v4.4.0

中文版

新特性

1. RL
   a. GRPO 支持 OPD RL 训练。参考文档：https://swift.readthedocs.io/zh-cn/latest/Instruction/Distillation.html
   b. GRPO/GKD 支持 Multi-teacher 训练。参考文档：https://swift.readthedocs.io/zh-cn/latest/Instruction/Distillation.html
   c. 新增多模态 vLLM 对齐测试。参考 issue：#9668
   d. GRPO/GKD Trainer 完成重构。
2. Megatron-SWIFT
   a. 新增 model_type 支持：GLM-5.2（glm_moe_dsa 新增 indexer_type = 'shared' 支持）、minicpmv4_6。
   b. NPU Megatron-Core 版本升级至 0.16。（感谢 @addsubmuldiv 的贡献）
   c. moe_router_load_balancing_type 支持同时指定多个类型。
   d. 修复 DeepSeek-V4 FP8 相关问题。
3. 训练
   a. 新增 DLLM 训练支持（DiffusionGemma）。训练脚本参考：https://github.com/modelscope/ms-swift/blob/main/examples/models/gemma4/diffusion_gemma.sh
   b. 新增 Qwen3-TTS 训练支持。训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/models/qwen3_tts
   c. Embedding 模型新增 MRL 支持，引入 mrl_dims 参数。
   d. Loss Scale 支持 ConcatLossScale，可将多种 loss scale 策略叠加使用，例如：last_round+ignore_empty_think+hermes。
   e. Qwen3.5、minicpmv4.6 支持 padding_free/packing。

新模型

1. 纯文本模型
   a. ZhipuAI/GLM-5.2
2. 多模态模型
   a. moonshotai/Kimi-K2.7-Code
   b. MiniMaxAI/MiniMax-M3
   c. PaddlePaddle/Unlimited-OCR
   d. google/diffusiongemma-26B-A4B-it

English Version

New Features

1. RL
   a. GRPO now supports OPD RL training. Reference doc: https://swift.readthedocs.io/zh-cn/latest/Instruction/Distillation.html
   b. GRPO/GKD now supports Multi-teacher training. Reference doc: https://swift.readthedocs.io/zh-cn/latest/Instruction/Distillation.html
   c. Added multimodal vLLM alignment tests. Reference issue: #9668
   d. GRPO/GKD Trainer has been refactored.
2. Megatron-SWIFT
   a. Added new model_type support: GLM-5.2 (glm_moe_dsa adds indexer_type = 'shared' support) and minicpmv4_6.
   b. NPU Megatron-Core upgraded to version 0.16. (Thanks to @addsubmuldiv for the contribution.)
   c. moe_router_load_balancing_type now supports specifying multiple types simultaneously.
   d. Fixed DeepSeek-V4 FP8 related issues.
3. Training
   a. Added DLLM training support (DiffusionGemma). Training script reference: https://github.com/modelscope/ms-swift/blob/main/examples/models/gemma4/diffusion_gemma.sh
   b. Added Qwen3-TTS training support. Training script reference: https://github.com/modelscope/ms-swift/tree/main/examples/models/qwen3_tts
   c. Added MRL support for Embedding models, introducing the mrl_dims parameter.
   d. Loss Scale now supports ConcatLossScale, enabling multiple loss scale strategies to be stacked, e.g.: last_round+ignore_empty_think+hermes.
   e. Qwen3.5 and minicpmv4.6 now support padding_free/packing.

New Models

1. Text-Only Models
   a. ZhipuAI/GLM-5.2
2. Multimodal Models
   a. moonshotai/Kimi-K2.7-Code
   b. MiniMaxAI/MiniMax-M3
   c. PaddlePaddle/Unlimited-OCR
   d. google/diffusiongemma-26B-A4B-it

What's Changed

• [doc] update ray doc by @hjh0119 in #9527
• support embedding MRL by @Jintao-Huang in #9528
• [1/N] Refactor SP by @Jintao-Huang in #9509
• [bugfix] fix megatron server mode by @hjh0119 in #9531
• Add NPU-compatible select_device() and new test cases for data preprocessing by @Ginray in #9512
• update gemma4 agent_template by @Jintao-Huang in #9532
• [megatron] update mcore requirements by @Jintao-Huang in #9536
• support concat_loss_scale by @Jintao-Huang in #9537
• Update doc mindspeed by @addsubmuldiv in #9542
• Update dataset by @Jintao-Huang in #9543
• fix swanlab get_run by @hjh0119 in #9547
• [model] support kimi_k27 by @Jintao-Huang in #9552
• [bugfix] pin MeanMetric all_reduce tensor to float32 to avoid cross-rank dtype mismatch by @HaozheZhang6 in #9551
• update npu scripts for qwen3 omni and 3.5 moe by @hazelduan in #9540
• [model] support diffusion_gemma4 by @Jintao-Huang in #9548
• revert select_device by @Jintao-Huang in #9559
• enable_thinking is set to False for hybrid thinking models. by @Jintao-Huang in #9557
• [template] Add disable ignore empty think by @Jintao-Huang in #9558
• [megatron] megatron remove num_samples by @Jintao-Huang in #9522
• [model] support minimax_m3 by @Jintao-Huang in #9553
• [bugfix] fix minimax_m3 by @Jintao-Huang in #9568
• Add Qwen3.5 NPU A3 16die script by @hazelduan in #9569
• [transformers] support qwen3_5 padding_free/packing by @Jintao-Huang in #9571
• fix gemma4-12b weight sync by @hjh0119 in #9572
• [doc] router replay by @hjh0119 in #9579
• [megatron] update cross_entropy_fusion_impl default value 'te' -> 'native' (#9578) by @Jintao-Huang in #9578
• Add logger initialization to main CLI by @rivergod in #9577
• [model] support GLM 5.2 by @Jintao-Huang in #9581
• [utils] update utils lazy_init by @Jintao-Huang in #9584
• [docs] update deepseek_v4 docs by @Jintao-Huang in #9585
• [docs] update docs by @Jintao-Huang in #9586
• [doc] sort grpo doc index by @hjh0119 in #9587
• [docs] update swift 4.3.1 image by @Jintao-Huang in #9574
• Fix DDP fp32 grad reduce arg mapping by @addsubmuldiv in #9590
• feat(megatron): expose muon_coefficient_type for the Muon optimizer by @yuchenwang3 in #9591
• [ci] introduce uv to accelerate pip install by @KadenZhang3321 in #9297
• fix(muon): muon_use_nesterov silently dropped (arg/config name mismatch) by @yuchenwang3 in #9599
• [bugfix] fix ddp_config by @Jintao-Huang in #9592
• [bugfix] fix gemma4 system by @Jintao-Huang in #9604
• Fix NPU Qwen3.5 grpo bugs by @addsubmuldiv in #9589
• [compat] compat moe_router_load_balancing_type mcore>=0.16 by @Jintao-Huang in #9603
• [1/N] refactor grpo/gkd by @hjh0119 in #9580
• [misc] rollout support sample-level chat template kwargs by @hjh0119 in #9623
• fix elastic code in transformers>=4.57.6 by @meichangsu1 in #9614
• [docs] update qwen3_5 docs by @Jintao-Huang in #9625
• Fix NPU MindSpeed 0.16 compatibility issues by @addsubmuldiv in #9624
• support dsa fused by @Jintao-Huang in #9616
• [BugFix] Fix IPC issue of async mode on AMD platform (ROCm) by @Treemann in #9627
• docs: fix GRPO pseudocode answer by @YingchaoX in #9630
• [feat] grpo multi-turn support openenv by @hjh0119 in #9628
• [docs] update contributing docs by @Jintao-Huang in #9634
• [docs] update deepseek_v4 vllm docs by @Jintao-Huang in #9597
• fix CI by @Jintao-Huang in #9636
• [bugfix] fix sp qwen3_5 by @Jintao-Huang in #9643
• update doc by @addsubmuldiv in #9650
• [model] support Unlimited_OCR by @z0o0ey in #9645
• [bugfix] fix Megatron Ray issue on AMD platform (ROCm) & update AMD support docs by @Treemann in #9653
• support minicpmv4_6 packing by @Jintao-Huang in #9654
• fix(dataset): skip empty rejected_messages instead of crashing by @he-yufeng in #9642
• [docs] update swift image by @Jintao-Huang in #9662
• [feat] support opd rl by @hjh0119 in #9641
• [model] support qwen3_tts by @Jintao-Huang in #9638
• Add support for Unlimited-OCR model with R-SWA by @z0o0ey in #9665
• align swift mm processor with vllm by @hjh0119 in #9669
• [bugfix] add missing offload_teacher_model param by @hjh0119 in #9674
• [bugfix] fix hf opsd & gkd use_vllm false by @hjh0119 in #9677
• [feat] support multi teacher by @hjh0119 in #9678
• [template] fix thinking list by @Jintao-Huang in #9679
• feat(megatron): expose optimizer_cuda_graph for the optimizer by @Nas01010101 in #9680
• fix ci by @Jintao-Huang in #9682

New Contributors

• @HaozheZhang6 made their first contribution in #9551
• @rivergod made their first contribution in #9577
• @yuchenwang3 made their first contribution in #9591
• @KadenZhang3321 made their first contribution in #9297
• @YingchaoX made their first contribution in #9630
• @z0o0ey made their first contribution in #9645
• @he-yufeng made their first contribution in #9642
• @Nas01010101 made their first contribution in #9680

Full Changelog: v4.3.0...v4.4.0