From 6afebc278ad9a56cc7d4462635851671668e675b Mon Sep 17 00:00:00 2001 From: zijiexia <37504505+zijiexia@users.noreply.github.com> Date: Tue, 26 May 2026 12:08:48 -0700 Subject: [PATCH] [docs] DeepSeek-V4 cookbook: note cu129 image for GB200 Pro DeepEP backend (#26413) Co-authored-by: Claude Opus 4.7 (1M context) --- .../autoregressive/deepseek-v4-deployment.jsx | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx b/docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx index fad5f99c8..006e28d73 100644 --- a/docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx +++ b/docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx @@ -909,6 +909,20 @@ export const DeepSeekV4Deployment = () => { `# NVSHMEM_HCA_LIST=\n` + cmd; } + // GB200 Pro with MegaMoE disabled runs the DeepEP a2a backend, which is + // currently only packaged in the CUDA 12.9 image — the default `:latest` + // ships CUDA 13 and does not include a compatible DeepEP build. + if ( + hardware === "gb200" && + isBig && + megamoe === "disabled" && + flags.some((f) => f.includes("--moe-a2a-backend deepep")) + ) { + cmd = + `# NOTE: for the DeepEP backend, use the cu129 docker image\n` + + `# (lmsysorg/sglang:latest-cu129) instead of the default \`:latest\`.\n` + + cmd; + } const withMultinode = multinode ? prependMultiNodeNote(cmd, nnodes) : cmd; // H200 Pro low-latency: show BOTH a single-node (TP=8 marlin) variant