slime

23 atoms · 跨 4 天 · 首见 2026-05-23 · 最近 2026-06-21

三色: 🟦 fact 22 · 🟥 take 1 · stance ▲0/▼0/◆0

来源: X 23

时态: fresh:23

标签: 好观点:12 · 好信源:11 · 好思考:3

预测 (forecast) (1)

• 2026-06-21 · slime used for rollouts on non-CUDA hardware
  I think slime was only used for rollouts on non-CUDA hardware
  tags: 好观点 · → daily

事实 (fact) (22)

• 2026-05-24 · 使用SGLang原生推理架构
  slime Source Code Walkthrough: SGLang-Native Inference Architecture (Part II)
  tags: 好观点·好信源 · → daily
• 2026-05-24 · 推理端有三条通信路径
  slime’s inference side has three communication paths
  tags: 好观点·好信源 · → daily
• 2026-05-23 · async training loop launches next rollout before training completes via ray.get placement
  _rollout_manager.generate.remote(...) returns an ObjectRef immediately... after ray.get receives the current rollout data, slime immediately launches the next rollout with .remote() before starting training_
  tags: 好观点·好思考 · → daily
• 2026-06-02 · slime v0.3.0 发布
  slime v0.3.0: Built for the Agent Era
  tags: 好信源 · → daily
  value: qty=v0.3.0
• 2026-06-02 · slime 新增 agent-native infrastructure
  Agent-Native Infrastructure
  tags: 好信源 · → daily
• 2026-06-02 · slime 新增 YAML-based multi-server configuration for SGLang deployment
  slime expanded SGLang deployment with YAML-based multi-server configurations
  tags: 好信源 · → daily
• 2026-06-02 · slime 不再 offload fp32 gradients and bf16 parameters in integrated training-serving workloads
  slime no longer offloads fp32 gradients and bf16 parameters in integrated training-serving workloads
  tags: 好信源 · → daily
• 2026-06-02 · slime 新增 compact and subagent workflows 支持
  slime v0.3.0 adds support for compact and subagent workflows
  tags: 好信源 · → daily
• 2026-06-02 · slime 重构 PPO implementation 使 actor 和 critic 共享 GPU resources
  slime rebuilt its PPO implementation so that actor and critic always share GPU resources
  tags: 好信源 · → daily
• 2026-06-02 · slime 支持 independent Megatron configurations for actor and critic
  It also supports independent Megatron configurations for actor and critic
  tags: 好信源 · → daily
• 2026-06-02 · slime 将 fully async training promoted 为 first-class feature
  fully async training has been promoted from an experimental example to a first-class feature
  tags: 好信源 · → daily
• 2026-06-02 · slime 发布 Coding Agent RL example
  released a complete Coding Agent RL example
  tags: 好信源 · → daily
• 2026-05-24 · 路由功能外包给SGLang生态系统
  slime fully outsources routing to the SGLang ecosystem
  tags: 好观点 · → daily
• 2026-05-24 · 控制路径使用同步HTTP客户端
  _Control path: synchronous requests. _make_request uses sync requests_
  tags: 好观点 · → daily
• 2026-05-24 · 数据路径使用异步HTTP客户端
  _Data path: async post with retry. The data path uses post, an async function with retry
  tags: 好观点 · → daily
• 2026-05-24 · 连接池容量与信号量一致
  connection pool capacity is intentionally aligned with semaphore capacity
  tags: 好观点 · → daily
• 2026-05-23 · async training introduces off-policy data trade-off due to one-step-old weights
  _Data for rollout_id + 1 is generated before training for rollout_id finishes, so it uses the old weights from before the current update_
  tags: 好思考 · → daily
• 2026-05-23 · uses InfoActor probe to map logical GPU order to physical GPU order
  _slime solves this with an InfoActor probe: temporarily launch one @ray.remote(num_gpus=1) Actor on each bundle, let it call ray.get_gpu_ids() from inside the process, collect physical GPU locations_
  tags: 好思考 · → daily
• 2026-05-23 · inference architecture uses SGLang HTTP Server as independent service
  slime uses service-based inference by design: SGLang runs as an independent HTTP Server, and the training side calls it asynchronously through sgl-router
  tags: 好观点 · → daily
• 2026-05-23 · supports disaggregated and colocated deployment via one ray placement group offset slice
  slime unifies colocated and disaggregated deployment through: one placement group + offset slicing
  tags: 好观点 · → daily
• 2026-05-23 · SGLangEngine is a Ray Actor reserving 0.2 GPU, not the real inference engine
  SGLangEngine is a Ray Actor that reserves only about 0.2 GPU. It does not run inference itself
  tags: 好观点 · → daily
• 2026-05-23 · rollout_manager and SGLangEngine are Ray Actors, ServerGroup and RolloutServer are dataclass logical objects
  RolloutServer & ServerGroup are not processes. They are dataclasses living inside the RolloutManager Ray Actor process
  tags: 好观点 · → daily

时间轴 (近 20)

• 2026-06-21 · forecast · slime used for rollouts on non-CUDA hardware · → daily
• 2026-06-02 · fact · slime v0.3.0 发布 · → daily
• 2026-06-02 · fact · slime 新增 agent-native infrastructure · → daily
• 2026-06-02 · fact · slime 新增 YAML-based multi-server configuration for SGLang de · → daily
• 2026-06-02 · fact · slime 不再 offload fp32 gradients and bf16 parameters in integ · → daily
• 2026-06-02 · fact · slime 新增 compact and subagent workflows 支持 · → daily
• 2026-06-02 · fact · slime 重构 PPO implementation 使 actor 和 critic 共享 GPU resource · → daily
• 2026-06-02 · fact · slime 支持 independent Megatron configurations for actor and c · → daily
• 2026-06-02 · fact · slime 将 fully async training promoted 为 first-class feature · → daily
• 2026-06-02 · fact · slime 发布 Coding Agent RL example · → daily
• 2026-05-24 · fact · 使用SGLang原生推理架构 · → daily
• 2026-05-24 · fact · 推理端有三条通信路径 · → daily
• 2026-05-24 · fact · 路由功能外包给SGLang生态系统 · → daily
• 2026-05-24 · fact · 控制路径使用同步HTTP客户端 · → daily
• 2026-05-24 · fact · 数据路径使用异步HTTP客户端 · → daily
• 2026-05-24 · fact · 连接池容量与信号量一致 · → daily
• 2026-05-23 · fact · inference architecture uses SGLang HTTP Server as independen · → daily
• 2026-05-23 · fact · async training loop launches next rollout before training co · → daily
• 2026-05-23 · fact · supports disaggregated and colocated deployment via one ray · → daily
• 2026-05-23 · fact · SGLangEngine is a Ray Actor reserving 0.2 GPU, not the real · → daily

← 实体目录 · 系统日志