Repository navigation
[diffusion] docs: state host-memory requirements for MiniMax-H3 resident recipes - #34909
Closed
Raymondlol wants to merge 1 commit into
Closed
Raymondlol wants to merge 1 commit into
Raymondlol wants to merge 1 commit into
Conversation
…ent recipes
Resident multi-GPU loading stages one full DiT copy per rank through host RAM
before placement. Measured on H200 (3 s sampling, isolated to the DiT load):
114.3 GiB at two ranks, 233.5 GiB at four (~58 GiB per rank, linear), with
container-wide usage peaking near 380 GiB for the recommended 4-GPU recipe.
Below that the load is OOM-killed by the OS with no server-side error.
The cookbook already states a host-memory requirement for the RTX 5090 recipe
("377 GiB host; use a 384 GiB-class machine"); this applies the same convention
to the datacenter recipes, and adds the host-RAM axis to the resident-vs-FSDP
framing, which previously discussed only GPU residency and collectives (FSDP
loads sharded: 12.8 GiB host for the same two-rank load).
Addresses sgl-project#34902.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Raymondlol
requested review from
JustinTong0323,
sogalin,
wisclmy0611 and
zijiexia
as code owners
August 15, 2026 05:06
github-actions
Bot
added
the
documentation
This was referenced Aug 15, 2026
Contributor
Author
|
Closing: the cookbook has moved on (layerwise offload is now the auto policy for H3, and the page has been restructured several times since), so this paragraph no longer fits. The measured host-RAM figures for resident multi-GPU loading remain in #34902 for anyone who needs them. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
Resident multi-GPU MiniMax-H3 loading stages a full DiT copy per rank through host RAM — measured on H200: 114.3 GiB at 2 ranks, 233.5 GiB at 4, container-wide peak near 380 GiB on the 4×H200 recipe — and below that the load is OOM-killed with no server-side error. The cookbook states a host-memory requirement only for the RTX 5090 recipe, and frames the resident-vs-FSDP trade in GPU terms only. Addresses #34902.
Modifications
--use-fsdp-inferenceconfiguration note.h200/residentcell; extend the two "Prefer Resident" FSDP warns with the host-RAM axis.Accuracy Tests
Not applicable; documentation-only.
Speed Tests and Profiling
Not applicable; documentation-only.
Validation
node docs/scripts/check_cookbook_configs.mjs— OKCI States
Latest PR Test (Base): ✅ Run #31866131714
Latest PR Test (Extra): ❌ Run #31866131593