File tree Expand file tree Collapse file tree 4 files changed +13
-10
lines changed
tutorial/opencode_build_openclaw_agent Expand file tree Collapse file tree 4 files changed +13
-10
lines changed Original file line number Diff line number Diff line change @@ -46,6 +46,7 @@ class AgentJetJob:
4646 experiment_dir: Directory where experiment outputs will be saved.
4747 project_name: Name of the project for organizing experiments.
4848 experiment_name: Unique name for this specific experiment run.
49+ logging: "swanlab", "tensorboard", etc
4950 n_gpu: Number of GPUs to use per node for training.
5051 model: Path or identifier of the model to train.
5152 algorithm: Advantage estimator algorithm (e.g., 'gae', 'vtrace').
Original file line number Diff line number Diff line change @@ -90,17 +90,16 @@ AgentJet具备全分布式蜂群训练(Swarm Training)能力,
9090
9191## 蜂群 Agentic RL 框架核心优势
9292
93- | 特性 | 经典训练框架 | AgentJet 蜂群框架 |
93+ | 特性 | 经典LLM RL训练框架 | AgentJet 蜂群框架 |
9494| ------| -------------| ------------------|
95- | ** 多模型异构训练** | ❌ 所有智能体共享同一模型 | ✅ 支持多个不同规模模型同时训练 |
95+ | ** 多模型异构训练** | 所有智能体共享同一可训练模型 | ✅ 支持多个不同规模模型同时训练 |
9696| ** 训练推理解耦** | ❌ 采样与训练紧耦合 | ✅ Server 训练,Client 采样,完全解耦 |
9797| ** 运行环境限制** | ❌ 受训练服务器环境限制 | ✅ Client 可在任意设备运行(笔记本/服务器) |
98- | ** 动态节点管理** | ❌ 无法动态添加/移除节点 | ✅ 训练中随时添加/移除 Client 节点 |
98+ | ** 动态节点管理** | ❌ 不支持边训练,边Debug | ✅ 训练中随时添加/移除 Client 节点 |
9999| ** 调试迭代速度** | ❌ 修改代码需重启训练(每次15分钟+) | ✅ 仅重启 Client(秒级),无需重载模型 |
100- | ** 容错能力** | ❌ 外部依赖故障导致训练中断 | ✅ Client 崩溃不影响训练,自动恢复 |
101- | ** 多任务混合训练** | ❌ 所有任务共享同一运行环境 | ✅ 不同 Client 运行不同任务环境 |
102- | ** Agent 框架兼容** | ❌ 需适配特定训练框架 | ✅ 兼容任何 OpenAI API 协议框架 |
103- | ** 本地开发体验** | ❌ 必须在训练服务器上调试 | ✅ 本地 IDE 调试,连接远程训练 |
100+ | ** 容错能力** | ❌ 外部故障导致训练中断会丢失进度 | ✅ Client 崩溃不影响训练,自动恢复 |
101+ | ** 多任务混合训练** | 所有任务共享同一运行环境 | ✅ 不同 Client 运行不同任务环境 |
102+ | ** 本地开发体验** | 在训练服务器上调试 | ✅ 本地 IDE 调试,连接远程训练 |
104103
105104## 灵活的蜂群训练模式
106105
Original file line number Diff line number Diff line change @@ -84,11 +84,13 @@ This starts the training proxy on `http://localhost:8090`.
8484
8585### Step 3: Configure OpenClaw to Use Training Endpoint
8686
87- OpenClaw needs to connect to the fake vLLM endpoint. Configure it to use ` http://localhost:8090 ` as the LLM backend.
87+ OpenClaw needs to connect to the fake vLLM endpoint.
88+
89+ Configure it to use ` http://localhost:8090 ` as the LLM backend.
8890
8991### Step 4: Send Training Requests
9092
91- Option A - Manual testing via OpenClaw CLI :
93+ Option A - Manual testing via OpenClaw Web / Cli :
9294
9395``` bash
9496openclaw agent --message " What are your thoughts on Paris?" --thinking high
Original file line number Diff line number Diff line change 4242 project_name = "openclaw-extraversion" ,
4343 experiment_name = "extraversion_training" ,
4444 n_gpu = 8 ,
45- model = '/mnt/data_cpfs/model_cache/modelscope/hub/Qwen/Qwen/Qwen2.5-3B -Instruct' ,
45+ model = '/mnt/data_cpfs/model_cache/modelscope/hub/Qwen/Qwen/Qwen2___5-7B -Instruct' ,
4646 batch_size = 32 ,
47+ logging = "swanlab" ,
4748 num_repeat = NUM_REPEAT ,
4849 max_prompt_length = 16000 , # at least 16000
4950 max_response_length = 8000 ,
You can’t perform that action at this time.
0 commit comments