[Video] Package H3 execution capabilities for Studio FA and explicit VSA

[yangzhuxinyzx]

Purpose

Package H3's existing exact row reduction with the other SM70 extensions and expose a CPU-safe capability descriptor for Studio. Supported TP4 V100 deployments can select column-major floating projections, query-tiled FA, peer residual rows, pageable host masters and shared VAEs without first-request CUDA compilation. Agree on available peer memory and accessibility across ranks; fall back to native FP32 reduction with a recorded reason when unsuitable.

The descriptor separately advertises the sparse binary inherited from #583 for explicit FastH3 VSA selection. Native CLI/AUTO defaults remain compatible. Original floating FL2VA and official Dense/VSA Data-Free adapters are required; this does not repurpose INT8 ConvRot or LightX2V models.

Test Plan

Check source/wheel extension registration, CPU capability probing and peer setup fallbacks, build the actual remote extensions, run focused sparse GPU checks, and compare the two explicit modes through real Studio frontend submissions on the allocated TP4 group.

Test Result

• 84 focused CPU tests passed (11 GPU skips on the CPU run).
• Five H3/SM70 extensions built locally and on the remote V100 host.
• Remote nine sparse GPU tests and three Studio-capability checks passed.
• Local pre-commit passed; GitHub CI runs against the final head.
• Real UI campaign, downloaded media and stage accounting are documented in docs/design/minimax_h3/STUDIO_FASTPATH.md.

Remote frontend evidence

On the allocated TP4 V100 SXM2 32 GB host (Torch 2.10.0+cu128, CUDA 12.8, 62 GiB host RAM, 300 W limits; no GPU settings changed), real UI submissions used original FL2VA, matched prompt/seed 42, 1280×736, 120 requested frames, 24 FPS, four actual denoiser intervals, and each backend's official FastH3 Data-Free adapter. One excluded warmup + three measured requests per mode.

Median	FA / Dense	Fast / VSA	Speedup
Denoise	51.58 s	29.01 s	1.78×
Full native generation	122.28 s	97.20 s	1.26×
Studio submit through saved result	124.25 s	99.71 s	1.25×

All eight outputs passed browser playback/download and complete FFmpeg decode with 124 aligned frames and audio. All ranks reported the requested backend and four DiT calls; peer residual route did not fall back. Measured MP4s were bitwise-repeatable within each mode. Adapter changes require reloading (VSA cold preparation: 944.26 s), which is excluded from warmed medians and shown separately in Studio. Tested native source: 5e50ef0df7; later native changes are documentation only. Detailed run IDs and report digest are in the checked-in implementation report.

VSA still fails the independent FP32 quality gate inherited from #583. This is explicit experimental execution, with no AUTO/default quality promotion and no promise of halving complete generation time. Dense and VSA use different official adapters; this is not output-equivalence evidence or an INT8/LightX2V comparison.