[Campaign] MiniMax-H3 Ref2VA on 910C: 4 logical devices, 5s/15s × lossless/lossy optimization #297

Current state: Draft #317, inspected head 6aef3438, publishes an author-reported campaign checkpoint, superseding the preflight-only status below. Overall verdict remains INCONCLUSIVE: one clean 5s lossless screening pair reports 1.57% lower complete-HTTP latency, below the 50% target; 15s clean timing and four-track final confirmation remain incomplete, and tested fast lossy profiles retain quality failures. This is evidence delivery, not campaign acceptance.
Decision needed: none
Next action: The existing campaign owner coordinates independent review of the published checkpoint and dispositions the remaining four-track evidence with the requester. The PR reports its direction ceiling exhausted; further search or final confirmation requires valid requester authorization and budget. Original preflight decisions and historical evidence below are preserved; this update grants no execution authority or budget extension.

Goal and scope

Optimize MiniMax-H3 Ref2VA on 4 logical Ascend 910C devices, with reproducible performance and independent quality validation across four required tracks:

The common reference is the original BF16 dense model, 50 steps. Four logical devices correspond to two dual-chip physical cards in this preflight. The 15s workload is a four-logical-device extension; it does not satisfy the original eight-device requirement.

Parent campaign: #293. Keep this 910C scope and its results separate from the eight-device 910B1 campaign in #294. No inherited speedup, quality PASS, or global-optimality claim.

Current evidence

The owner-provided local preflight report records one same-configuration warmup and three measured requests per duration, timed from HTTP request to receipt of the complete MP4 response:

• 5s / 124 frames: 292.148 s median.
• 15s / 362 frames: 1,702.939 s median.

Media decoding and file-hash checks passed; the devices were released. The report includes environment repairs, stage timings, reproduction commands and evidence locations. These are reported preflight results; the underlying local artifacts have not been independently replayed here. No optimization A/B or lossless/lossy quality acceptance has been completed.

The old 5s prompt and reference image are semantically mismatched. Preserve those samples as performance/repeatability evidence and establish aligned inputs for formal quality comparisons. The two durations currently use different prompts, so their latency ratio is not a controlled length-scaling result. Private host identifiers, logs and paths are not published here.

Work and deliverables

• Freeze formal baselines. Pin inputs, model/source/dependency identities, parallelism, memory policy, resource allocation, timeout and timing boundaries. Establish semantically aligned quality baselines for both durations. Keep preflight and failed attempts.
• Freeze independent quality contracts. Specify media completeness, numerical checks, video/audio and reference/temporal consistency, scorer/weight identities and human evaluation. Measure baseline variability before judging candidates.
• Profile the bottlenecks. Collect steady-state DiT operator/shape, communication and offload evidence separately from timing samples. Report rank coverage and memory definitions. High DiT time alone does not identify the dominant operator.
• Evaluate individual candidates. Record hypothesis, source attribution, actual enabled path, code/config changes, quality and E2E effect. Separate implementation optimizations preserving the numerical contract from approximate caching, quantization and sparsity. Model replacement, distillation and step-count changes require explicit scope decisions.
• Validate all four final configurations. Compare each directly with its matched original baseline using a prospectively frozen independent-server, ordering, warmup and repeat protocol. Do not add speedups from separate experiments or infer tail latency from three samples. Preserve negative and inconclusive results.
• Deliver code and reusable evidence. Provide necessary engine changes and upstream relationships, reproduction instructions, sanitized manifests and sample references, quality reports, final configurations and rejected experiments. Separate software delivery from campaign acceptance.

Current coordination and recorded decisions

The existing campaign owner's claim and decision record reports the requester selections below. The original questions following this section are retained as design history; the selected options are no longer open alternatives. Detailed contracts and execution readiness remain pending.

• Ownership: Dong1017 is assigned; the Codex agent in the requester's H3 session coordinates the campaign. Reuse that work rather than creating a parallel implementation.
• Quality selections: strict bit-exact for lossless; reduced-vBench relative BF16 score error <1% for lossy, with no human checks during RSI. Exact comparison points, automated checks, baseline variability, inputs and validator identities still require freezing.
• Search and target: the claim records an initial ceiling of 12 candidate directions and 16 wall-hours / 64 logical-device-hours including failures and retries, and at least 50% matched-B0 E2E latency reduction per track. Final four-track confirmation requires a separately established budget.
• Preparation and next action: the existing owner reports the isolated control checkout and source-mode CLI/import checks complete. Distribution installation remains incomplete after bounded dependency/wheel-build timeouts; no installed-wheel qualification is claimed. Next: prepare separate engine environments and verify source/package identities, then establish aligned no-approximation B0 and freeze automated contracts. Candidate search still awaits exact gates, allocation window and budget accounting. No new model execution or search is reported; local draft evidence is not a published bundle or campaign acceptance.

Initial decision questions (superseded where recorded above)

1. Lossless definition: bit-exact at specified comparison points, or explicitly bounded numerical tolerance? Specify tensors/outputs, dtype, statistics and limits. Repeatable MP4 hashes alone do not establish whole-path equivalence.
2. Lossy gates: freeze video, audio, reference and temporal checks and allowed degradation. Confirm how the cited vBench relative-score requirement and ten-prompt human checks apply; thresholds are not yet accepted and are not inherited from 910B1.
3. Execution protocol: freeze resource reservation, run ceilings, independent-server repetitions and A/B ordering before observing candidate outcomes. The current one-warmup/three-measurement counts describe preflight, not an automatically accepted final statistical protocol.

Stop candidate promotion on quality failure, resource interference, request failure or incomplete cleanup. Report each track independently as supported, failed or inconclusive; missing checks remain open. Final acceptance belongs to the requester.

Upstream reuse assessment (proposed candidates)

Static inspection of vLLM-Omni #9698, head 922f1e9f, suggests useful candidates; none is qualified on this 910C campaign yet.

• Lossy tracks: Cache-DiT and cross-step reference-KV reuse require separate 5s/15s quality and performance checks. Device residency changes where KV is stored, not the approximation introduced by cross-step reuse. Compare host/device residency at the same refresh policy, then compare with the no-cache BF16 baseline. Do not add reported gains from different upstream experiments.
• Lossless track guard: quality=lossless disables Cache-DiT, while reference KV has a separate configuration entry. Audit actual enabled paths rather than relying on the request label.
• All four tracks: terminal DLO cleanup, worker shutdown and request-owned diagnostics are lifecycle/observability candidates. Verify normal, failed and cancelled requests, actual worker exits and repeat startup; longer shutdown budgets are not inference speedups.
• Migration first: select a source pin and establish a no-approximation compatibility baseline before candidate measurements. The preflight pin predates the new residency option. Verify all-rank HBM at USP4 and both durations, retain cache ownership protections, and use a common demonstrated steady-state warmup protocol.

References

• Parent: [Campaign] MiniMax-H3 Ref2VA Ascend NPU latency optimization with video/audio quality gates #293; 910C coordination
• Related 910B1 evidence: Qualify MiniMax-H3 Ref2VA five-second Ascend optimization #294
• Candidate upstream implementation: vLLM-Omni #9698, applicability to 910C to be assessed.