minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

docs(cookbook): add GB10 (DGX Spark) MXFP4 cells for Ling-3.0-flash (#36364)

Browse Source
This commit is contained in:
Xinyuan Tong
2026-08-27 12:15:17 -07:00
committed by GitHub
parent 20a491d1d3
commit 11de5e2281
2 changed files with 57 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/src/snippets/configs/inclusionAI/ling-3.0-flash-benchmarks.jsx
+28
View File
@@ -165,6 +165,34 @@ export const benchmarks = [
notes: "Full GSM8K stop rate 99.62%; default decode CUDA Graph captured 40 shapes through batch 314.",
},
// ====================================================================
// GB10 + MXFP4 (TP1, DGX Spark sm121)
// ====================================================================
{
match: { hw: "dgx-spark", variant: "default", quant: "mxfp4", strategy: "low-latency", spec: "dspark", nodes: "single" },
sglang_version: "PR #33561 @ 2f85329efe",
speed: [
{ workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
ttft_ms: 2172.03, tpot_ms: 9.48, tokens_per_sec_per_gpu: 580 },
{ workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 16 },
ttft_ms: 71292.20, tpot_ms: 38.09, tokens_per_sec_per_gpu: 1157 },
],
accuracy: { gsm8k_pct: 96.36 },
notes: "Full GSM8K stop rate 99.55%; single-chip memory ceiling makes c=16 prefill the bottleneck.",
},
{
match: { hw: "dgx-spark", variant: "default", quant: "mxfp4", strategy: "high-throughput", spec: "off", nodes: "single" },
sglang_version: "PR #33561 @ 2f85329efe",
speed: [
{ workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 1 },
ttft_ms: 2068.91, tpot_ms: 25.76, tokens_per_sec_per_gpu: 324 },
{ workload: { dataset: "random", isl: 8192, osl: 1024, max_concurrency: 16 },
ttft_ms: 21649.69, tpot_ms: 106.51, tokens_per_sec_per_gpu: 1129 },
],
accuracy: { gsm8k_pct: 96.82 },
notes: "Full GSM8K stop rate 99.62%.",
},
// ====================================================================
// GB300 + BF16 (TP4)
// ====================================================================
docs/src/snippets/configs/inclusionAI/ling-3.0-flash.jsx
+29 -1
View File
@@ -1,7 +1,7 @@
export const config = {
modelName: "Ling-3.0-flash",
supportedHardware: ["h20-3e", "h200", "h800", "h100", "b200", "gb300"],
supportedHardware: ["h20-3e", "h200", "h800", "h100", "b200", "gb300", "dgx-spark"],
groupHardware: false,
variants: [{ id: "default", label: "Ling-3.0-flash" }],
@@ -46,6 +46,7 @@ export const config = {
"h100": "lmsysorg/sglang:dev-Ling-3.0-flash",
"b200": "lmsysorg/sglang:dev-Ling-3.0-flash",
"gb300": "lmsysorg/sglang:dev-Ling-3.0-flash",
"dgx-spark": "lmsysorg/sglang:dev-Ling-3.0-flash",
},
dockerHostNetworkWhen: (_sel, { flags }) =>
@@ -374,6 +375,19 @@ sgl-eval run gsm8k \\
"--port {{PORT}}",
],
},
{
match: { hw: "dgx-spark", variant: "default", quant: "mxfp4", strategy: "low-latency", spec: "dspark", nodes: "single" },
verified: true,
flags: [
"--model-path {{MODEL_NAME}}",
"--tp 1",
"--moe-runner-backend flashinfer_mxfp4",
...DSPARK_FLAGS,
"--mem-fraction-static 0.85",
"--host {{HOST_IP}}",
"--port {{PORT}}",
],
},
{
match: { hw: "h20-3e", variant: "default", quant: "bf16", strategy: "high-throughput", spec: "off", nodes: "single" },
verified: false,
@@ -567,6 +581,20 @@ sgl-eval run gsm8k \\
"--port {{PORT}}",
],
},
{
match: { hw: "dgx-spark", variant: "default", quant: "mxfp4", strategy: "high-throughput", spec: "off", nodes: "single" },
verified: true,
flags: [
"--model-path {{MODEL_NAME}}",
"--tp 1",
"--moe-runner-backend flashinfer_mxfp4",
"--mem-fraction-static 0.85",
"--tool-call-parser ling3",
"--reasoning-parser ling3",
"--host {{HOST_IP}}",
"--port {{PORT}}",
],
},
// Hybrid KDA must pass prefix keys to Mooncake; otherwise storage writes are empty.
// Cold uncached extends above chunked_prefill_size skip write-through for that influx.