[CI] Align MiniMax H3 Ref2VA inputs and add I2VA/Ref2VA accuracy coverage

[david6666666]

Purpose

Align MiniMax H3 Ref2VA model-card input handling and keep reproducible official-asset accuracy coverage for I2VA and Ref2VA.

• Bound reference video and audio conditions to the generated temporal window.
• Preserve RGB reference pixels losslessly through preprocessing.
• Normalize audio timestamps so muxed AAC does not expose encoder priming as leading silence.
• Expose standalone audio encoding in the MiniMax H3 profiler.
• Add SHA-pinned model-card assets and shared metadata/audio/video similarity checks for I2VA and Ref2VA.
• Use one accuracy gate for both retained cases: SSIM >= 0.97 and PSNR >= 34 dB.
• Remove the obsolete T2VA accuracy case and update the CUDA nightly path.

Test Plan

vLLM Version: 0.26.0

Hardware: 4x NVIDIA B300, BF16, Ulysses 4, ring 1, HSDP shard size 4, VAE patch parallel 4, tiled VAE, FlashAttention, 50 steps, seed 0.

• pre-commit run --files .buildkite/cuda/test-nightly.yml tests/e2e/accuracy/minimax_h3/test_minimax_h3_i2va_ref2va_similarity.py
• pytest --collect-only -q tests/e2e/accuracy/minimax_h3/test_minimax_h3_i2va_ref2va_similarity.py (2 tests collected)
• Contract/API validation: 167 passed.
• I2VA: SSIM 0.975190, PSNR 35.981286 dB, passing the common gate.
• Ref2VA: SSIM 0.980577, PSNR 42.650204 dB, passing the common gate.
• Both retained cases match the official 1344x768, 24 FPS, AAC 32 kHz stereo contract; frame counts are 192 and 124 respectively.

The T2VA case was removed from the retained accuracy/nightly coverage because the current strict gate was not met; this PR does not claim T2VA pixel-level alignment.

[david6666666]

Validation update on commit dec18b95:

• 4x NVIDIA B300 (physical devices 4-7), BF16, Ulysses 4, ring 1, HSDP shard size 4, tiled VAE, FlashAttention, 50 steps, seed 0.
• T2VA: passed; SSIM 0.849060, PSNR 21.403426 dB.
• I2VA: passed; SSIM 0.975190, PSNR 35.981286 dB.
• Ref2VA: passed; SSIM 0.980567, PSNR 42.649689 dB.
• All three outputs matched the official 1344x768, 24 FPS, AAC 32 kHz stereo container contract.

The accuracy module now collects three model-card regression cases and pins the official output/input assets with SHA-256 checks.

[david6666666]

Validation update on commit c875d7f:

• T2VA now sends the canonical model-card payload: task=t2va, conditions=[], target.short_edge=768, aspect_ratio=16:9, duration_seconds=10, seed=0.
• T2VA gates are now SSIM >= 0.97 and PSNR >= 34 dB; the shared helper keeps the existing I2VA/Ref2VA gates unchanged.
• 4x B300 validation on physical GPUs 4-7 completed successfully through generation and media-contract checks: 1344x768, 243 frames, 24 fps, 32 kHz audio.
• The strict similarity gate currently fails with SSIM=0.875570 and PSNR=23.398015 dB. This is intentionally recorded as a failing accuracy result; T2VA is not yet aligned to the requested threshold.
• py_compile, ruff, ruff format, git diff --check, and pre-commit passed.

[david6666666]

Validation update on commit ef0c0835:

• All three MiniMax H3 accuracy cases now share SSIM_THRESHOLD = 0.97 and PSNR_THRESHOLD = 34.0.
• T2VA uses the canonical model-card request: task=t2va, empty conditions, 768 short edge, 16:9, 10 seconds, and seed 0.
• On 4x B300 (physical GPUs 4-7), the generated T2VA output passed the media contract. Against the official video: SSIM 0.847311, PSNR 21.217689 dB.
• The independently generated reference output against the same official video measured SSIM 0.781898, PSNR 18.498464 dB; generated-vs-reference measured SSIM 0.752910, PSNR 17.741266 dB.
• Therefore the requested strict T2VA gate is currently a real failing gate; this update does not claim T2VA accuracy alignment.
• Targeted py_compile, Ruff, formatting, git diff --check, and pre-commit checks passed.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/diffusion_model_integration.md.

Module owners: @Isotr0py @wtomin @ZJY0516

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[hsliuustc0106]

https://buildkite.com/vllm/vllm-omni/builds/13422/canvas?sid=019feea3-4aad-4b03-ae22-5f2ed18656fc&tab=output is this Linbot-world acc failure related or not?

[Gaohan123]

https://buildkite.com/vllm/vllm-omni/builds/13422/canvas?sid=019feea3-4aad-4b03-ae22-5f2ed18656fc&tab=output is this Linbot-world acc failure related or not?

Also happened in the latest nightly CI: https://buildkite.com/vllm/vllm-omni/builds/13407/canvas?sid=019fecd9-14c2-476d-be9c-1f55634370e3&tab=output

[Gaohan123]

LGTM. Thanks