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...

Patch release v3.11.3

Full Changelog: v3.11.2...v3.11.3

Patch release v3.11.2

Full Changelog: v3.11.1...v3.11.2

Patch release v3.11.1

Full Changelog: v3.11.0...v3.11.1

v3.11.0

中文版

新特性

1. Megatron-SWIFT
   a. 支持 GRPO Megatron 训练，训练文档参考：https://swift.readthedocs.io/zh-cn/latest/Megatron-SWIFT/GRPO.html
   b. FP8 blockwise 训练支持，支持FP8加载和导出权重，训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/megatron/fp8
   c. MTP 训练支持，训练脚本参考：https://github.com/modelscope/ms-swift/blob/main/examples/megatron/lora/mtp.sh
   d. 新模型支持：GPT-OSS，Llama4，InternVL3.5-GPT-OSS等。
   e. 支持 --save_strategy epoch 策略存储模型。
   f. 兼容 megaron-core 0.12-0.15 版本。
2. RL
   a. 新算法 SAPO 支持，文档参考：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/AdvancedResearch/SAPO.html
   b. 新算法 CISPO 支持，文档参考：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/AdvancedResearch/CISPO.html
   c. 缓解训推不一致的算法支持，包括 TIS/MIS 与 rollout off-policy metrics 记录，文档参考：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/AdvancedResearch/training_inference_mismatch.html
   d. tree-rollout 支持，文档参考：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/AdvancedResearch/treepo.html （感谢招商银行团队 @li2zhi 的贡献）
   e. gkd 训练支持使用 liger_kernel loss（--use_liger_kernel true）。
   f. 新增 GRPO loss_type，文档参考：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/DeveloperGuide/loss_types.html
3. 训练
   a. cached dataset 重构，更好支持大型数据集离线 tokenize 场景，脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/cached_dataset
   b. 预训练场景 --truncation_strategy split 策略支持，将长文本切成多条数据样本避免 tokens 浪费。
   c. packing_num_proc 参数支持。
   d. Qwen2.5-VL系列模型兼容使用 "qwen_vl_utils>=0.14"。
   e. MFU 日志插件支持。(感谢 @y2logic 的贡献)
4. 国产化硬件（感谢昇腾和招商银行技术团队的贡献）
   a. Megatron-SWIFT 支持昇腾 NPU，文档参考：https://swift.readthedocs.io/zh-cn/latest/BestPractices/NPU-support.html
   b. 昇腾NPU混合算子支持 Qwen2、Qwen3、Qwen3-MoE 系列模型，加速训练过程。

新模型

1. 纯文本模型：
   a. moonshotai/Kimi-K2-Thinking
2. 多模态模型：
   a. SenseNova/SenseNova-SI-InternVL3-2B系列
   b. mistralai/Ministral-3-3B-Instruct-2512系列
   c. mistralai/Mistral-Small-3.2-24B-Instruct-2506

English Version

New Features

1. Megatron-SWIFT
   a. GRPO training support on Megatron, documentation: https://swift.readthedocs.io/en/latest/Megatron-SWIFT/GRPO.html
   b. FP8 blockwise training support, including FP8 weight loading and exporting. Training scripts: https://github.com/modelscope/ms-swift/tree/main/examples/megatron/fp8
   c. MTP training support, training script: https://github.com/modelscope/ms-swift/blob/main/examples/megatron/lora/mtp.sh
   d. New model support: GPT-OSS, Llama4, InternVL3.5-GPT-OSS, etc.
   e. Support for saving strategy --save_strategy epoch.
   f. Compatible with megaron-core versions 0.12–0.15.
2. RL
   a. New algorithm SAPO supported, documentation: https://swift.readthedocs.io/en/latest/Instruction/GRPO/AdvancedResearch/SAPO.html
   b. New algorithm CISPO supported, documentation: https://swift.readthedocs.io/en/latest/Instruction/GRPO/AdvancedResearch/CISPO.html
   c. Algorithms for mitigating training–inference mismatch, including TIS/MIS and rollout off-policy metrics. Docs: https://swift.readthedocs.io/en/latest/Instruction/GRPO/AdvancedResearch/training_inference_mismatch.html
   d. Tree-rollout support, docs: https://swift.readthedocs.io/en/latest/Instruction/GRPO/AdvancedResearch/treepo.html (Thanks to CMB team @li2zhi for the contribution)
   e. GKD training supports liger_kernel loss (--use_liger_kernel true).
   f. New GRPO loss types added, docs: https://swift.readthedocs.io/en/latest/Instruction/GRPO/DeveloperGuide/loss_types.html
3. Training
   a. Cached dataset refactoring for better offline tokenization of large datasets. Scripts: https://github.com/modelscope/ms-swift/tree/main/examples/train/cached_dataset
   b. Pretraining --truncation_strategy split support, splitting long text into multiple samples to avoid token waste.
   c. Added packing_num_proc parameter support.
   d. Qwen2.5-VL series models compatible with "qwen_vl_utils>=0.14".
   e. MFU logging plugin support (Thanks to @y2logic).
4. Domestic Hardware Support (Thanks to Ascend and CMB technical teams)
   a. Megatron-SWIFT supports Ascend NPU, documentation: https://swift.readthedocs.io/en/latest/BestPractices/NPU-support.html
   b. Ascend NPU mixed operators support Qwen2, Qwen3, Qwen3-MoE series models, accelerating training.

New Models

1. Text-only models:
   a. moonshotai/Kimi-K2-Thinking
2. Multimodal models:
   a. SenseNova/SenseNova-SI-InternVL3-2B series
   b. mistralai/Ministral-3-3B-Instruct-2512 series
   c. mistralai/Mistral-Small-3.2-24B-Instruct-2506

What's Changed

• bump version 3.11.0.dev by @Jintao-Huang in #6560
• [model] support Kimi-K2 by @Jintao-Huang in #6562
• [bugfix] fix pp vit_lr by @Jintao-Huang in #6565
• [bugfix] fix tools parse in gkd/grpo server mode by @hjh0119 in #6568
• [bugfix] fix grpo with reward model by @hjh0119 in #6567
• [bugfix] fix mcore-bridge vpp by @Jintao-Huang in #6581
• qwen2.5-vl compat qwen_vl_utils version by @Jintao-Huang in #6584
• [bugfix] fix packing_length by @Jintao-Huang in #6594
• [dataset] support packing_num_proc by @Jintao-Huang in #6592
• Fix emb loss scale by @tastelikefeet in #6597
• [megatron] compat megatron-core 0.12-0.14 by @Jintao-Huang in #6599
• [kto] fix kto loss_type=apo_zero_unpaired by @Jintao-Huang in #6601
• Fix command line display for UI by @slin000111 in #6603
• Support Megatron GRPO by @hjh0119 in #6025
• [megatron] fix train_iters by @Jintao-Huang in #6611
• [bugfix] fix modelscope patch_hub by @Jintao-Huang in #6612
• [template] support add_eos by @Jintao-Huang in #6613
• [dataset] refactor cached_dataset by @Jintao-Huang in #6561
• [bugfix]fix add_eos in gkd/grpo for truncated sample encode by @hjh0119 in #6618
• Support GKD Liger Kernel Loss by @hjh0119 in #6619
• Support generative reranker right pad by @0russwest0 in #6573
• update swift image 3.10.1 by @Jintao-Huang in #6622
• [model] support mistral 2506 by @Jintao-Huang in #6624
• update peft version by @Jintao-Huang in #6621
• [bugfix] Fix multinode write conflict mcore-bridge (deepseek-v3) by @Jintao-Huang in #6626
• Initialize chord dataset after accelerator setup in GRPOTrainer by @tongchen126 in #6638
• [bugfix] fix megatron grpo max_epochs by @hjh0119 in #6646
• [bugfix] fix megatron grpo server mode sync weight by @hjh0119 in #6648
• [megatron] fix save barrier by @Jintao-Huang in #6653
• [bugfix] fix megatron grpo rollout_group by @hjh0119 in #6655
• [bugfix] fix chatml chat template by @Jintao-Huang in #6656
• [bugfix] fix train_type full freeze_llm by @Jintao-Huang in #6651
• [mcore-bridge] optimize gpt_bridge comm by @Jintao-Huang in #6659
• [algo] support cispo algorithm by @hjh0119 in #6572
• [model] Support SenseNova-SI by @hjh0119 in #6657
• [megatron] fix swift export merge_lora by @Jintao-Huang in #6664
• [bugfix] memory log is missing on Ascend NPU by @baymax591 in #6647
• update doc by @tastelikefeet in #6665
• [bugfix] Fix GKD with TRL >= 0.24 & GKD Liger by @hjh0119 in #6663
• [template] support truncation_strategy spllit (swift pt) by @Jintao-Huang in #6672
• [bugfix] fix qwen3_omni seq_cls by @Jintao-Huang in #6673
• [bugfix] getattr error for activation_offloading in RM training by @hjh0119 in #6677
• [bugfix] fix liger-kernel version check by @hjh0119 in #6679
• [bugfix] fix qwen3_vl image_list fps by @Jintao-Huang in #6696
• [bugfix] fix logprobs in vllm sampling params by @hjh0119 in #6698
• [megatron] support global_aux_loss by @Jintao-Huang in #6699
• [bugfix] fix megatron grpo local jsonl writer by @hjh0119 in #6700
• fix type_type=rm eval trl>=0.25 by @Jintao-Huang in #6701
• add npu fsdp example by @addsubmuldiv in #6697
• add npu deepspeed example by @addsubmuldiv in https://github.com/model...