|
| 1 | +# Submission notes: `ascend-video-diffusion-optimization` |
| 2 | + |
| 3 | +InferRSI submission record for the packaged Ascend skill. |
| 4 | + |
| 5 | +## Packaged location |
| 6 | + |
| 7 | +```text |
| 8 | +src/inferrsi/data/skill_library/hardware/ascend/ascend-video-diffusion-optimization/ |
| 9 | +├── SKILL.md |
| 10 | +├── metadata.json |
| 11 | +├── assets/campaign.template.json |
| 12 | +├── references/field-notes-video-diffusion.md |
| 13 | +└── scripts/evaluate_target.py |
| 14 | +``` |
| 15 | + |
| 16 | +## Commit and PR |
| 17 | + |
| 18 | +```bash |
| 19 | +git checkout -b skill/ascend-video-diffusion-optimization |
| 20 | +git add src/inferrsi/data/skill_library/hardware/ascend/ascend-video-diffusion-optimization |
| 21 | +git add src/inferrsi/data/skill_library/skill-index.json |
| 22 | +git add tests/test_ascend_video_diffusion_skill.py |
| 23 | +git commit -s -m "skills: add goal-driven Ascend video diffusion optimization workflow" |
| 24 | +``` |
| 25 | + |
| 26 | +- DCO sign-off is required (`git commit -s`); do not invent a signing identity. |
| 27 | +- Expected CI lane: draft (lint and CPU smoke). No model downloads, GPU jobs, or |
| 28 | + campaign thresholds were added. |
| 29 | +- Suggested PR title: `skills: add goal-driven Ascend video diffusion optimization workflow` |
| 30 | + |
| 31 | +Suggested PR body: |
| 32 | + |
| 33 | +> Adds a portable workflow skill for latency / RTF / throughput / peak-HBM work on |
| 34 | +> Ascend video diffusion pipelines (condition encoders, DiT, MoE, attention, VAE). |
| 35 | +> The skill drives a measure-change-verify loop and ends in a verdict that follows |
| 36 | +> the repository claim contract. |
| 37 | +> |
| 38 | +> - `SKILL.md`: goal contract, stage baseline, optimization loop, candidate |
| 39 | +> selection, correctness protection, final acceptance. |
| 40 | +> - `references/field-notes-video-diffusion.md`: first-hand field notes |
| 41 | +> and rejected experiments. Marked unverified and non-reproducible. |
| 42 | +> - `scripts/evaluate_target.py`: optional local pre-check for target, improvement, |
| 43 | +> variance, quality and resource gates. Stdlib only, no accelerator dependency. |
| 44 | +> It is explicitly not the source of a public verdict. |
| 45 | +> - `assets/campaign.template.json`: schema template with synthetic values; real |
| 46 | +> gates come from the acceptance issue. |
| 47 | +> |
| 48 | +> No campaign revisions or numeric gates are included. No public performance claim is made. |
| 49 | +
|
| 50 | +## Contract alignment |
| 51 | + |
| 52 | +| Rule | How this package complies | |
| 53 | +| --- | --- | |
| 54 | +| No pins or thresholds in skills | No revisions, SHAs, campaign IDs, or gate values are committed; the template holds synthetic values only and points at the acceptance issue. | |
| 55 | +| Verdicts are `PASS` / `FAIL` / `INCONCLUSIVE` | The evaluator emits only those three, plus `INVALID` for unreadable input, which is documented as a tool error rather than a verdict. | |
| 56 | +| Microbench ≠ service gain | Section 6 requires the full workload, cold and warm measurement, and issue-scoped evidence. The evaluator's output carries a scope note that a public `PASS` also needs pinned revisions, frozen validators and a complete evidence tree. | |
| 57 | +| Validation levels L0–L4 | Section 1 links the repository's level definitions and evidence layout. | |
| 58 | +| One plane per step | Section 3 requires one declared change per iteration. | |
| 59 | +## Open questions for the maintainer |
| 60 | + |
| 61 | +1. `metadata.json` → `origin`. Every shipped example uses `"external"` for imported |
| 62 | + material. This skill is first-party, so `["first-party"]` was used; confirm the |
| 63 | + allowed values and whether the key should be omitted instead. |
| 64 | +2. Optional bilingual edition. The package is English-only, matching the existing |
| 65 | + `skills/` candidates. Say whether a `zh-CN` mirror under the engine skill library |
| 66 | + is also expected. |
| 67 | + |
| 68 | +## Local validation performed |
| 69 | + |
| 70 | +Run: |
| 71 | + |
| 72 | +```bash |
| 73 | +pytest -q tests/test_ascend_video_diffusion_skill.py |
| 74 | +``` |
| 75 | + |
| 76 | +The CPU-only tests verify catalog discovery, trigger metadata, packaged |
| 77 | +resources, relative links, and evaluator behavior for `PASS`, `FAIL`, |
| 78 | +`INCONCLUSIVE`, and malformed-input `INVALID` results. They do not establish |
| 79 | +model-level quality or end-to-end performance; those remain campaign evidence. |
0 commit comments