v0.4.0

Highlights

• Initial DeepSeek V4 support, covering Flash FSDP2 + EP training and DeepSeek V4 tool-call parsing and cleanup in #190 and #218
• Expand Qwen3.5 training with padding-free / packed-sequence support and Qwen3.5 MoE GatedDeltaNet sequence-parallel support in #186 and #222
• Add Gemma 4 multimodal training support #199
• Strengthen LoRA training with rsLoRA for Multi-LoRA, FSDP2 support for Multi-LoRA SFT, and Expert Parallelism LoRA SFT examples for DeepSeek V4 and Qwen3.5 MoE in #187, #155, and #198
• Improve NPU acceleration and stability with fused operators, Qwen3.5 FLA patches, Group MatMul EP scoping, and sequence-parallel compatibility fixes in #194, #204, #205, #206, and #208

New Features

• Add padding-free and packed-sequence support for Qwen3.5 by @meichangsu1 in #186
• Add rsLoRA support to Multi-LoRA by @xichengpro in #187
• Add FSDP2 support for Multi-LoRA SFT by @kevssim in #155
• Add DeepSeek V4 Flash FSDP2 + EP training support by @meichangsu1 in #190
• Add NPU fused operators: RMSNorm, RoPE, SwiGLU, and SDPA by @ys2025-AI in #194
• Add multi-turn rollout support by @tastelikefeet in #193
• Add support for client-specified checkpoint saving paths by @vx120 in #196
• Add LoRA SFT support for Expert Parallelism, with DeepSeek V4 and Qwen3.5 MoE examples by @kevssim in #198
• Add Qwen3.5 NPU FLA and fused-operator patches by @ys2025-AI in #204
• Add LoRA capacity query support by @kevssim in #201
• Optimize Native FSDP memory_efficient_init weight loading for multi-node EP/FSDP jobs and add multi-node scripts by @meichangsu1 in #207
• Add Gemma 4 support by @EvineR666 in #199
• Add DeepSeek V4 tool-call parsing and cleanup support by @meichangsu1 in #218
• Add Gemma 4 12B cookbook by @EvineR666 in #219
• Add automatic device detection by @vx120 in #220
• Add Qwen3.5 MoE GatedDeltaNet sequence-parallel support by @meichangsu1 in #222
• Refactor server configuration and observability by @Yunnglin in #210

Bug Fixes

• Fix cache reset behavior for multimodal models by @hjh0119 in #189
• Fix Qwen3.5 GatedDeltaNet padding-free compatibility and create_causal_mask compatibility after cache_positions removal in transformers >5.3.0 by @meichangsu1 in #202
• Fix transformers 5.9 AttentionMask wrapper compatibility in sequence parallel by @ys2025-AI in #206
• Fix SP path overriding the NPU-patched chunk_gated_delta_rule by @ys2025-AI in #208
• Fix NPU Group MatMul patch scope so it only applies in EP scenarios by @0hujun in #205
• Fix adapter saving to use the MultiLora state dict by @meichangsu1 in #215

更新内容

亮点功能

• 首发支持 DeepSeek V4，覆盖 Flash FSDP2 + EP 训练，以及 DeepSeek V4 tool call 解析与清理 in #190 and #218
• 扩展 Qwen3.5 训练能力，新增 padding-free / packed-sequence 支持和 Qwen3.5 MoE GatedDeltaNet sequence parallel 支持 in #186 and #222
• 新增 Gemma 4 多模态训练支持 in #199
• 增强 LoRA 训练能力，覆盖 Multi-LoRA 的 rsLoRA、Multi-LoRA SFT 的 FSDP2 支持，以及 DeepSeek V4 / Qwen3.5 MoE 的 EP LoRA SFT 示例 in #187, #155, and #198
• 增强 NPU 加速与稳定性，覆盖融合算子、Qwen3.5 FLA patch、Group MatMul EP 以及 sequence-parallel 兼容性修复 in #194, #204, #205, #206, and #208

新特性

• 支持 Qwen3.5 padding-free / packed-sequence 训练 by @meichangsu1 in #186
• Multi-LoRA 支持 rsLoRA by @xichengpro in #187
• Multi-LoRA SFT 支持 FSDP2 by @kevssim in #155
• 支持 DeepSeek V4 Flash FSDP2 + EP 训练 by @meichangsu1 in #190
• 新增 NPU 融合算子：RMSNorm、RoPE、SwiGLU、SDPA by @ys2025-AI in #194
• 支持 multi-turn rollout by @tastelikefeet in #193
• 支持客户端指定服务端路径保存 checkpoint by @vx120 in #196
• EP 支持 LoRA SFT，并新增 DeepSeek V4 和 Qwen3.5 MoE 示例 by @kevssim in #198
• 新增 Qwen3.5 NPU FLA 与融合算子补丁 by @ys2025-AI in #204
• 支持查询 LoRA capacity 信息 by @kevssim in #201
• 优化 Native FSDP memory_efficient_init 多节点 EP/FSDP 权重加载，并新增多节点脚本 by @meichangsu1 in #207
• 新增 Gemma 4 支持 by @EvineR666 in #199
• 新增 DeepSeek V4 tool call 解析与清理支持 by @meichangsu1 in #218
• 新增 Gemma 4 12B cookbook by @EvineR666 in #219
• 新增自动显卡设备检测 by @vx120 in #220
• 支持 Qwen3.5 MoE GatedDeltaNet sequence parallel by @meichangsu1 in #222
• 服务端配置与可观测性重构 by @Yunnglin in #210

Bug 修复

• 修复多模态模型 cache reset 问题 by @hjh0119 in #189
• 修复 Qwen3.5 GatedDeltaNet padding-free 训练兼容性，并兼容 transformers >5.3.0 中 cache_positions 移除后的 create_causal_mask 逻辑 by @meichangsu1 in #202
• 修复 sequence parallel 中 transformers 5.9 AttentionMask wrapper 兼容问题 by @ys2025-AI in #206
• 修复 SP 路径覆盖 NPU patch 后的 chunk_gated_delta_rule 问题 by @ys2025-AI in #208
• 修复 NPU Group MatMul patch 作用范围，限定仅在 EP 场景启用 by @0hujun in #205
• 修复保存 adapter 时未使用 MultiLora state dict 的问题 by @meichangsu1 in #215

New Contributors

• @tpx818 made their first contribution in #65
• @wangxingjun778 made their first contribution in #68
• @hzher made their first contribution in #92
• @xichengpro made their first contribution in #123
• @vx120 made their first contribution in #118
• @0hujun made their first contribution in #183
• @a550580874 made their first contribution in #176
• @ys2025-AI made their first contribution in #194
• @EvineR666 made their first contribution in #199

Full Changelog: https://github.com/modelscope/twinkle/commits/v0.4.0

v0.3.0

中文版本

新特性

1. 全面支持padding_free参数，可用于sft、dpo、grpo等各类训练中，在InputProcessor构造时传入padding_free=True即可生效
2. 支持resume-from-checkpoint，参考这里

Bug修复

1. 更新了lora dtype和模型dtype不同导致的训练问题
2. 修复了npu gemm算子的支持
3. 修复npu下fsdp生效时megatron gather报错的问题

English Version

New Features

1. Full support for the padding_free parameter, which can be used in various training types such as SFT, DPO, GRPO, etc. It takes effect by passing padding_free=True when constructing InputProcessor.
2. Support for resume-from-checkpoint. Refer to here.

Bug Fixes

1. Fixed a training issue caused by mismatched LoRA dtype and model dtype.
2. Fixed support for the NPU GEMM operator.
3. Fixed an error where Megatron gather failed when FSDP was enabled on NPU.

What's Changed

• Update docker file by @tastelikefeet in #180
• fix: model dtype is not same as lora dtype in FSDP train by @0hujun in #183
• fix: when setting fsdp size unuse megatron for gather in npu by @0hujun in #185
• npu gemm patch by @a550580874 in #176
• Support dpo/grpo/gkd/sft padding_free by @tastelikefeet in #181
• [feat] Resume from ckpt by @kevssim in #135

New Contributors

• @0hujun made their first contribution in #183
• @a550580874 made their first contribution in #176

Full Changelog: v0.2.1...v0.3.0

v0.2.1

中文版本

新功能

1. 支持了Qwen/Qwen3.6-27B的魔搭官方服务，详情查看：https://www.modelscope.cn/organization/twinkle-kit

Bug修复

1. 修复了expert权重同步错误的问题
2. 修复了多lora场景下GRPO MoE训练崩塌的问题
3. 修复了对多模态输入的序列切分问题
4. 修复了pp > 1 和tp>1时服务器不正常的问题
5. 修复了多处remote_function不正确的问题
6. 修复了服务器训练模型上传和模型训练共用pipeline导致阻塞的问题
7. 修复了采样器模块的一些bug

English Version

New Features

1. Added support for the official ModelScope service on Qwen/Qwen3.6-27B. For details, see: https://www.modelscope.cn/organization/twinkle-kit

Bug Fixes

1. Fixed an issue with incorrect expert weight synchronization.
2. Fixed a training collapse issue with GRPO MoE in multi-LoRA scenarios.
3. Fixed a sequence splitting issue for multimodal inputs.
4. Fixed abnormal server behavior when pp > 1 and tp > 1.
5. Fixed multiple incorrect remote_function implementations.
6. Fixed a blocking issue caused by the model upload and model training pipelines sharing the same pipeline on the server side.
7. Fixed several bugs in modules such as the Sampler.

What's Changed

• add base_layer suffix for expert weights by @hjh0119 in #159
• update cookbook and doc 0415 by @Yunnglin in #157
• Docs support Q3.6 by @tastelikefeet in #158
• Fix multi lora device by @tastelikefeet in #160
• Fix MoE multi-lora training by @tastelikefeet in #161
• Fix model id and upload to hub by @Yunnglin in #162
• Add notebooks by @tastelikefeet in #164
• Npu adapt megatron by @addsubmuldiv in #153
• Fix save by @tastelikefeet in #165
• A small refactor by @tastelikefeet in #166
• A small refactor, move 4d mask to processor by @tastelikefeet in #167
• Fix some potential bugs by @tastelikefeet in #168
• Fix some bugs by @tastelikefeet in #169
• fix mm tokentypeids splitting by @tastelikefeet in #170
• Fix model pp > 1 and tp > 1 errors by @Yunnglin in #171
• Fix moe weight sync by @tastelikefeet in #172
• update notebooks by @Yunnglin in #174
• Modify remote_function decorators in multi_lora_transformers by @xichengpro in #173
• support cp ,fix qwen3.5 gdn sp by @meichangsu1 in #138
• support qwen3.6 grpo & in-place add lora by @hjh0119 in #163
• Fix multi lora by @tastelikefeet in #177
• support q3.6-27b by @tastelikefeet in #178
• Fix sampler and grpo by @Yunnglin in #179

Full Changelog: v0.2.0...v0.2.1

v0.2.0

中文

新特性

1. 重构了服务部分，目前的多租户服务支持tinker/twinkle双client语法规则。
2. 支持GKD和On-policy蒸馏，请查看cookbook。
3. megatron的底层替换为mcore_bridge库，并支持了对应的多模态训练。
4. 支持了DPO算法，请查看cookbook。
5. 支持了Qwen3.5系列的多模态任务的训练。
6. 新增了服务端可用的Dockerfile。

Bug修复

1. 0.2.0 bug修复较多，请查看如下的修复列表。

English

New Features

1. Refactored the service layer; the multi-tenant service now supports both tinker/twinkle dual client syntax rules.
2. Added support for GKD and On-policy distillation — see cookbook.
3. Replaced the underlying Megatron backend with the mcore_bridge library, with support for corresponding multimodal training.
4. Added support for the DPO algorithm — see cookbook.
5. Added support for multimodal task training on the Qwen3.5 series.
6. Added a server-side Dockerfile.

Bug Fixes

1. A significant number of bugs have been fixed in 0.2.0 — please refer to the fix list below.

What's Changed

• Fix tinker loss device mismatch by @addsubmuldiv in #115
• Refact server by @Yunnglin in #111
• [Fix] EP+FSDP checkpoint save for MoE expert parameters by @kevssim in #116
• Support GKD and on-policy distillation by @tastelikefeet in #112
• Support patcher on samplers by @tastelikefeet in #119
• fix vllmsampler client by @tastelikefeet in #122
• [Fix] Prevent client from importing ray via twinkle.server.common.serialize by @xichengpro in #123
• update Qwen3.5 grpo demo by @hjh0119 in #124
• Fix mm server by @tastelikefeet in #125
• fix tp get logps by @hjh0119 in #126
• fix serve_multiplexed_model_id and mm data process by @Yunnglin in #120
• [feat] fsdp2 memory_efficient_init by @kevssim in #117
• fix megatron weights sync by @hjh0119 in #128
• Support DPO by @tastelikefeet in #130
• Fix npu qwen3moe grpo by @vx120 in #118
• fix dpo with lazy_dataset by @tastelikefeet in #136
• support transformers multi-modal grpo by @hjh0119 in #131
• fix import by @kevssim in #137
• Refactor megatron to mcore_bridge by @tastelikefeet in #134
• Fix bugs by @tastelikefeet in #139
• Change online model to qwen3.5-27b by @tastelikefeet in #140
• Merge to main by @tastelikefeet in #141
• fix megatron multi-lora converter by @hjh0119 in #144
• Remerge release/0.2 to main by @tastelikefeet in #146
• support rl vit lora with vLLM by @hjh0119 in #147
• Add server metrics monitor and DPO client by @Yunnglin in #132
• Fix multi lora saving by @tastelikefeet in #148
• fix transformers model loading by @tastelikefeet in #150
• fix short math grpo cookbook by @Yunnglin in #149
• fix docker file by @Yunnglin in #151
• fix cookbook by @tastelikefeet in #152
• fix tensor collect by @Yunnglin in #154
• fix multi lora training by @tastelikefeet in #156

New Contributors

• @xichengpro made their first contribution in #123
• @vx120 made their first contribution in #118

Full Changelog: v0.1.3...v0.2.0

v0.1.3

中文版本

新特性

1. 增加了client模式的便捷安装脚本，并提升了文档描述
2. 支持transformers分支的ep+fsdp分片

Bug修复

1. 修复加载本地数据集失败的问题
2. 修复服务化启动时http_options错误传递到模型的问题

English Version

New features

1. Add a shell installation script to support the client mode, and improve the description of documentation
2. Support ep+fsdp sharding of transformers

BugFix

1. Fix a bug that causes an error on local dataset loading
2. Fix an error that the http_options argument was mis-transfered to the model in the server mode

What's Changed

• Fix loading local datasets by @tastelikefeet in #108
• [fix] http_options leaking to model init & NPU tensor serialization failure over HTTP by @kevssim in #109
• Fix docs and add new start scripts by @tastelikefeet in #113
• [feat]support ep_fsdp by @kevssim in #71

Full Changelog: v0.1.2...v0.1.3

v0.1.2

中文

新特性

1. 支持Qwen3.5系列的transformers模型多模态训练，包含图片和视频
2. 支持数据集预处理的batched=True操作，提升速度

Bug修复

1. 修复NPU下权重同步卡死的问题

English

New Features

1. Support multi-modal training of Qwen3.5 transformers framework, containing images and videos
2. Support batched=True when preprocess datasets

BugFix

1. Fix the hang problem of NPU weight synchronization

What's Changed

• Update cookbok to qwen35 by @tastelikefeet in #98
• Support Qwen3.5 mm by @tastelikefeet in #100
• Support batched preprocessing by @tastelikefeet in #101
• fix video mm by @tastelikefeet in #105
• Fix GRPO weight-sync hangs and HCCL resource exhaustion on NPU by @addsubmuldiv in #102
• add new cookbook with qwen3.5 by @tastelikefeet in #106
• fix cookbook by @tastelikefeet in #107

Full Changelog: v0.1.1...v0.1.2

v0.1.1

Twinkle 0.1.1 version Release

中文

• 支持Qwen3.5-2B~Qwen3.5-9B等Dense模型

English

• Support model series of Qwen3.5-2B~Qwen3.5-9B

Full Changelog: v0.1...v0.11

v0.1

中文

Twinkle框架的0.1版本发布！

新功能

1. 🎉完整的数据集、DataLoader、Loss、Transformers和Megatron模型、Advantage、Sampler等组件的支持
2. 🎉支持PT、SFT、RL等多种训练Stage，并支持单卡、多机多卡、Ray、Client-Server等多种训练模式
3. 🎉支持了首版的多租户复用训练，并完整开源了server端实现。使用ray serve实现了多副本可扩缩容部署，并支持粘滞路由
4. 🎉在魔搭官方网站上，提供了在线服务，用户可以使用该服务免费训练Qwen/Qwen3-30B-A3B-Instruct-2507，并推送模型到ModelHub上

English

Twinkle Framework Version 0.1 Released!

New Features

1. 🎉 Full support for components including Dataset, DataLoader, Loss, Transformers and Megatron models, Advantage, Sampler, and more
2. 🎉Support for multiple training stages such as PT, SFT, and RL, with various training modes including single-GPU, multi-node multi-GPU, Ray, and Client-Server
3. 🎉 First version of multi-tenant shared training is now supported, with the server-side implementation fully open-sourced. Multi-replica scalable deployment is implemented using Ray Serve, with support for sticky routing
4. 🎉 An online service is now available on the ModelScope official website, where users can train Qwen/Qwen3-30B-A3B-Instruct-2507 for free and push models to ModelHub

What's Changed

• Squash to main by @tastelikefeet in #46
• rename cmb by @tpx818 in #65
• docs: update README and remove ulysses_size from ep_fsdp_qwen3_moe.py by @meichangsu1 in #64
• add contrbutors by @yingdachen in #66
• fix lora fetch by @tastelikefeet in #67
• Update documentation links in README.md by @wangxingjun778 in #68
• Fix router by @tastelikefeet in #69
• Fix doc links and add tests by @tastelikefeet in #70
• Refactor code by @tastelikefeet in #72
• Fix compat tinker and update doc by @Yunnglin in #73
• [compat] gpt_bridge compat transformers_5 by @Jintao-Huang in #75
• Fix server state adapter limit by @Yunnglin in #74
• Fix some bugs by @tastelikefeet in #77
• [model] support Qwen3.5 series models by @hjh0119 in #76
• fix single gpu bug by @tastelikefeet in #78
• [bugfix] fix dense model get layer spec by @hjh0119 in #80
• fix grad norm bug by @tastelikefeet in #81
• Update readme by @yingdachen in #83
• Add custom route for sticky session by @Yunnglin in #82
• [bugfix] fix 4d attention mask device by @hjh0119 in #85
• add more comment for node resouces by @tastelikefeet in #79
• Update doc and fix bugs by @tastelikefeet in #84
• Fix logps by @tastelikefeet in #86
• recover cp sequence before loss by @hjh0119 in #88
• [bugfix] fix logps with PP by @hjh0119 in #89
• Fix megatron loss by @tastelikefeet in #90
• Dev feature by @hzher in #92
• Fix proxy by @Yunnglin in #87
• fix TEGroupedLinear by @tastelikefeet in #94
• [bugfix] fix grpo loss by @hjh0119 in #93
• fix numpy version by @tastelikefeet in #95
• [bugfix] fix contiguous by @hjh0119 in #96
• Add a sample script by @tastelikefeet in #97

New Contributors

• @tastelikefeet made their first contribution in #46
• @tpx818 made their first contribution in #65
• @wangxingjun778 made their first contribution in #68
• @hzher made their first contribution in #92

Full Changelog: https://github.com/modelscope/twinkle/commits/v0.1