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

[diffusion] docs: verify the DGX Spark H3 recipe (#37456)

Browse Source 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Mick
2026-09-06 23:10:27 +08:00
committed by GitHub
co-authored by Claude Fable 5
parent be00a543a7
commit ade1da017f
2 changed files with 24 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/src/snippets/configs/MiniMaxAI/minimax-h3.jsx
+15 -14
View File
@@ -24,10 +24,11 @@ const CONSUMER_24G = ["rtx4090", "rtx3090"];
// their recipes are derived from the tier logic, not verified runs.
const WORKSTATION_48G = ["rtx6000ada"];
const WORKSTATION_96G = ["rtxpro6000"];
// GB10 unified memory: 128 GB shared between CPU and GPU, so the VRAM/host
// split that shapes every tier above does not exist. The whole 108 GB
// deployment fits, and an offloaded component's "copy to device" is an
// in-memory copy on the coherent bus. No hard-cap anchor was measured.
// GB10 unified memory: 128 GB shared between CPU and GPU (121.7 GB visible
// to torch), so the VRAM/host split that shapes every tier above does not
// exist. The 134 GiB deployment still exceeds the pool, and the loader's
// automatic placement handles that split better than any explicit flag set:
// verified on DGX Spark, see unified128Flags().
const UNIFIED_128G = ["dgx-spark"];
const CONSUMER_SINGLE = [
...CONSUMER_12G,
@@ -73,14 +74,12 @@ function consumerFlags(s) {
}
function unified128Flags() {
// Unified memory holds the whole deployment: the memory manager pins every
// component (the budget sees ~128 GB), and streamed copies move at memory
// speed on the coherent bus. video_vae=36 still buys the fast decode.
return [
"--performance-mode memory",
"--layerwise-offload-components dit,text_encoder,vae",
"--layerwise-resident-layers video_vae=36",
];
// No flags: the deployment (134 GiB) exceeds the pool, automatic offload
// engages on its own and pins 42 of 50 DiT layers. Measured on DGX Spark,
// the explicit discrete-GPU recipe (--performance-mode memory + offload
// components + video_vae=36) ran the same denoise 2.1x slower (25.8 vs
// 12.1 s/it) -- do not carry discrete-card flags onto unified memory.
return [];
}
function workstation96Flags() {
@@ -117,8 +116,10 @@ function consumerHints(s) {
}
if (UNIFIED_128G.includes(s.hw)) {
return [
"derived recipe, not yet verified: the 128 GB unified pool holds the whole 108 GB deployment, so the memory manager pins every component and offload copies run at memory speed on the coherent bus",
"expect step times above the discrete-GPU rows: the GB10's ~273 GB/s memory bandwidth is the denoise ceiling, not the placement",
"verified on DGX Spark at 480P: ~12.1 s per denoise step steady-state, ~40 s decode, ~12 min per warm request -- with no flags at all; adding the discrete-GPU offload flags measured 2.1x slower on the same box",
"the text encoder runs ~5.5 min per request and does not warm up: it is steady-state compute on this chip, not a stall -- budget for it",
"expect ~12 min of server load before the first request; the first request itself runs at full speed (no JIT tax was measured)",
"step times sit above the discrete-GPU rows because the GB10's ~273 GB/s memory bandwidth is the denoise ceiling, not the placement",
];
}
if (WORKSTATION_96G.includes(s.hw)) {