Skip to content

[diffusion] model: support minimax-h3 - #33275

Merged
mickqian merged 15 commits into
mainfrom
codex/native-video-audio-pipeline
Aug 2, 2026
Merged

mickqian merged 15 commits into
mainfrom
codex/native-video-audio-pipeline

Conversation

@mickqian

@mickqian mickqian commented Aug 2, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

  • add a fully native joint video/audio generation pipeline with text, endpoint-frame, and reference-media conditioning
  • integrate distributed execution, mixed-precision loading, caching, and online quantization hooks
  • add focused runtime contracts, PR-blocking GPU coverage, and an interactive deployment cookbook

Scope

  • native serving and OpenAI-compatible video API integration
  • tensor and sequence parallel execution
  • video/audio encoding and decoding paths

CI States

Latest PR Test (Base): 🚫 Run #30748870644
Latest PR Test (Extra): ❌ Run #30748870571

Add native MiniMax-H3 serving for FL2VA and Ref2VA variants, including text/image/video/audio conditioning, parallel execution, lossless performance paths, quantization and Cache-DiT integration, focused release contracts, and an interactive deployment cookbook.

Co-authored-by: zhenaozhenfu <zhenaozhenfu@minimaxi.com>
Co-authored-by: BBuf <1182563586@qq.com>
Co-authored-by: andyluo7 <andy.luo@amd.com>
Co-authored-by: Zijie Xia <zijie_xia@icloud.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: chao-xue <877184285@qq.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
@mintlify

mintlify Bot commented Aug 2, 2026 •

Copy link
Copy Markdown
Contributor

Preview deployment for your docs. Learn more about Mintlify Previews.

Project Status Preview Updated (UTC)
lmsysorg 🟢 Ready View Preview Aug 2, 2026, 10:08 AM

💡 Tip: Enable Workflows to automatically generate PRs for you.

@gemini-code-assist

Copy link
Copy Markdown
Contributor

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

@github-actions github-actions Bot added documentation Improvements or additions to documentation quant LLM Quantization Multi-modal multi-modal language model diffusion SGLang Diffusion jit-kernel labels Aug 2, 2026
@mickqian

mickqian commented Aug 2, 2026

Copy link
Copy Markdown
Collaborator Author

/tag-and-rerun-ci

@github-actions github-actions Bot added the run-ci CI: run the baseline test suite on this PR label Aug 2, 2026
@mickqian mickqian changed the title [Diffusion] Add a native video-audio generation pipeline [diffusion] optimization: support cuda-ipc zero-staging all-to-all for 2-rank Ulysses Aug 2, 2026
@mickqian mickqian changed the title [diffusion] optimization: support cuda-ipc zero-staging all-to-all for 2-rank Ulysses [diffusion] model: support minimax-h3 Aug 2, 2026
@mickqian
mickqian merged commit 70fe2e0 into main Aug 2, 2026
187 of 261 checks passed
@mickqian
mickqian deleted the codex/native-video-audio-pipeline branch August 2, 2026 14:32
BBuf added a commit to BBuf/sglang that referenced this pull request Aug 5, 2026
…nd reuse a2a staging buffers

On the generic USPAttention path (heads at dim=2), the Ulysses input
exchange issued three all_to_all_single collectives (Q, K, V), each with
a permute+contiguous on both sides, and every USP collective allocated
fresh send/recv buffers on each of the 2x40 exchanges per denoise step.

- _usp_input_all_to_all_qkv: pack Q/K/V destination-major with the
  existing Triton relayout kernel (reading through strides, no
  contiguous pre-pass), exchange in ONE collective, unpack with views +
  one contiguous pass. Bit-exact fallback to the unpacked path for
  ineligible inputs (CPU, GQA-mismatched shapes, non-contiguous, fp32,
  torch.compile).
- _a2a_staging_buffer: reuse USP collective staging buffers keyed by
  (role, shape, dtype, device); bypassed under autograd, torch.compile
  and CUDA graph capture.
- pack_qkv_destination_major gains an optional out= so the pack kernel
  writes straight into the reusable staging buffer.

Pure data movement, bit-exact by construction and by test: 4-rank NCCL
micro-benchmark asserts torch.equal across variants, and same-seed
Wan2.2-T2V-A14B 4xH200 ulysses=4 frames are byte-identical to main.

Perf (4xH200 NVLink, bf16): input exchange 2.950 -> 2.253 ms (720p),
1.331 -> 1.033 ms (480p), 2.112 -> 1.654 ms (H3 768p), 1.28-1.31x; pack
kernel vs permute reference 2.7x. Wan2.2-A14B 480p x81f x20 steps,
ulysses=4, 9 sequential requests drop-first median: e2e 32.653 -> 31.965
s per request (-2.1%), denoise step median 1535.3 -> 1483.3 ms (-3.4%)
on an idle machine; an independent steady-contention sandwich pairing
(base/opt/base) reproduces -1.7% e2e / -2.3% step.

Adapted from NVlabs Sana sol-engine (Apache-2.0) MiniMax-H3 / Wan2.2
recipes; the relayout kernels were already ported in sgl-project#33275.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
saturn-acc pushed a commit to saturn-acc/sglang that referenced this pull request Aug 16, 2026
Co-authored-by: zhenaozhenfu <zhenaozhenfu@minimaxi.com>
Co-authored-by: BBuf <1182563586@qq.com>
Co-authored-by: andyluo7 <andy.luo@amd.com>
Co-authored-by: Zijie Xia <zijie_xia@icloud.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: chao-xue <877184285@qq.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Atituiset pushed a commit to Atituiset/sglang that referenced this pull request Sep 10, 2026
Co-authored-by: zhenaozhenfu <zhenaozhenfu@minimaxi.com>
Co-authored-by: BBuf <1182563586@qq.com>
Co-authored-by: andyluo7 <andy.luo@amd.com>
Co-authored-by: Zijie Xia <zijie_xia@icloud.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: chao-xue <877184285@qq.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

This branch was successfully deployed

1 active deployment
staging - docs_new — f589dfdc Deployed Aug 2, 2026 by mintlify[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

diffusion SGLang Diffusion documentation Improvements or additions to documentation jit-kernel Multi-modal multi-modal language model quant LLM Quantization run-ci CI: run the baseline test suite on this PR

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant