Skip to content

Consolidated release: gateway sharding, hub mode, math training, H-cluster practice - #4

Closed
AltmanD wants to merge 1 commit into
datamllab:mainfrom
AltmanD:main
Closed

AltmanD wants to merge 1 commit into
datamllab:mainfrom
AltmanD:main

Conversation

@AltmanD

@AltmanD AltmanD commented Nov 12, 2026

Copy link
Copy Markdown
Collaborator
  • 网关弹性扩展:静态路由 + 弹性自动扩容(多机优先)+ 自动缩容
  • 新增 Hub 部署模式:节点经 HTTP 主动拉取任务,适配 gpu 机器配备防火墙的集群
  • 数学任务算法验证:GRPO / DAPO / NPR / OPD / PRM 蒸馏,统一接入共享 Slime 训练循环
  • Controller 与 sandbox 优化:预检、端口分配、启动队列、清理准入
  • 配置工具链优化:选项 CLI 向导 + Web 控制台 + 训练推理一致性
  • H 集群最佳实践:brainbox 适配器、rjob 资源申请、反向隧道

…uster practice

- Elastic gateway scaling: static routing + remote-first auto-growth +
  automatic retirement; validated on 4xH200
- Hub deployment mode: pull-based nodes over HTTP for one-way-firewall
  clusters; validated end-to-end with real SGLang inference (reward=1)
- Unified training algorithms (GRPO, DAPO, NPR, OPD, PRM distillation)
  on the shared Slime pipeline, validated on math tasks
- Controller: deployment preflight, port allocation, startup queue,
  cleanup admission
- Config tooling: options CLI wizard + web console + consistency checks
- H-cluster practice: brainbox adapter, rjob provisioning, reverse tunnels

Tests: 795 passed / 2 skipped / 3 pre-existing failures.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant