diff --git a/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-V4.mdx b/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-V4.mdx
index e99e29499..0bbb2ad6f 100644
--- a/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-V4.mdx
+++ b/docs_new/cookbook/autoregressive/DeepSeek/DeepSeek-V4.mdx
@@ -100,6 +100,10 @@ Please refer to the [official SGLang installation guide](../../../docs/get-start
| NVIDIA H200 |
lmsysorg/sglang:deepseek-v4-hopper |
+
+ | NVIDIA H100 |
+ lmsysorg/sglang:dev |
+
diff --git a/docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx b/docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx
index 080a8e7ec..79b79b863 100644
--- a/docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx
+++ b/docs_new/src/snippets/autoregressive/deepseek-v4-deployment.jsx
@@ -7,6 +7,7 @@ export const DeepSeekV4Deployment = () => {
// GB200 → FP4 weights, Flash TP=4 / Pro TP=8 2-node
// GB300 → FP4 weights, Flash TP=4 / Pro TP=4 single-node
// H200 → FP8 weights, Flash TP=4 / Pro TP=16 2-node
+ // H100 → FP4 weights (Marlin), Flash TP=8 single-node / Pro TP=16 2-node
// Model variant → HF slug:
// Flash (285B) → deepseek-ai/DeepSeek-V4-Flash
// Pro (1.6T) → deepseek-ai/DeepSeek-V4-Pro
@@ -32,6 +33,7 @@ export const DeepSeekV4Deployment = () => {
{ id: "gb300", label: "GB300 (FP4)", default: false },
{ id: "h200", label: "H200 (FP8)", default: false },
{ id: "h200-fp4", label: "H200 (FP4)", default: false },
+ { id: "h100", label: "H100 (FP4)", default: false },
],
},
modelSize: {
@@ -71,14 +73,17 @@ export const DeepSeekV4Deployment = () => {
},
};
- // Recipes that are not supported on the H200 (FP4) Marlin path.
- const H200_FP4_UNSUPPORTED_RECIPES = new Set(["cp", "pd-disagg"]);
+ // Recipes that are not supported on the Marlin (FP4) Hopper paths
+ // (H200 FP4, H100 FP4).
+ const MARLIN_UNSUPPORTED_RECIPES = new Set(["cp", "pd-disagg"]);
+ const MARLIN_HARDWARE = new Set(["h200-fp4", "h100"]);
+ const MARLIN_LABEL = { "h200-fp4": "H200 (FP4)", h100: "H100 (FP4)" };
const resolveItems = (option, vals) => {
- if (option.name === "recipe" && vals && vals.hardware === "h200-fp4") {
+ if (option.name === "recipe" && vals && MARLIN_HARDWARE.has(vals.hardware)) {
return option.items.map((it) =>
- H200_FP4_UNSUPPORTED_RECIPES.has(it.id)
- ? { ...it, disabled: true, disabledReason: "Not supported on H200 (FP4)" }
+ MARLIN_UNSUPPORTED_RECIPES.has(it.id)
+ ? { ...it, disabled: true, disabledReason: `Not supported on ${MARLIN_LABEL[vals.hardware]}` }
: it
);
}
@@ -119,12 +124,13 @@ export const DeepSeekV4Deployment = () => {
const handleRadioChange = (optionName, value) => {
setValues((prev) => {
const next = { ...prev, [optionName]: value };
- // Switching to H200 (FP4) while cp / pd-disagg is selected: fall back
- // to low-latency since those recipes are not supported on this path.
+ // Switching to a Marlin (FP4) Hopper path while cp / pd-disagg is
+ // selected: fall back to low-latency since those recipes are not
+ // supported on Marlin.
if (
optionName === "hardware" &&
- value === "h200-fp4" &&
- H200_FP4_UNSUPPORTED_RECIPES.has(next.recipe)
+ MARLIN_HARDWARE.has(value) &&
+ MARLIN_UNSUPPORTED_RECIPES.has(next.recipe)
) {
next.recipe = "low-latency";
}
@@ -177,6 +183,10 @@ export const DeepSeekV4Deployment = () => {
// single-node TP=4 / TP=8 deployment fits Flash / Pro on Hopper.
"h200-fp4|small": { slug: "deepseek-ai/DeepSeek-V4-Flash", tp: 4, multinode: false },
"h200-fp4|big": { slug: "deepseek-ai/DeepSeek-V4-Pro", tp: 8, multinode: false },
+ // H100 (FP4) also uses the Marlin runner, but Hopper memory pressure forces
+ // a higher TP: Flash fits at TP=8 single-node, Pro needs TP=16 across 2 nodes.
+ "h100|small": { slug: "deepseek-ai/DeepSeek-V4-Flash", tp: 8, multinode: false },
+ "h100|big": { slug: "deepseek-ai/DeepSeek-V4-Pro", tp: 16, multinode: true, nnodes: 2 },
};
// Per (hardware, modelSize) PD role TP (from allinone _PD_SPEC).
const PD_TP_SPEC = {
@@ -238,6 +248,12 @@ export const DeepSeekV4Deployment = () => {
"h200-fp4|big|low-latency",
"h200-fp4|big|balanced",
"h200-fp4|big|max-throughput",
+ "h100|small|low-latency",
+ "h100|small|balanced",
+ "h100|small|max-throughput",
+ "h100|big|low-latency",
+ "h100|big|balanced",
+ "h100|big|max-throughput",
]);
// Recipes whose command is intentionally not yet provided (e.g. blocked by an
// upstream limitation). Showing a minimal placeholder is friendlier to users
@@ -331,6 +347,60 @@ export const DeepSeekV4Deployment = () => {
: `${BEING_VERIFIED_NOTE}\n${commentOutCommand(fp4Cmd)}`;
}
+ // H100 (FP4) Marlin path: also Hopper + Marlin, but unlike H200 (FP4) the
+ // memory budget forces a higher TP and (for Pro) a 2-node deployment.
+ // Flash: TP=8, single node
+ // Pro: TP=16, 2 nodes, env SGLANG_SHARED_EXPERT_TP1=1
+ // + mem-fraction-static 0.9; cg=8 / max-run=32 on low-lat/balanced
+ // low-latency: MTP 3 / 1 / 4 (steps / topk / draft-tokens)
+ // balanced: MTP 1 / 1 / 2
+ // max-throughput: MTP disabled
+ if (hardware === "h100") {
+ const verifyKey = `${hardware}|${modelSize}|${recipe}`;
+ if (TBD_RECIPES.has(verifyKey)) return TBD_PLACEHOLDER;
+
+ const h100Env = isBig ? ["SGLANG_SHARED_EXPERT_TP1=1"] : [];
+ const h100EnvBlock = h100Env.length ? h100Env.join(" \\\n") + " \\\n" : "";
+
+ const h100Flags = [
+ " --trust-remote-code",
+ ` --model-path ${slug}`,
+ ` --tp ${tp}`,
+ ];
+ if (multinode) h100Flags.push(...multiNodeFlags(nnodes));
+ h100Flags.push(" --moe-runner-backend marlin");
+ if (recipe === "low-latency") {
+ h100Flags.push(" --speculative-algo EAGLE");
+ h100Flags.push(" --speculative-num-steps 3");
+ h100Flags.push(" --speculative-eagle-topk 1");
+ h100Flags.push(" --speculative-num-draft-tokens 4");
+ } else if (recipe === "balanced") {
+ h100Flags.push(" --speculative-algo EAGLE");
+ h100Flags.push(" --speculative-num-steps 1");
+ h100Flags.push(" --speculative-eagle-topk 1");
+ h100Flags.push(" --speculative-num-draft-tokens 2");
+ }
+ if (isBig) {
+ h100Flags.push(" --mem-fraction-static 0.9");
+ // max-throughput leaves cg/max-run at engine defaults; low-lat + balanced
+ // cap them tight to keep latency predictable on Hopper.
+ if (recipe !== "max-throughput") {
+ h100Flags.push(" --cuda-graph-max-bs 8");
+ h100Flags.push(" --max-running-requests 32");
+ }
+ }
+ if (toolcall === "enabled") h100Flags.push(" --tool-call-parser deepseekv4");
+ if (reasoningParser === "enabled") h100Flags.push(" --reasoning-parser deepseek-v4");
+ h100Flags.push(" --host 0.0.0.0");
+ h100Flags.push(" --port 30000");
+
+ const h100Cmd = `${h100EnvBlock}sglang serve \\\n${h100Flags.join(" \\\n")}`;
+ const h100WithNote = multinode ? prependMultiNodeNote(h100Cmd, nnodes) : h100Cmd;
+ return VERIFIED_RECIPES.has(verifyKey)
+ ? h100WithNote
+ : `${BEING_VERIFIED_NOTE}\n${commentOutCommand(h100WithNote)}`;
+ }
+
// ---- env ----
// _LAUNCH_HEAD always prepends these:
// Per-hardware env (whitelist #1: NVSHMEM removed for B200).