From abd1761956644c9069313ce3d3829a6b2842a72f Mon Sep 17 00:00:00 2001 From: Hao Liu Date: Mon, 24 Aug 2026 08:42:44 +0800 Subject: [PATCH] docs: remove stale delta NCCL recommendation Related to #2209 Signed-off-by: Hao Liu --- docs/en/advanced/external-rollout-engines.md | 3 ++- docs/zh/advanced/external-rollout-engines.md | 3 ++- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/docs/en/advanced/external-rollout-engines.md b/docs/en/advanced/external-rollout-engines.md index b12b3cb8c6..8efd245fd9 100644 --- a/docs/en/advanced/external-rollout-engines.md +++ b/docs/en/advanced/external-rollout-engines.md @@ -14,9 +14,10 @@ This page is a roadmap. Use it to decide when to use `--rollout-external-engine- | Trainer and external engines cannot form an NCCL group, but can see the same filesystem path | `--update-weight-mode full --update-weight-transport disk` | | Full checkpoints are too heavy for large-model cross-cluster or cross-DC sync | `--update-weight-mode delta --update-weight-transport disk` | | Rollout serving can use an independent SGLang environment, or even different GPU models/vendors | external engines + disk transport | -| You want to validate delta wire/apply logic inside one datacenter | `--update-weight-mode delta --update-weight-transport nccl` | | You need frozen reference, reward, or tool-side models | Prefer `update_weights: false` in [SGLang Config](sglang-config.md#3-multi-model-serving) | +Delta mode supports disk transport only. Use full mode when syncing weights over NCCL. + ## What External Engine Does First launch SGLang servers independently: diff --git a/docs/zh/advanced/external-rollout-engines.md b/docs/zh/advanced/external-rollout-engines.md index c077f82608..53b4571b10 100644 --- a/docs/zh/advanced/external-rollout-engines.md +++ b/docs/zh/advanced/external-rollout-engines.md @@ -14,9 +14,10 @@ External rollout engine 指的是：SGLang engine 不由 slime 训练任务启 | 训练器和 external engine 不能建立 NCCL group，但能共享同一路径的文件系统 | `--update-weight-mode full --update-weight-transport disk` | | 大模型跨集群或跨数据中心同步，full checkpoint 太重 | `--update-weight-mode delta --update-weight-transport disk` | | rollout serving 想使用独立 SGLang 环境，甚至不同型号或不同厂家的 GPU | external engine + disk transport | -| 想验证 delta wire/apply 逻辑，但仍在同一数据中心内 | `--update-weight-mode delta --update-weight-transport nccl` | | 需要 reference、reward、tool-side model 等冻结模型 | 优先用 [SGLang Config](sglang-config.md#3-多模型服务) 的 `update_weights: false` | +delta mode 仅支持 disk transport。通过 NCCL 同步权重时请使用 full mode。 + ## External Engine 做了什么 使用 external engine 时，先独立启动 SGLang server：