Enable DeepGemm warmup in DeepSeek-V4 cookbook (#23883)

This commit is contained in:
Baizhou Zhang
2026-04-27 18:41:12 -07:00
committed by GitHub
parent b1e1fe8eee
commit 3177fa7951
@@ -257,7 +257,6 @@ export const DeepSeekV4Deployment = () => {
// ---- env ---- // ---- env ----
// _LAUNCH_HEAD always prepends these: // _LAUNCH_HEAD always prepends these:
const COMMON_ENV = ["SGLANG_JIT_DEEPGEMM_PRECOMPILE=0"];
// Per-hardware env (whitelist #1: NVSHMEM removed for B200). // Per-hardware env (whitelist #1: NVSHMEM removed for B200).
const HW_ENV = { const HW_ENV = {
h200: ["SGLANG_DSV4_FP4_EXPERTS=0"], // allinone _ENV_H200 h200: ["SGLANG_DSV4_FP4_EXPERTS=0"], // allinone _ENV_H200
@@ -467,8 +466,8 @@ export const DeepSeekV4Deployment = () => {
flags.push(" --host 0.0.0.0"); flags.push(" --host 0.0.0.0");
flags.push(" --port 30000"); flags.push(" --port 30000");
// Assemble: [HW env] [recipe env] [common env] \ sglang serve \ flags... // Assemble: [HW env] [recipe env] \ sglang serve \ flags...
const envAll = [...HW_ENV, ...recipeEnv, ...COMMON_ENV]; const envAll = [...HW_ENV, ...recipeEnv];
const envBlock = envAll.length ? envAll.join(" \\\n") + " \\\n" : ""; const envBlock = envAll.length ? envAll.join(" \\\n") + " \\\n" : "";
const base = `${envBlock}sglang serve \\\n${flags.join(" \\\n")}`; const base = `${envBlock}sglang serve \\\n${flags.join(" \\\n")}`;
// GB200 multinode may need machine-specific NVSHMEM / Gloo env vars; // GB200 multinode may need machine-specific NVSHMEM / Gloo env vars;
@@ -527,7 +526,6 @@ export const DeepSeekV4Deployment = () => {
// NCCL_MNNVL_ENABLE / NCCL_CUMEM_ENABLE may also be needed depending on the // NCCL_MNNVL_ENABLE / NCCL_CUMEM_ENABLE may also be needed depending on the
// GB300 cluster's NVLink/IB topology — see §3.2 "Configuration Tips" note. // GB300 cluster's NVLink/IB topology — see §3.2 "Configuration Tips" note.
const MNNVL_ENV = isGB300 ? ["SGLANG_MOONCAKE_CUSTOM_MEM_POOL=True"] : []; const MNNVL_ENV = isGB300 ? ["SGLANG_MOONCAKE_CUSTOM_MEM_POOL=True"] : [];
const COMMON_ENV = ["SGLANG_JIT_DEEPGEMM_PRECOMPILE=0"];
const buildRole = (mode, port, distPort) => { const buildRole = (mode, port, distPort) => {
const roleEnv = []; const roleEnv = [];
@@ -552,7 +550,7 @@ export const DeepSeekV4Deployment = () => {
if (hardware === "h200" && modelSize === "big") { if (hardware === "h200" && modelSize === "big") {
roleEnv.push("SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=128"); roleEnv.push("SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=128");
} }
const envAll = [...HW_ENV, ...roleEnv, ...MNNVL_ENV, ...COMMON_ENV]; const envAll = [...HW_ENV, ...roleEnv, ...MNNVL_ENV];
const envBlock = envAll.length ? envAll.join(" \\\n") + " \\\n" : ""; const envBlock = envAll.length ? envAll.join(" \\\n") + " \\\n" : "";
const flags = []; const flags = [];