update docs

Author: binary-husky

README.md

@@ -28,7 +28,7 @@
 
 ## ✈️ Fast Introduction
 
-### Classic Mode
+### **1 - Classic Mode**
 
 Let's begin with the simplest example: a math agent with a tool call. This is a simple & centralized training example.
 
@@ -37,11 +37,12 @@ Let's begin with the simplest example: a math agent with a tool call. This is a
     ```python
     ajet --conf ./tutorial/example_math_agent/math_agent.yaml --backbone='verl'
     ```
+
 <div align="center">
-<img width="640" alt="image" src="https://serve.gptacademic.cn/publish/shared/Image/classic+swarm+revise.jpg"/>
+<img width="600" alt="image" src="https://serve.gptacademic.cn/publish/shared/Image/new_classic.png"/>
 </div>
 
-### Swarm Mode
+### **2 - Swarm Mode**
 
 Let's begin with the simplest AgentJet Swarm example: also a math agent. In this case, you can use any GPU-less laptop to train the model remotely.
 
@@ -50,6 +51,17 @@ Let's begin with the simplest AgentJet Swarm example: also a math agent. In this
     ```python
     AJET_SWARM_URL="http://swarm-server-ip:10086" python ./tutorial/example_math_swarm/math.py
     ```
+
+
+<div align="center">
+<img width="600" alt="image" src="https://serve.gptacademic.cn/publish/shared/Image/new_swarm.png"/>
+</div>
+
+<div align="center">
+<img width="600" alt="image" src="https://serve.gptacademic.cn/publish/shared/Image/swarming.gif"/>
+</div>
+
+
 <div align="center">
 <img width="600" alt="image" src="https://github.com/user-attachments/assets/41ed1e71-8b18-4c4c-b5e2-833399317337"/>
 </div>

docs/en/workflow.md docs/en/classic_workflow.mddocs/en/workflow.md renamed to docs/en/classic_workflow.md

@@ -1,6 +1,6 @@
-# Trainable Workflow
+# Classic Workflow
 
-This tutorial introduces how to define a trainable workflow.
+This tutorial introduces how to define a trainable workflow in **classic mode** — single-machine training launched with the `ajet` CLI, where rollout and training share the same GPU cluster. For distributed training where rollout workers run on separate (potentially GPU-less) machines, see [Swarm Workflow](../swarm_workflow/).
 
 !!! info ""
     AgentJet provides two **convenient** and **mutually compatible** ways to wrap your Workflow:
@@ -241,16 +241,10 @@ Here's a complete example with multiple agent roles (Werewolves game):
     - You can flexibly switch training targets by modifying `trainable_targets`
 
 
-## Swarm
-
-Wrapping and training your agent on a machine without GPU.
-
-Working in progress and coming soon.
-
-
 ## Next Steps
 
 <div class="card-grid">
+<a href="../swarm_workflow/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/lucide:network.svg" class="card-icon card-icon-general" alt=""><h3>Swarm Workflow</h3></div><p class="card-desc">Distributed training with rollout workers on separate machines.</p></a>
 <a href="../data_pipeline/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:database.svg" class="card-icon card-icon-data" alt=""><h3>Data Pipeline</h3></div><p class="card-desc">Configure data loading from files, HuggingFace, or environments.</p></a>
 <a href="../task_judger/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:check-decagram.svg" class="card-icon card-icon-general" alt=""><h3>Task Judger</h3></div><p class="card-desc">Set up reward functions to evaluate agent performance.</p></a>
 </div>

docs/en/configuration.md

@@ -2,7 +2,7 @@
 
 AgentJet uses YAML files to configure every aspect of a training run.
 This page is a **lookup reference** for every configuration key.
-For conceptual introductions, see [Workflow](../workflow/), [Data Pipeline](../data_pipeline/), and [Task Judger](../task_judger/).
+For conceptual introductions, see [Classic Workflow](../classic_workflow/), [Swarm Workflow](../swarm_workflow/), [Data Pipeline](../data_pipeline/), and [Task Judger](../task_judger/).
 
 <br/>
 
@@ -1273,7 +1273,8 @@ ajet --conf my_config.yaml --backbone=debug
 ## Next Steps
 
 <div class="card-grid">
-<a href="../workflow/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/lucide:workflow.svg" class="card-icon card-icon-general" alt=""><h3>Workflow</h3></div><p class="card-desc">How to define your trainable agent workflow.</p></a>
+<a href="../classic_workflow/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/lucide:workflow.svg" class="card-icon card-icon-general" alt=""><h3>Classic Workflow</h3></div><p class="card-desc">How to define your trainable agent workflow.</p></a>
+<a href="../swarm_workflow/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/lucide:network.svg" class="card-icon card-icon-general" alt=""><h3>Swarm Workflow</h3></div><p class="card-desc">Distributed training with rollout on separate machines.</p></a>
 <a href="../data_pipeline/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/lucide:database.svg" class="card-icon card-icon-general" alt=""><h3>Data Pipeline</h3></div><p class="card-desc">Task readers and data formats in detail.</p></a>
 <a href="../task_judger/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/lucide:scale.svg" class="card-icon card-icon-general" alt=""><h3>Task Judger</h3></div><p class="card-desc">Built-in and custom reward functions.</p></a>
 <a href="../example_math_agent/" class="feature-card"><div class="card-header"><img src="https://api.iconify.design/mdi:calculator-variant.svg" class="card-icon card-icon-math" alt=""><h3>Math Agent Example</h3></div><p class="card-desc">See all configuration applied in a real training run.</p></a>

docs/en/example_train_multi_model.md

@@ -202,7 +202,8 @@ Compared to traditional single-model shared parameter training, non-shared param
 ### Cross References
 
 - **[AgentJet Swarm Training Mode](../swarm.md)**: Deep dive into AgentJet's swarm architecture design philosophy and core advantages
-- **[Trainable Workflow](../workflow.md)**: Learn how to define multi-agent workflows in AgentJet
+- **[Classic Workflow](../classic_workflow.md)**: Learn how to define multi-agent workflows in AgentJet
+- **[Swarm Workflow](../swarm_workflow.md)**: Distributed training with rollout on separate machines
 - **[Task Judger](../task_judger.md)**: Understand reward function design principles and customization methods
 - **[Math Agent Example](../example_math_agent.md)**: Learn the basics of single-agent training
 

docs/en/example_train_multi_model.zh.md

@@ -203,7 +203,8 @@ sequenceDiagram
 ### 交叉引用
 
 - **[AgentJet Swarm 训练模式](../swarm.md)**：深入了解 AgentJet 蜂群架构的设计理念和核心优势
-- **[可训练工作流](../workflow.md)**：学习如何在 AgentJet 中定义多智能体工作流
+- **[经典工作流](../classic_workflow.md)**：学习如何在 AgentJet 中定义多智能体工作流
+- **[蜂群工作流](../swarm_workflow.md)**：分布式训练，rollout 在独立机器上运行
 - **[任务评判器](../task_judger.md)**：了解奖励函数的设计原理和自定义方法
 - **[数学 Agent 示例](../example_math_agent.md)**：学习单智能体训练的基础示例
 

docs/en/swarm_intro_blog_en.md

@@ -245,7 +245,7 @@ Explore our rich example library to quickly start your journey:
     - 👩🏻‍⚕️ [**Learn to ask questions like a doctor**](https://modelscope.github.io/AgentJet/en/example_learning_to_ask)
     - 🎴 [**Write and solve a countdown game using AgentScope**](https://modelscope.github.io/AgentJet/en/example_countdown)
     - 🚶 [**Solve the Frozen Lake puzzle using AgentJet**](https://modelscope.github.io/AgentJet/en/example_frozenlake)
-- **Timeline Automatic Merging Capability**: Supports [multi-agent workflows](https://modelscope.github.io/AgentJet/en/workflow/) and adopts context merging technology to accelerate training by **1.5x to 10x** in multi-turn (or multi-agent) conversation scenarios. (Similar to the "tree structure" processing capability mentioned in the minimax forge technical report.)
+- **Timeline Automatic Merging Capability**: Supports [multi-agent workflows](https://modelscope.github.io/AgentJet/en/classic_workflow/) and adopts context merging technology to accelerate training by **1.5x to 10x** in multi-turn (or multi-agent) conversation scenarios. (Similar to the "tree structure" processing capability mentioned in the minimax forge technical report.)
 - **Reliable and Reproducible**: We continuously track the framework's performance on multiple [different tasks + major Git versions + different training backends](https://benchmark.agentjet.top/) (data continuously aggregated), what you see is what you get, hidden bugs are discovered in seconds.
 - **Token Consistency Automatic Alert & Repair**: By default, AgentJet automatically performs Re-tokenization drift repair based on the Token ID returned by the vLLM engine.
 - **Multi-Training Backend Support**: Supports multiple training backends including VERL, and is working on supporting other training backends like TRL.

docs/en/swarm_intro_blog_zh.md

@@ -259,7 +259,7 @@ AgentJet 在工程质量上投入了大量精力，确保每一次训练结果
     - 👩🏻‍⚕️ [**学习像医生一样提问**](https://modelscope.github.io/AgentJet/en/example_learning_to_ask)
     - 🎴 [**使用 AgentScope 编写倒计时游戏并求解**](https://modelscope.github.io/AgentJet/en/example_countdown)
     - 🚶 [**使用 AgentJet 解决冰湖行走谜题**](https://modelscope.github.io/AgentJet/en/example_frozenlake)
-- **时间线自动合并能力**: 支持 [多智能体工作流](https://modelscope.github.io/AgentJet/en/workflow/) 并采用上下文合并技术，在多轮（或多智能体）对话场景中将训练加速 **1.5x 到 10x**。(类似于minimax forge技术报告中提到的“树形结构”处理能力。)
+- **时间线自动合并能力**: 支持 [多智能体工作流](https://modelscope.github.io/AgentJet/en/classic_workflow/) 并采用上下文合并技术，在多轮（或多智能体）对话场景中将训练加速 **1.5x 到 10x**。(类似于minimax forge技术报告中提到的“树形结构”处理能力。)
 - **可靠可复现**：我们持续跟踪框架在多项 [不同任务 + 主要 Git 版本 + 不同训练后端](https://benchmark.agentjet.top/) 上的性能（数据持续汇总中），所见即所得，隐形Bug秒发现。
 - **Token一致性自动告警&修复**：AgentJet默认情况下会自动根据 vLLM 引擎返回的 Token ID 进行 Re-tokenization 漂移修复。
 - **多训练后端支持**：支持包括 VERL 在内的多个训练后端，正着手支持 TRL 等其他训练后端。