minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

docs(DeepSeek-V4): bump GB300 Pro PD decode --mem-fraction-static 0.83 → 0.9 (#23698)

Browse Source
This commit is contained in:
fzyzcjy
2026-04-25 16:35:44 +08:00
committed by GitHub
parent 11d77a60df
commit 880599cd43
1 changed files with 5 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx
+5 -3
View File
@@ -500,11 +500,13 @@ export const DeepSeekV4Deployment = () => {
// weights at tp=4 take ~224 GB/card on a 273 GB GB300; runtime needs
// headroom for DeepEP buffer + mooncake KV recv + CG private pool.
// Cookbook defaults (mem-frac 0.874, cg_max_bs 512, max-running 256)
// OOM during CG capture. Verified working on 2026-04-25 (journal
// 2026-04-25-001 Cell D, Δ10).
// OOM during CG capture. mem-frac sweep at 0.83 / 0.87 / 0.89 / 0.91
// all pass static smoke; 0.9 picked as the default — leaves
// ~14 GB / GPU post-CG headroom for mooncake transfer + activation
// peaks while giving ~1M-token KV pool.
if (isGB300 && modelSize === "big") {
flags.push(" --max-running-requests 128");
flags.push(" --mem-fraction-static 0.83");
flags.push(" --mem-fraction-static 0.9");
flags.push(" --cuda-graph-max-bs 128");
} else {
flags.push(" --max-running-requests 256");