Repository navigation
[Doc] cookbook: Add Nemotron 3.5 Super VL - #44510
Merged
Merged
Conversation
Contributor
|
Preview deployment for your docs. Learn more about Mintlify Previews.
💡 Tip: Enable Automations to automatically generate PRs for you. |
Jiminator
marked this pull request as ready for review
October 15, 2026 15:24
Jiminator
requested review from
JustinTong0323,
sogalin,
wisclmy0611 and
zijiexia
as code owners
October 15, 2026 15:24
zijiexia
approved these changes
Oct 15, 2026
This branch was successfully deployed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Add an interactive Nemotron 3.5 Super VL cookbook at
/cookbook/autoregressive/NVIDIA/Nemotron3.5-Super. The SGLang model source is on thenemotron3-super-dev-20261015branch.Modifications
bfloat16andfloat32. High Throughput omits MTP; Low Latency enables EAGLE MTP.--moe-runner-backend tritonon H200. Omit the explicit attention backend, maximum running requests, and static memory fraction from B200 recipes; omit the last two limits from H200 recipes.Accuracy Tests
The bfloat16 Mamba-state source runs cover all ten supported hardware/quantization/strategy cells. All eight non-Spark float32 selections completed 300-question text-only GSM8K runs with zero request errors. Spark float32 testing was stopped before completing the cohort.
The completed results used the early-access NGC weights and source at
2c7511f6f7cf1cdf9ae568ad0fb25f291edb10da. The command panel uses the publiclmsysorg/sglang:nemotron3-5-supertag and the requested public Hugging Face checkpoint IDs. Those public artifacts have not been verified by the accuracy runs.Speed Tests and Profiling
Not applicable to this documentation change. The strategy labels do not assert measured throughput or latency gains.
Validation
node docs/scripts/check_cookbook_configs.mjspassed.mint validatepassed.mint broken-linkspassed.git diff --checkpassed.CI States
Latest PR Test (Base): ✅ Run #38761659422
Latest PR Test (Extra): ❌ Run #38761659039
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.