acceptance: MiniMax-H3 T2VA lossless latency on four L20X GPUs #438

Activity

1. amy-why-3459 commented on Nov 14, 2026

   

   amy-why-3459

   on Nov 14, 2026 – with Cursor

   Collaborator

   More actions

   Thanks for opening this. Before a candidate is charged against it, here are a few points I think the acceptance contract should pin down:

   1. Owner and roadmap row. Sibling campaigns, e.g. evidence/h3-ref2va-910c-4logical-20261101-297/README.md, record an accountable owner, a claim/plan receipt and who makes the final acceptance decision. Please add those here. As far as I can see, the [Roadmap] vLLM-Omni Model Optimization #416 tables don't have an H3 T2VA × L20X row yet, so this task should be linked there too.

   2. L3 feasibility first. Byte identity of decoded video and audio for all 20 prompt/seed pairs assumes the baseline itself is bit-reproducible on L20X. Please record a baseline-vs-baseline run with two independent server starts before any candidate is judged. Otherwise a no-op candidate could already fail L3. The decoder and its version, the pixel and sample formats, and the stream selection should be fixed by the pinned validator (673a1fd1…), not left to the environment.

   3. L4 measurement protocol. "≥10% lower arithmetic mean over 20 valid requests" needs:

      • the warmup count;
      • whether baseline and candidate are interleaved or A/B/A on the same node and session;
      • how many repeat rounds are run;
      • whether a confidence bound or a noise floor from repeated baseline runs is required beyond the point estimate.

      Please also define exactly where "submit" and "downloaded MP4" are timestamped (client clock, including transfer and write), and what "matched storage conditions" means (same mount or tmpfs, same output path).

   4. Execution identity. Please record the GPU topology (PCIe or NVLink), the driver, CUDA and PyTorch versions, and the baseline's parallel layout across the four cards (TP/SP/CFG split). One-plane-per-iteration changes are only comparable if that layout is fixed, or is itself the declared plane.

   5. Workload pinning. Please state the resolution, duration, FPS and frame count that h3-v2 release-v1 implies for this task, and list the 20 prompt/seed pairs or their digest in plan/, so that L3 and L4 use the same set.

   Agreed that the initial verdict is INCONCLUSIVE, since there is no qualified candidate or gain yet.
2. Dong1017

   added

   area:harnessHarness module: profile, A/B loop, quality gates, evidence reportsHarness module: profile, A/B loop, quality gates, evidence reports

   kind:acceptanceCampaign acceptance issue: frozen gates + evidence tree; PASS/FAIL/INCONCLUSIVECampaign acceptance issue: frozen gates + evidence tree; PASS/FAIL/INCONCLUSIVE

   status:proposedWaiting for the decision named in the status blockWaiting for the decision named in the status block

   on Nov 15, 2026
3. Dong1017 commented on Nov 15, 2026

   

   Dong1017

   on Nov 15, 2026

   Collaborator

   More actions

   Dong-bot first triage (steward).

   Classification: kind:acceptance + area:harness + status:proposed. This is a campaign acceptance Issue under parent #416 (MiniMax-H3 T2VA × four L20X), not yet accepted for candidate search.

   What is already clear: BF16 checkpoint / 49 denoise / dense lossless / C1; pinned baseline, InferenceX and model revisions; L4 mean-latency ≥10% bar; L3 decoded A/V byte identity with named validator SHA-256; evidence path evidence/h3-t2va-l20x-lossless/; initial verdict INCONCLUSIVE.

   Before accept / charge a candidate (agreeing with Amy’s open pins):

   1. Name accountable owner, claim/plan receipt, and who makes the final acceptance decision; link a row in [Roadmap] vLLM-Omni Model Optimization #416 if missing.
   2. Confirm L3 byte-identity feasibility against the frozen baseline before treating it as a hard gate.
   3. Add the three-line status block (Current state / Decision needed / Next action).

   Next action: Author (or campaign owner) answer Amy’s contract pins and add the status block. Steward will re-read for status:accepted only after those are pinned; no candidate charged against this Issue until then.