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

docs: prepend SGLANG_JIT_DEEPGEMM_PRECOMPILE=0 for H200 FP8 Flash max-throughput (#25152)

Browse Source 
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
zijiexia
2026-05-13 00:50:15 -07:00
committed by GitHub
co-authored by Claude Opus 4.7
parent c32f2dc1ac
commit 7d515c6d1f
1 changed files with 3 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx
+3
View File
@@ -391,6 +391,9 @@ export const DeepSeekV4Deployment = () => {
} }
} else if (recipe === "max-throughput") { } else if (recipe === "max-throughput") {
if (hardware === "h200") { if (hardware === "h200") {
if (!isBig) {
recipeEnv.push("SGLANG_JIT_DEEPGEMM_PRECOMPILE=0");
}
recipeEnv.push(isBig recipeEnv.push(isBig
? "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=128" ? "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=128"
: "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=256"); : "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=256");