Repository navigation
[diffusion] MiniMax-H3: warm up at the served clip shape - #37945
Merged
Merged
Conversation
The synthetic warmup request rendered the default 5-second 768p clip regardless of --warmup-num-frames / --warmup-resolutions, so the first served forward at another clip length paid allocator growth and first-call kernel setup in every DiT block (2 to 3 s per forward on a 14 s clip). The warmup target now honors both flags. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
kevin-mii
requested review from
AgainstEntropy,
BBuf,
HaiShaw,
mickqian,
ping1jing2 and
yichiche
as code owners
September 4, 2026 06:33
5 tasks done
…armup Only the aspect ratio comes from --warmup-resolutions. BCG seeds that flag with an area-capped default (832x464 for H3), which the previous revision turned into a 464-short-edge warmup instead of the released 768p canvas. Reuse the H3 fps / duration / short-edge constants, reject frame counts outside the 4-15 s admission range with a flag-named error, and drop the positivity check server args already perform. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Admission already range-checks duration_seconds, and H3 only serves the named aspect ratios, so the hook just converts the two flags and snaps to the nearest ratio. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
kevin-mii
added a commit
to kevin-mii/sglang
that referenced
this pull request
Sep 4, 2026
fp8 per-tensor becomes its own quantization method, fp8_per_tensor (fp8_per_tensor.py subclassing the online fp8 method), registered by name; fp8.py, PipelineConfig and the loader keep their shapes and VDN-H3 admission maps --quantization fp8 to it. The out-of-place QK-norm + RoPE JIT variant is dropped (clone + the existing in-place kernel, ~1%). The H3 served-shape warmup fix moves to its own PR (sgl-project#37945). One benchmark preset; no package re-export of the VDN config. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Collaborator
|
/tag-and-rerun-ci |
StevenChenSE
pushed a commit
to StevenChenSE/sglang
that referenced
this pull request
Sep 6, 2026
…roject#37945) Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
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
2 participants
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
MiniMax-H3's synthetic warmup rendered the default 5-second 768p clip regardless of
--warmup-num-frames/--warmup-resolutions. The first served forward at another clip length then paid allocator growth and first-call kernel setup in every DiT block: 2 to 3 s per forward on a 14 s clip (measured on 4x and 8x B200).Modifications
MiniMaxH3SamplingParamsderives the warmup duration from--warmup-num-frames(24 fps; admission range-checks the duration as for any request) and the aspect ratio from--warmup-resolutions(_synthetic_warmup_target), keeping the released 5-second 16:9 default otherwise. The short edge stays at the released 768: breakable CUDA graphs seed--warmup-resolutionswith an area-capped default (832x464 for H3) that must not shrink the canvas. Unit testtest_synthetic_warmup_target_honors_warmup_flags.Split out of #37903, which depends on it for its served-shape warmup numbers.
Checklist
🤖 Generated with Claude Code
CI States
Latest PR Test (Base): ✅ Run #33854829199
Latest PR Test (Extra): ❌ Run #33854829117
Latest PR Test (AMD ROCm 7.2): ⏳ Run #33854829201