👋Welcome ms-swift v4

[Jintao-Huang]

ms-swift v4将会对整体框架进行重构，这会引入不少break change。以下为会导致break change的重构内容：

1. 目录结构重构与依赖关系优化：swift.llm目录将拆分为swift.template, swift.dataset, swift.model, swift.pipelines。
2. model_type与template接耦，与transformers model_type对齐，将去除qwen3_thinking, qwen3_no_thinking, qwen3_moe_thinking等额外的model_type。
3. megatron训练 training loop重写，支持RAY。使用megatron-core而丢弃使用megatron-lm依赖。
4. transformers v5.0新特性引入。

如果有更多重构/新特性支持的建议，可以在下方留言，该文档会持续更新，谢谢！

ms-swift v4 will undergo a comprehensive framework refactoring, which will introduce several breaking changes. The following are refactoring items that will cause breaking changes:

1. Directory structure refactoring and dependency optimization: The swift.llm directory will be split into swift.template, swift.dataset, swift.model, and swift.pipelines.
2. Decoupling model_type from template, aligning with transformers model_type. Additional model_types such as qwen3_thinking, qwen3_no_thinking, qwen3_moe_thinking, etc. will be removed.
3. Megatron training loop rewrite to support RAY. Will use megatron-core and drop the megatron-lm dependency.
4. Introduction of transformers v5.0 new features.

If you have more suggestions for refactoring/new feature support, please leave a comment below. This document will be continuously updated. Thank you!

[edc3000]

加入对remote的LLM-as-judge奖励的支持，不用同步等待奖励计算时间，减少GPU气泡

[chuxiliyixiaosa]

pt推理时，读取数据与模型推理是串行的，当推理图像、视频等数据时，会导致推理效率低下。

[hjh0119]

@edc3000 maybe async reward can help #7252

[Jintao-Huang]

pt推理时，读取数据与模型推理是串行的，当推理图像、视频等数据时，会导致推理效率低下。

是的 会优化一下

[aJupyter]

与其他框架RL速度的对比

[LiXinYuECNU]

使用swift推理，当某个样本数据加载报错会直接中断其他的推理

[zyjcsf]

希望可以增加对Omni模型talker部分的微调训练支持哈

[xiyu-Luoy]

hope support vllm-omni

[651961]

希望支持一下MiMo-V2-Flash的MOPD（Multi-Teacher On-Policy Distillation），on policy distillation的方法很有潜力，希望能做的更简单易用。

[JoshonSmith]

希望可以增加对Omni模型talker部分的微调训练

[howitry]

希望可以优化流式加载数据集场景下的训练速度，目前流式加载数据集的速度不够理想：（1）不支持dataloader_num_workers>1；（2）dataloader只会在rank==0处理数据并分发到其他rank，这样训练速度会随着机器的增加而变慢，非常不利于流式加载场景下的多机多卡训练。

[hpsun1109]

希望支持一下扩散模型训练

[hpsun1109]

希望支持Nemotron-3-Nano-30B-A3B等mamba架构模型

[inzamam-iqbal]

Supporting Pluginization with call backs for megatron training will be very useful. Currently when megatron the training is not very customizable unless we make code changes to the ms-swift package it self. Thank you.