[Video] Measure H3 workflows with actual denoise schedules

[yangzhuxinyzx]

Purpose

The H3 acceptance evaluator hardcodes 49 denoiser calls and one shape, so it cannot validate official four/eight-step adapters or other legal workflows. Native requests now retain actual video/audio sigmas, selected attention backends, completed step/block counts and per-step CUDA/CPU timing. Throughput still divides each rank's useful, unpadded FLOPs by the slowest rank's full synchronized denoise time.

A native benchmark command performs one complete warmup plus three consecutive full requests on one leased GPU group. Evaluation rejects incomplete warmup, changed engine/configuration, missing blocks/steps, inconsistent work counts, profiler timings and quality-capture runs. Results retain source/kernel fingerprints, end-to-end latency, memory and existing NVML measurements. Sparse/cache variants are explicitly rejected until their real execution accounting is implemented.

Stacked on #571 (codex/v100-h3-general-sm70-20260908-165458), which owns common prepared execution and residual changes. Integration commit 5195f31b8d retains both tested parent histories; the PR diff contains the acceptance/measurement scope. The campaign root is onecat/main at 4f19ef7a20db60bb0685e599bd3f4dd156202eed. The open-PR search found no overlapping workflow-accounting PR.

Test Plan

• CPU: actual 49/4/8-call and DMD2 schedules, TP1/2/4, strict >80 and CV thresholds, completed same-session warmup, invalid/inconsistent execution records, native service/API regressions.
• GPU: independently count a real dense DiT block with suffix padding; verify event instrumentation preserves outputs and removes hooks after a request.
• Full model: compare instrumented 720p five-second four-step outputs to frozen mainline, then collect matching unprofiled FA/FI measurements.

Test Result

Python 3.12.13, Torch 2.10.0+cu128, V100 SXM2 32GB, task-owned binaries/caches and native GPU leases:

.venv/bin/python -m pytest tests/video/test_h3_acceptance.py tests/video/test_h3_service.py tests/video/test_h3_workflows.py tests/video/test_h3_omni_api.py -q
81 passed
.venv/bin/python -m pytest tests/video/test_h3_work_counter.py -q
1 GPU test passed: exact useful FLOPs, two steps, bitwise outputs, hook cleanup

All commit hooks pass, including mypy and ruff. The first GPU5 attempt was rejected by an existing lease before launching a test; the successful test used GPU0 after the owned generation released it. Raw evidence and exact launch environments: /data/minimax-h3/sm70-general-20260909/; see docs/design/minimax_h3/WORKFLOW_METRICS.md.

Full 720p five-second four-step instrumentation validation passed against frozen mainline: video/audio latents and all 124 pre-encoding frames match bitwise, video SSIM 1.0 and all audio numerical gates pass (metrics-quality.json).

FA completed one full warmup plus three full requests at source 82362a4312: denoise 65.880184 / 65.898529 / 65.965556 seconds, CV 0.055668%, peak allocation 19,501,498,880 bytes/card. An audit found the legacy counter included identical column-parallel LoRA A replicas. Commit 9bcde29ada excludes those replicas and reports them separately with balanced per-rank row ownership. Row-A/B partial products remain distinct work. Added CPU tests validate logical adapter work across TP1/2/4 and uneven tails; legacy/inconsistent counters are rejected.

The retained FA files are unchanged. fa-720p-three-runs/audited-counts/ records a header-shape audit with original JSON/source hashes and unchanged times: corrected median 47.091839–47.091855 useful TFLOP/s/card. This supersedes the overcounted 47.524847 figure. It is not a new run of the revised counter. Matching FI measurements completed on 9bcde29ada with the revised runtime counter: 70.625556 / 70.545638 / 70.525989 seconds, median 43.989720–43.989736 useful TFLOP/s/card, CV 0.061019%. All rank work counts match the audited FA records exactly. No >80 TFLOP/s, full coverage or independent official/human quality acceptance is claimed. Keep Draft pending those gates. GPU event spans include stream dependencies and host feeding gaps; separate Nsight Systems profiles now retain exact exclusive-wall closure and overlapping service attribution (fa-profile-breakdown/, fi-profile-breakdown/). FA attention/GEMM/communication consume 31.439/16.041/8.453 seconds; FI 36.119/15.919/8.527 seconds. These profiled runs are excluded from performance acceptance. The two backend full latents are not interchangeable under the 1% gate, despite small single-operator reference error; independent official quality remains required.

AI assistance: implemented and initially validated with OpenAI Codex. Human line-by-line review and independent validation are required before promotion.