[Video] Integrate H3 SM70 execution, metrics and VSA diagnostics

[yangzhuxinyzx]

Purpose

Integrate the H3 SM70 execution and VSA stack into main through this existing PR, including the already-contained dependencies #571, #578 and #581. Original floating and W8A16 weights share prepared FP16 projection and adapter-aware execution; explicit residual sharding/peer reduction, layer offload and shared VAE host storage extend the measured native workflows. Workflow-aware counters report actual sampling work, sparse computation, per-step timing and source/kernel provenance.

FastH3 VSA retains the official four-step Data-Free algorithm, top-k=64, prefix-dense queries, per-query video blocks, boundary handling and learned compression. Request-owned layout buffers and deferred GPU counters remove synchronization and repeated allocation without changing the native result. Public validation and integer statistics remain intact. The independent CUTLASS SIMT FP32 implementation is an acceptance-only diagnostic with no runtime/default/AUTO registration. A dedicated evaluator checks the 31.3-second stage separately from numerical/human quality and the future 80-TF target.

The final integration head is 1c39c173116269bf69493941da36f7c7f9dc281d; its runtime/tests match the tested d6669c4ce1a915d59d4baaa258a437ccc03268b0 exactly and it additionally includes the final TP2 dependency evidence. The synchronization with main at 24220ca0eb4a02b2376cf48feb430bb4d5c5c3d7 retains its mmap host weights, media progress and physical-device masking alongside the H3 changes. The merge itself changes no numerical CUDA kernel. The user authorized source integration on 2026-09-10; this does not mark unfinished acceptance gates as passed. The separate cache implementation in #584 is not included.

Test Plan

• Validate the combined main/H3 configuration, API, host-storage and staging behavior; exercise mmap with layer offload, adapter increments and aliased storage on V100.
• Retain sparse boundary/padding/input-rejection, exact layout, deferred accounting and memory/synchronization checks.
• Compare the primary fixed-noise full sampling with independent FP32 mathematics, then measure one full warmup and three unprofiled TP4 requests under the documented workload contract.

Test Result

• Synchronized candidate: full CPU video suite 359 passed, 251 GPU/opt-in skipped. Focused leased-V100 integration suite 73 passed, plus 24 QK/RoPE GPU checks passed. Ten existing GPU-only cases now skip explicitly on hosts without CUDA. All applicable local pre-commit checks pass.
• Retained sparse/layout evidence: 23 counter/operator GPU checks and 30 layout GPU checks pass. Nine sparse and eight layout cases pass both memcheck and synccheck with zero errors. Full primary native output and rank/layer/step accounting are preserved bitwise.
• Acceptance-only exact diagnostic: 20 GPU checks pass, including query-batch boundaries; all pass memcheck and synccheck with zero errors. The full 609-block operator input, initial sampling inputs, all four steps, final video/audio latents and all 124 decoded RGB frames match the frozen independent FP32 reference bitwise. Audio spectral cosine is 0.999999999999945 and RMS ratio 0.999999998643185.
• Native FP16 VSA formal denoise median 30.990756 s, CV 0.071956%, full-request median 68.157817 s versus retained matched Dense 87.426512 s. It passes timing only and fails independent FP32 quality.
• Exact FP32 diagnostic formal denoise median 60.353224 s, CV 0.028716%, full-request median 140.666471 s, lowest-rank median effective throughput 28.042311 TFLOP/s/card, peak allocation bound 23.173655 GiB/card. Primary numerical gates pass; denoise and request-speed targets fail. Host staging/encoding remained unsettled in the first two requests, and cross-campaign background host state was uncontrolled, so the request difference is not attributed entirely to sparse arithmetic.

No configuration passes the combined VSA quality/speed stage. Seeds 43/44, 243-frame and 15-second boundaries, TP1/TP2 and representative Dense/LightX2V/Ref2VA full-output regressions remain incomplete for this VSA stage. Official original-hardware comparison is deferred. Human audiovisual review is pending, including object-count consistency also present in the identical reference. No default/AUTO promotion is included.

Reproduction commands, exact source/binary identities, playable candidate/reference MP4/WAV paths and remaining bottlenecks are recorded in docs/design/minimax_h3/VSA_QUALITY_SPEED.md. Both retained MP4s pass complete FFmpeg decode. Merge validation artifacts are in the documented vsa-merge-main-20260910 directory. Historical performance remains tied to its original measured source and binaries; integration tests do not establish new performance or quality acceptance.

Assisted-by: OpenAI Codex