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

[DeepSeek V4] Add W4A4 MegaMoE server flag (#35918)

Browse Source
This commit is contained in:
Baizhou Zhang
2026-08-21 18:44:18 -07:00
committed by GitHub
parent 0be2a209ac
commit 3b5909de0e
15 changed files with 147 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/src/snippets/_playground.jsx
+14 -9
View File
@@ -611,23 +611,21 @@ export const Playground = ({ config }) => {
// ---- Axis: MoE Parallelism ----------------------------------------------
// Backend single-select + EP numeric knob; either is optional. Picking the
// "megamoe" backend reveals a Quantization sub-select (W4A8 / W4A4) in the same
// row — W4A4 adds the FP4-activations env vars.
// row — W4A4 adds the FP4-activations server flag.
moe: {
initState: () => ({ backend: null, ep: null, mmQuant: null }),
// Prefer --moe-a2a-backend over --moe-runner-backend when both present.
// mmQuant is derived from the base env (FP4 activations present → W4A4).
// mmQuant is derived from the base flag (FP4 activations present → W4A4).
deriveFromBase: (cell, fc, h) => {
const flags = (cell && cell.flags) || [];
const baseEnv = (cell && cell.env) || [];
const a2a = h.findFlagArg(flags, "--moe-a2a-backend");
const runner = h.findFlagArg(flags, "--moe-runner-backend");
const fp4Acts = baseEnv.some(
(e) => e.startsWith("SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS"));
const w4a4 = h.hasFlag(flags, "--enable-w4a4-mxfp4-megamoe");
return {
backend: a2a || runner || null,
ep: h.parseIntFlagAny(flags, h.EP_HEADS),
mmQuant: fp4Acts ? "w4a4" : "w4a8",
mmQuant: w4a4 ? "w4a4" : "w4a8",
};
},
@@ -651,15 +649,19 @@ export const Playground = ({ config }) => {
if (opt?.env?.length) env = [...env, ...opt.env];
}
// MegaMoE owns the MoE path: when the effective backend is megamoe, strip the
// DeepEP dispatch + any prior megamoe env, then re-add the selected quant's
// env. When the backend is explicitly switched away from megamoe, only drop
// the megamoe quant env (leave DeepEP dispatch intact).
// DeepEP dispatch + any prior MegaMoE quant settings, then re-add the
// selected quant's flags/env. When the backend is explicitly switched
// away from MegaMoE, only drop the MegaMoE quant settings (leave DeepEP
// dispatch intact).
const mq = fc.megamoeQuant;
if (mq) {
const quantKeys = [];
const quantFlagHeads = [];
for (const o of (mq.options || [])) {
for (const e of (o.env || [])) quantKeys.push(e.split("=")[0]);
for (const f of (o.flags || [])) quantFlagHeads.push(f.split(/[\s=]/)[0]);
}
flags = h.stripFlagsByFirstToken(flags, quantFlagHeads);
const effBackend = value.backend !== null
? value.backend : (derived && derived.backend);
if (effBackend === "megamoe") {
@@ -667,6 +669,9 @@ export const Playground = ({ config }) => {
const quant = value.mmQuant != null
? value.mmQuant : ((derived && derived.mmQuant) || "w4a8");
const opt = (mq.options || []).find((o) => o.id === quant);
if (opt?.flags?.length) {
flags = h.insertAfter(flags, h.ANCHOR_NEAR_MOE, opt.flags);
}
if (opt?.env?.length) env = [...env, ...opt.env];
} else if (value.backend !== null) {
env = h.stripEnvByPrefix(env, quantKeys);
docs/src/snippets/configs/Qwen/qwen3.8.jsx
+2 -5
View File
@@ -178,11 +178,8 @@ export const config = {
{ id: "w4a8", label: "W4A8",
env: ["SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320"] },
{ id: "w4a4", label: "W4A4",
env: [
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320",
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS=1",
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND=1",
] },
flags: ["--enable-w4a4-mxfp4-megamoe"],
env: ["SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320"] },
],
},
// WideEP — the launch post's large-scale-EP claim — so the range goes past
docs/src/snippets/configs/deepseek-ai/deepseek-v4.jsx
+2 -5
View File
@@ -258,11 +258,8 @@ sgl-eval run aime25 \\
{ id: "w4a8", label: "W4A8",
env: ["SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320"] },
{ id: "w4a4", label: "W4A4",
env: [
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320",
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS=1",
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND=1",
] },
flags: ["--enable-w4a4-mxfp4-megamoe"],
env: ["SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320"] },
],
},
ep: { label: "EP", values: [
docs/src/snippets/configs/moonshotai/kimi-k3.jsx
+2 -5
View File
@@ -513,11 +513,8 @@ export const config = {
{ id: "w4a8", label: "W4A8",
env: ["SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320"] },
{ id: "w4a4", label: "W4A4",
env: [
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320",
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS=1",
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND=1",
] },
flags: ["--enable-w4a4-mxfp4-megamoe"],
env: ["SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320"] },
],
},
ep: { label: "EP", values: [