Repository navigation
docs(cookbook): add GB10 (DGX Spark) MXFP4 cells for Ling-3.0-flash - #36364
Merged
Merged
Conversation
TP1 + flashinfer_mxfp4 on a single GB10 chip, both strategies: - low-latency + DSPARK (external draft) - high-throughput + spec off Measured on GB10 (sm121): speed bench (ISL 8192 / OSL 1024, 80 exact-length requests, c=1 and c=16) and accuracy notes for both cells; full GSM8K runs are in flight and will fill the remaining accuracy field.
JustinTong0323
requested review from
sogalin,
wisclmy0611 and
zijiexia
as code owners
August 25, 2026 18:52
github-actions
Bot
added
the
documentation
Collaborator
Author
|
Full GSM8K (1319) results are in for both cells:
Both consistent with the H200/B200 mxfp4 cells (95.3-96.7%). Cookbook data now complete: deployment configs, speed benches (c=1/16), and full accuracy for both GB10 cells. |
- hw id gb10 -> dgx-spark so the selector renders 'DGX Spark' (the id already exists in the shared HARDWARE_CATALOG alongside b300/gb300) - docker image -> lmsysorg/sglang:dev-Ling-3.0-flash (multi-arch incl. arm64; verified on GB10: BailingMoeV3 + FlashInfer CUTLASS MXFP4 import cleanly). lmsysorg/sglang:latest predates sgl-project#33561 and lacks the model.
zijiexia
approved these changes
Aug 27, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Adds GB10 (DGX Spark, sm121) as a hardware option for the Ling-3.0-flash cookbook, MXFP4 quant, TP1 on the single chip:
--moe-runner-backend flashinfer_mxfp4)All numbers measured on a real GB10 box with the #33561 head (
2f85329efe, includes the SwigluStep clamp fix required for correct output on the SM12x CUTLASS path):DSPARK decode is ~2.7x faster than greedy at bs=1 on GB10. The spec-off cell's GSM8K is 100% on a deterministic 20-question window; full 1319-question GSM8K runs for both cells are in flight and the dspark accuracy field will be updated when they land.
Verified:
mint validatepasses; localmint devrenders both cells (config + benchmark table) under hw=gb10.CI States
Latest PR Test (Base): ✅ Run #33037260605
Latest PR Test (Extra): ❌ Run #33037260462
Latest PR Test (AMD ROCm 7.2): ➖ No AMD PR run found for this commit.