit参数，自动清理过期 checkpoint，并保留指标最优和最新的权重。
   e. Qwen3-Next/Qwen3.5 新增apply_wd_to_qk_layernorm参数，支持对 qk layernorm 应用权重衰减。
   f. 多模态MoE模型lora支持 --target_modules all-router 配置。
3. RL
   a. 支持GDPO算法计算优势，使用参数--scale_rewards gdpo。（感谢 @Auraithm 的贡献）
   b. GKD 支持使用 top-k logits 计算KL以节约显存，使用参数 --gkd_topk_logits。
   c. GKD 支持使用 teacher server，避免显式加载教师模型。
4. 训练
   a. 新增 muon clip 优化器支持，训练示例：https://github.com/modelscope/ms-swift/blob/main/examples/train/optimizer/muonclip.sh （感谢 @vx120 的贡献）
   b. 依赖更新：兼容最新依赖 python3.12, transformers 5.2.0, vllm 0.15.1, trl 0.28, liger-kernel 0.7.0等。
   c. generative reranker lm_head 部分计算优化，降低显存占用。
   d. fsdp2支持激活 cpu offload；deepspeed elastic支持。（感谢招商 @meichangsu1 的贡献）

新模型

1. 纯文本模型
   a. Qwen/Qwen3-Coder-Next
   b. ZhipuAI/GLM-4.7-Flash, ZhipuAI/GLM-5
   c. MiniMaxAI/MiniMax-M2.1
   d. Tencent-YouTu-Research/Youtu-LLM-2B
   e. IQuestLab/IQuest-Coder-V1-40B-Instruct
   f. allenai/OLMoE-1B-7B-0924-Instruct系列（感谢 @qianhao0713 的贡献）
2. 多模态模型
   a. Qwen/Qwen3.5-35B-A3B, Qwen/Qwen3.5-9B 系列。训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/models/qwen3_5
   b. Qwen3-VL-Embedding, Qwen3-VL-Reranker。训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/embedding/qwen3, https://github.com/modelscope/ms-swift/tree/main/examples/train/reranker/qwen3
   c. deepseek-ai/DeepSeek-OCR-2
   d. ZhipuAI/GLM-OCR
   e. PaddlePaddle/PaddleOCR-VL-1.5
   f. OpenBMB/MiniCPM-o-4_5
   g. stepfun-ai/Step3-VL-10B
   h. google/medgemma-4b-it 系列

English Version

New Features

1. Architecture Optimization
   a. Directory structure refactoring and dependency optimization with modular design to enhance architecture scalability and customizability.
   b. Decoupling of model_type and template to simplify support for models with multiple templates under the same model_type.
   c. Rewritten Megatron-SWIFT training loop using megatron-core instead of megatron-lm dependency. (Compatible with Ascend NPU)
2. Megatron-SWIFT
   a. New model support: Qwen3.5 series, GLM4.7-Flash, MiniMax-M2.1, OLMoE.
   b. Embedding task support. Training example: https://github.com/modelscope/ms-swift/tree/main/examples/megatron/embedding
   c. Reranker task support. Training example: https://github.com/modelscope/ms-swift/tree/main/examples/megatron/reranker
   d. Added save_total_limit parameter to automatically clean up expired checkpoints while retaining the best-performing and latest weights.
   e. Added apply_wd_to_qk_layernorm parameter for Qwen3-Next/Qwen3.5 to support weight decay on qk layernorm.
   f. Multi-modal MoE model LoRA supports --target_modules all-router configuration.
3. RL
   a. Support for GDPO algorithm to compute advantages using parameter --scale_rewards gdpo. (Thanks to @Auraithm)
   b. GKD supports using top-k logits to compute KL for memory savings with parameter --gkd_topk_logits.
   c. GKD supports using teacher server to avoid explicitly loading the teacher model.
4. Training
   a. Added Muon-CLIP optimizer support. Training example: https://github.com/modelscope/ms-swift/blob/main/examples/train/optimizer/muonclip.sh (Thanks to @vx120)
   b. Dependency updates: Compatible with latest dependencies including python3.12, transformers 5.2.0, vllm 0.15.1, trl 0.28, liger-kernel 0.7.0, etc.
   c. Optimized generative reranker lm_head computation to reduce memory usage.
   d. FSDP2 supports CPU offload activation; DeepSpeed elastic support. (Thanks to @meichangsu1)

New Models

1. Text-only Models
   a. Qwen/Qwen3-Coder-Next
   b. ZhipuAI/GLM-4.7-Flash, ZhipuAI/GLM-5
   c. MiniMaxAI/MiniMax-M2.1
   d. Tencent-YouTu-Research/Youtu-LLM-2B
   e. IQuestLab/IQuest-Coder-V1-40B-Instruct
   f. allenai/OLMoE-1B-7B-0924-Instruct series (Thanks to @qianhao0713)
2. Multi-modal Models
   a. Qwen/Qwen3.5-35B-A3B, Qwen/Qwen3.5-9B series. Training scripts: https://github.com/modelscope/ms-swift/tree/main/examples/models/qwen3_5
   b. Qwen3-VL-Embedding, Qwen3-VL-Reranker. Training scripts: https://github.com/modelscope/ms-swift/tree/main/examples/train/embedding/qwen3, https://github.com/modelscope/ms-swift/tree/main/examples/train/reranker/qwen3
   c. deepseek-ai/DeepSeek-OCR-2
   d. ZhipuAI/GLM-OCR
   e. PaddlePaddle/PaddleOCR-VL-1.5
   f. OpenBMB/MiniCPM-o-4_5
   g. stepfun-ai/Step3-VL-10B
   h. google/medgemma-4b-it series

What's Changed

• [misc] update swift patch_conv3d by @Jintao-Huang in #7320
• add npu megatron multi-node example by @addsubmuldiv in #7321
• [bugfix] fix megatron convert by @Jintao-Huang in #7323
• [model] Support Qwen3-VL-Embedding/Qwen3-VL-Reranker by @Jintao-Huang in #7329
• [reranker] refactor reranker by @Jintao-Huang in #7334
• [bugfix] fix video base64 torchcodec by @Jintao-Huang in #7338
• [bugfix] fix modelopt by @Jintao-Huang in #7339
• [docs] Update swift image 3.12 by @Jintao-Huang in #7332
• [bugfix] fix get_chunked_inputs slice by @hjh0119 in #7346
• fix find node ip by @tastelikefeet in #7350
• Fix multi-modal reranker doc by @tastelikefeet in #7354
• [bugfix] fix app_args by @Jintao-Huang in #7367
• [bugfix] fix qwen2_vl video by @Jintao-Huang in #7376
• [bugfix] fix vllm moe model load_weights by @hjh0119 in #7362
• [v4] refactor ms-swift v4 by @Jintao-Huang in #7238
• feat: support scale rewards "gdpo" by @Auraithm in #7348
• [infer] infer backend pt -> transformers by @Jintao-Huang in #7379
• [docs] update docs & update Copyright by @Jintao-Huang in #7384
• Fix device mismatch in _forward_qwen3_vl_or_qwen3_omni when computing visual_pos_masks by @yaqiangsun in #7372
• add npu qwen3-next example and warning of ep size by @addsubmuldiv in #7390
• [bugfix] fix deepseek_v3_1 thinking template by @Jintao-Huang in #7388
• [docs] update docs & update dataset 'loss' by @Jintao-Huang in #7402
• [bugfix] Fix ref adapters trainable params 0 by @Jintao-Huang in #7403
• [readme] update error timeline of news by @shizhengLi in #7404
• [bugfix] fix sp reranker by @Jintao-Huang in #7405
• [v4] fix ci by @Jintao-Huang in #7559
• [refactor] reorganize reward and rollout modules into dedicated direct… by @hjh0119 in #7397
• [grpo] speedup grpo train stage encode with concurrent by @Cccei000 in #7391
• Update the NPU-supported features table by @addsubmuldiv in #7562
• [bugfix] fix attn_impl by @Jintao-Huang in #7564
• [v4] refactor ms-swift v4 (pipelines/arguments/swiftmixin/callback/tuner_plugin) by @Jintao-Huang in #7385
• [bugfix] fix minimax tp by @Jintao-Huang in #7788
• fix inputs_embeds for hunyuanOCR by @slin000111 in #7803
• [bugfix] fix deepspeed distributed weight offload code by @Silas-11 in #7802
• [generative_reranker] generative reranker logits memory optimization by @Jintao-Huang in #7816
• update requirements by @Jintao-Huang in #7819
• [misc] update issue template by @Jintao-Huang in #7818
• [bugfix] fix dpo by @Jintao-Huang in #7824
• update wechat by @tastelikefeet in #7827
• [bugfix] fix deepspeed optimizer offload code by @Silas-11 in #7821
• [model] support glm4_moe_lite by @Jintao-Huang in #7829
• [bugfix] fix hunyuan ocr by @Jintao-Huang in #7831
• [megatron] support glm_moe_lite by @Jintao-Huang in #7833
• chore: epochs -> epoch by @zzc0430 in #7825
• [optimizer] Set loss mask to compute the loss for multi-turn reasoning by @Simon-ss7 in #7838
• [bugfix] fix recompute_granularity none by @Jintao-Huang in https://gi...

Patch release v3.12.6

What's Changed

• [bugfix] fix grpo move_model_batches by @hjh0119 in #8091

Full Changelog: v3.12.5...v3.12.6

Patch release v3.12.5

Full Changelog: v3.12.4...v3.12.5

Patch release v3.12.4

Full Changelog: v3.12.3...v3.12.4

Patch release v3.12.3

Full Changelog: v3.12.2...v3.12.3

Patch release v3.12.2

Full Changelog: v3.12.1...v3.12.2

v3.12.1

What's Changed

• [bugfix] fix glm4_7 agent_template by @Jintao-Huang in #7256
• [bugfix] fix DeepSeek-OCR vllm deploy by @hjh0119 in #7258
• [feat] add async reward function support for GRPO training by @hjh0119 in #7252
• [model] support medgemma by @slin000111 in #7261
• [megatron] Support MiniMaxAI/MiniMax-M2.1 by @Jintao-Huang in #7262
• Support muonclip optimizer by @vx120 in #7191
• add task_type by @slin000111 in #7265
• [bugfix] fix mtp save by @Jintao-Huang in #7267
• [feat] support megatron grpo entropy mask & log by @hjh0119 in #7263
• [model] support iquestcoder by @Jintao-Huang in #7271
• [bugfix] fix reward model adapters by @hjh0119 in #7293
• Fix the issue of repeated inference in multi-turn scheduler. by @Simon-ss7 in #7279
• [bugfix] auto-enable async engine for vLLM encode tasks by @hjh0119 in #7301
• [bugfix] fix vllm_engine load_format by @Jintao-Huang in #7302
• fix npu megatron cp by @addsubmuldiv in #7299
• [misc] Remove unnecessary clone operations during weight synchronization by @hjh0119 in #7308
• [model] support youtu-llm by @hjh0119 in #7306
• [megatron] fix gpt_bridge oom by @Jintao-Huang in #7310
• [misc] fix youtu agent template type-checking by @hjh0119 in #7311
• [bugfix] Fix duplicate 'load_format' argument being passed in rollout by @hjh0119 in #7312

New Contributors

• @Simon-ss7 made their first contribution in #7279

Full Changelog: v3.12.0...v3.12.1

v3.12.0

中文版

新特性

1. Megatron-SWIFT
   a. GKD算法支持Megatron训练，文档参考：https://swift.readthedocs.io/zh-cn/latest/Megatron-SWIFT/GKD.html
   b. 新模型支持：GLM4 Dense; GLM4.7; GLM4.6v-Flash, GLM-4.1V。
   c. save_safetensors 支持断点续训，将 Mcore-Bridge 加载和存储方式作为推荐方式。
   d. 非 padding-free 训练模式支持更多训练阶段：GRPO/DPO/KTO/RM/序列分类。
   e. group_by_length 参数支持，将数据集长度大致相同的样本分组在一起（含随机因素），加速非packing模式下训练速度。
   f. 支持 --report_to 参数，将训练日志在 wandb/swanlab 中记录并可视化。
   g. Qwen3-Next 使用 Zero-Centered RMSNorm，与 transformers 对齐。
   h. train_dataloader_shuffle 参数支持，控制训练数据集是否随机。
   i. template.encode 新增重试机制，避免 megatron 训练因网络问题获取图片/视频报错而卡住。
2. RL
   a. 增加 Off-Policy Sequence Masking (from DeepSeek-V3.2)，文档参考：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/AdvancedResearch/training_inference_mismatch.html#off-policy-sequence-masking
   b. GRPO 增加参数 num_generations_eval 设置 eval 阶段的生成数量。
   c. 优化 GKD loss 计算的显存峰值。
   d. GRPO/GKD server mode 支持使用 ipv6 地址。
   e. 支持使用 structured_outputs_regex 进行结构化输出采样。
3. 训练
   a. embedding/reranker/序列分类任务支持序列 packing 和序列并行。训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/sequence_parallel
   b. 支持 --fsdp fsdp2 使用 ms-swift 内置的 FSDP2 配置文件。
   c. loss_scale 支持3种基本策略：'default'、'last_round'、'all'与其他策略的混合使用，例如：'last_round+ignore_empty_think'。
   d. cached_dataset 支持 embedding/reranker/序列分类训练任务，训练脚本参考https://github.com/modelscope/ms-swift/tree/main/examples/train/cached_dataset
   e. thinking template 重构，ThinkingTemplate 功能合入 Template，新增enable_thinking, add_non_thinking_prefix参数。
   f. 新增 SWIFT_PATCH_CONV3D 环境变量，避免 torch2.9 环境跑 conv3d 缓慢的问题。
   g. 支持 swanlab_notification_method 参数，在训练完成/发生错误时，指定 swanlab 通知方式。
   h. dataloader_prefetch_factor 参数默认值从10修改为2。
4. 国产化硬件（感谢昇腾和招商银行技术团队的贡献）
   a. 新增更多训练脚本：https://github.com/modelscope/ms-swift/tree/main/examples/ascend
   b. Qwen3-VL 混合算子支持，具体查看这个PR：#7079
   c. 更新 Megatron-SWIFT NPU 性能采集/精度采集相关文档，参考这里：https://swift.readthedocs.io/zh-cn/latest/Megatron-SWIFT/Ascend.html

新模型

1. 纯文本模型：
   a. ZhipuAI/GLM-4.7系列
   b. iic/QwenLong-L1.5-30B-A3B
   c. gongjy/MiniMind2 （感谢 @PiggerZZM 的贡献）
2. 多模态模型：
   a. ZhipuAI/GLM-4.6V; ZhipuAI/GLM-4.6V-Flash系列
   b. Tencent-Hunyuan/HunyuanOCR

English Version

New Features

1. Megatron-SWIFT
   a. GKD algorithm supports Megatron training. Documentation reference: https://swift.readthedocs.io/en/latest/Megatron-SWIFT/GKD.html
   b. New model support: GLM4 Dense; GLM4.7; GLM4.6v-Flash, GLM-4.1V.
   c. save_safetensors supports checkpoint resumption, with Mcore-Bridge loading and storage method as the recommended approach.
   d. Non-padding-free training mode supports more training stages: GRPO/DPO/KTO/RM/sequence classification.
   e. group_by_length parameter support, grouping samples with similar lengths in the dataset together (with random factors) to accelerate training speed in non-packing mode.
   f. Support for --report_to parameter to log and visualize training logs in wandb/swanlab.
   g. Qwen3-Next uses Zero-Centered RMSNorm, aligned with transformers.
   h. train_dataloader_shuffle parameter support to control whether training dataset is shuffled.
   i. Added retry mechanism to template.encode to prevent megatron training from freezing due to network issues when fetching images/videos.
2. RL
   a. Added Off-Policy Sequence Masking (from DeepSeek-V3.2). Documentation reference: https://swift.readthedocs.io/en/latest/Instruction/GRPO/AdvancedResearch/training_inference_mismatch.html#off-policy-sequence-masking
   b. GRPO adds num_generations_eval parameter to set the number of generations during eval stage.
   c. Optimized memory peak for GKD loss calculation.
   d. GRPO/GKD server mode supports using ipv6 addresses.
   e. Support for structured output sampling using structured_outputs_regex.
3. Training
   a. Embedding/reranker/sequence classification tasks support sequence packing and sequence parallelism. Training script reference: https://github.com/modelscope/ms-swift/tree/main/examples/train/sequence_parallel
   b. Support for --fsdp fsdp2 to use ms-swift built-in FSDP2 configuration file.
   c. loss_scale supports 3 basic strategies: 'default', 'last_round', 'all' and their hybrid use with other strategies, e.g., 'last_round+ignore_empty_think'.
   d. cached_dataset supports embedding/reranker/sequence classification training tasks. Training script reference: https://github.com/modelscope/ms-swift/tree/main/examples/train/cached_dataset
   e. Thinking template refactored, ThinkingTemplate functionality merged into Template, added enable_thinking and add_non_thinking_prefix parameters.
   f. Added SWIFT_PATCH_CONV3D environment variable to avoid slow conv3d execution in torch2.9 environment.
   g. Support for swanlab_notification_method parameter to specify swanlab notification method when training completes/errors occur.
   h. dataloader_prefetch_factor parameter default value changed from 10 to 2.
4. Domestic Hardware (Thanks to Ascend and CMB technical teams)
   a. Added more training scripts: https://github.com/modelscope/ms-swift/tree/main/examples/ascend
   b. Qwen3-VL hybrid operator support, see this PR: #7079
   c. Updated Megatron-SWIFT NPU performance collection/accuracy collection documentation, reference: https://swift.readthedocs.io/en/latest/Megatron-SWIFT/Ascend.html

New Models

1. Text-only models:
   a. ZhipuAI/GLM-4.7 series
   b. iic/QwenLong-L1.5-30B-A3B
   c. gongjy/MiniMind2 (Thanks to @PiggerZZM's contribution)
2. Multimodal models:
   a. ZhipuAI/GLM-4.6V; ZhipuAI/GLM-4.6V-Flash series
   b. Tencent-Hunyuan/HunyuanOCR

What's Changed

• [model] Support GLM4.6-V by @Jintao-Huang in #6948
• [model] support glm4_6v flash by @Jintao-Huang in #6959
• [bugfix] fix truncation_strategy left by @Jintao-Huang in #6961
• [bugfix] fix megatron save_checkpoint by @Jintao-Huang in #6963
• [feat] GKD support truncation strategy delete to resample by @hjh0119 in #6964
• [misc] megatron grpo check rollout_logps by @hjh0119 in #6970
• [misc] set default group_port for vllm client by @hjh0119 in #6972
• [grpo] support Off-Policy Sequence Masking by @hjh0119 in #6978
• [megatron, misc] support check_latest_model by @hjh0119 in #6988
• [bugfix] fix reranker_padding_free by @Jintao-Huang in #6989
• [megatron] fix eval_iters 1 by @Jintao-Huang in #6990
• Add dense_npu.sh for megatron lora training in huawei npu by @vx120 in #6976
• fix system swift pt by @Jintao-Huang in #7003
• [bugfix] fix qwen_vl_utils torchvision base64 by @Jintao-Huang in #7004
• [bugfix] fix liger_kernel flash_attn by @Jintao-Huang in #7005
• [bugfix] fix qwen3_vl bridge by @Jintao-Huang in #7006
• [bugfix] fix reranker padding_free & fix seq_cls omni padding_free by @Jintao-Huang in #7007
• [npu] add npu qwen3_omni sft example for mindspeed backend by @tongtong0613 in #7008
• [bugfix] qwen-omni3 vllm infer with USE_AUDIO_IN_VIDEO by @hjh0119 in #7009
• [bugfix] fix grpo sleep_level 2 causes gibberish outputs by @hjh0119 in #7017
• add npu vllm-ascend docs and examples by @addsubmuldiv in #7013
• [compat] fix mcore012 compat torch new by @Jintao-Huang in #7021
• [megatron] Megatron support random/non-random dataloader by @Jintao-Huang in #7016
• [bugfix] megatron add retry to avoid hang by @Jintao-Huang in #7023
• [trainer] refactor acc metrics by @Jintao-Huang in #7026
• [infer] update embddding/reranker demo by @Jintao-Huang in #7029
• [train] support embeding/reranker packing & support reranker/embedding cache_dataset by @Jintao-Huang in #6987
• update readme by @Jintao-Huang in #7033
• [misc] update swift image by @Jintao-Huang in #7039
• [bugfix] remove add_eos for rm in grpo by @hjh0119 in #7040
• [npu] Fix device mismatch in weight sync for HCCL communicator by @singing4you in #7036
• collect npu profiling data by @OneMondy in #6977
• [bugfix] fix null_ref_context by @Jintao-Huang in #7042
• [model] support hunyuan_ocr by @slin000111 in #7038
• update flash_attn version; fix mcore 0.15 hang by @Jintao-Huang in #7043
• [bugfix] fix grpo multi turn log_entropy by @hjh0119 in #7044
• [bugfix] fix dataloader megatron by @Jintao-Huang in #7050
• [grpo] support num_generations_eva...