|
| 1 | +# Cam / Alec — H3 interface handoff (2026-09-28) |
| 2 | + |
| 3 | +Decision-oriented status after reconciling InferenceX [#2916](https://github.com/SemiAnalysisAI/InferenceX/pull/2916) to current `main` and checking App [#1193](https://github.com/SemiAnalysisAI/InferenceX-app/pull/1193) against that contract. No third product PR. |
| 4 | + |
| 5 | +## SHAs |
| 6 | + |
| 7 | +| Side | Branch | Head | Base | |
| 8 | +| --- | --- | --- | --- | |
| 9 | +| Backend #2916 | `feat/h3-cross-hardware` (repair also on `cursor/h3-main-reconcile-7116`) | `1e648b146fc1dede7e4758392787fd8b1f1de30c` | `main` @ `e0127239f073f3cd64f0732acdff62354af7b49a` | |
| 10 | +| App #1193 | `feat/videogenx-hardware-dashboard` | `ac73782b325d3b5a7764243d6a8aac2f7434663a` | `master` (unchanged this pass) | |
| 11 | + |
| 12 | +#2916 is **MERGEABLE** vs current main (was CONFLICTING). Still draft / blocked on reviews and required checks — not GPU-qualified. |
| 13 | + |
| 14 | +## Contract matrix (what is compatible now) |
| 15 | + |
| 16 | +| Surface | Backend #2916 after reconcile | App #1193 consumer | Verdict | |
| 17 | +| --- | --- | --- | --- | |
| 18 | +| **YAML / planner** | Empty `config-keys` + `workflow-dispatch: h3-video.yml`; planner validates workflow then selects **no** LLM jobs | N/A (App does not plan sweeps) | Compatible | |
| 19 | +| **Runner / paths** | Workflows, `PYTHONPATH`, and staged power packaging read `inferencex-e2e/`; staged package still exposes `infx/...` | Local demo still serves extracted CI bundles (`scripts/serve-h3-artifact.py`) | Compatible; path fix was backend-only | |
| 20 | +| **Results** | `result.json` `schema_version: "1.0.0"`, `bundle_type: h3_benchmark_result` ([RESULTS.md](./RESULTS.md), [result.schema.json](./result.schema.json)) | `bundle.ts` requires exactly `1.0.0` / `h3_benchmark_result`; history projection is additive nullable | Compatible — no App edit required for this repair | |
| 21 | +| **Media** | Original media + SHA256 inventory retained; export/fidelity workflows unchanged in behavior | Compare / `h3-media` publish path read published blobs; chart uses history projection, not raw bundles | Compatible; publication/warm path not re-run here | |
| 22 | +| **Async completion** | Per-request outcomes (`completed` / `invalid_media` / timeouts / etc.), job IDs, submit→terminal timings | History keeps scheduled / completed / failed / valid counts separate; nulls stay null | Compatible | |
| 23 | +| **AMD** | Prep/inventory/serving CI paths + Python 3.10 staged power; generation still gated | Roster keeps MI355X as **Not measured** | Compatible as “prep-only / not measured”; no AMD generation claim | |
| 24 | + |
| 25 | +## What still needs qualification (do not claim) |
| 26 | + |
| 27 | +1. **GPU execution on the reconciled head** — no new allocation this pass; prior CPU green at old head does not qualify this tip. |
| 28 | +2. **AMD generation smoke** — staging/prep only; do not treat MI355X as measured. |
| 29 | +3. **Healthy H100 replacement / A/A noise floor / AMD monitor trust** — still unproven for article claims. |
| 30 | +4. **Native H3 SRT adapter** — separate migration slice; not in #2916/#1193 closure. |
| 31 | +5. **App remote Unit/E2E** — #1193 still draft / blocked; checks were skipped, not passed. |
| 32 | + |
| 33 | +## Recovered plateau → deployment-scaling story (do not re-open as “unknown”) |
| 34 | + |
| 35 | +Later evidence already supersedes “first discover deployment tradeoffs”: |
| 36 | + |
| 37 | +- Retained batch-one / single-replica cells: C1/C2/C4 throughput within about ±1.5% while latency grows with concurrency → concurrency mostly measures queueing on that layout ([App videogenx-dashboard.md](https://github.com/SemiAnalysisAI/InferenceX-app/blob/feat/videogenx-hardware-dashboard/docs/videogenx-dashboard.md)). |
| 38 | +- App already keys the curve by **deployment** (`participating GPUs` × `tp_size` × `ulysses_degree`), treats `concurrency > replicas` as queueing, and keeps queued cells out of the chart. |
| 39 | +- Fixture still has **one** measured deployment per measured SKU (4 GPUs / TP2×Ulysses2). Sealed 2-GPU and H200 4/8-GPU campaign evidence exists in the Sept 26 work package; the remaining engineering sweep is the authorized GPUs-per-video matrix on a new backend head — not another plateau hunt. |
| 40 | + |
| 41 | +## App #1193 compatibility verdict |
| 42 | + |
| 43 | +**Code + fixture read: compatible with repaired #2916.** No mandatory App patch for the `inferencex-e2e/` path realignment or the preserved `result.json` 1.0.0 contract. |
| 44 | + |
| 45 | +Unverified here (needs App owner / browser when UI changes): live blob publication, Cypress E2E against non-fixture backends, and any UI polish still open on #1193. Prefer App-owner fixes inside #1193 over a third PR. |
| 46 | + |
| 47 | +## Next concrete decision (pick one) |
| 48 | + |
| 49 | +1. **Merge-ready backend review of #2916** at `1e648b14` (CPU contract green; still draft; GPU still outstanding), then App #1193 review/merge sequence; **or** |
| 50 | +2. **Authorize one GPU qualification on this head** — prefer a single H200 C1 (or one GPUs-per-video cell) before any broad campaign; **or** |
| 51 | +3. **Defer GPU** and only land the planner/workflow foundation, keeping article claims inside already retained deployment-scaling evidence. |
| 52 | + |
| 53 | +Recommendation: (1) if the goal is to unblock review of the existing drafts; schedule (2) only when someone will own the exact cell and acceptance receipt. |
| 54 | + |
| 55 | +## Validation run this pass (CPU only) |
| 56 | + |
| 57 | +```text |
| 58 | +# H3 harness (matches test-h3-video.yml) |
| 59 | +cd experimental/video-generation |
| 60 | +PYTHONPATH=$PWD/../../inferencex-e2e |
| 61 | +uv run --no-project --python 3.12 \ |
| 62 | + --with 'av==16.1.0' --with 'numpy==2.3.5' \ |
| 63 | + --with 'pytest>=8,<9' --with 'jsonschema>=4,<5' python -m pytest -q |
| 64 | +# → 661 passed |
| 65 | +
|
| 66 | +# Staged power on retained runtime Python |
| 67 | +uv run --no-project --python 3.10 --with 'pytest==8.4.2' python -m pytest -q \ |
| 68 | + experimental/video-generation/tests/test_ci.py::test_staged_package_recomputes_power_without_the_repository |
| 69 | +# → 1 passed |
| 70 | +
|
| 71 | +# Planner workflow-dispatch contract |
| 72 | +cd inferencex-e2e && uv run pytest -q \ |
| 73 | + infx/tests/matrix/test_process_changelog.py -k 'workflow_dispatch or manual' \ |
| 74 | + infx/tests/matrix/test_validation.py -k 'workflow_dispatch or manual' \ |
| 75 | + infx/tests/test_installed_package.py -k 'workflow_dispatch or manual' |
| 76 | +# → 18 passed |
| 77 | +``` |
| 78 | + |
| 79 | +## Owner notes |
| 80 | + |
| 81 | +- Repair commits authored/committed as Wenyao Gao `<wenyao.gao28@gmail.com>`. |
| 82 | +- Sibling repair branch `cursor/h3-main-reconcile-7116` matches `feat/h3-cross-hardware`; do not open a third H3 product PR. |
| 83 | +- No Slack/email/public review comments; no GPU dispatch; no claim of GPU qualification. |
0 commit comments