# Docker Hub publication and portable commands

Final end-to-end command testing is complete on the MI50 host (`.20`) and V620 host (`.50`). All 48 executable configurations were exercised: 191 command tests succeeded and one benchmark reproduced the documented V620 FP16 TP2/MTP capacity limitation. All 48 serving commands completed actual inference. This is the owner-authorized public review package: [open the chart](https://localaiservers.com/benchmarks/qwen-mi50-v620-20261110/qwen-mi50-v620-release-chart.html). See the chart's separate Docker Hub command-validation column, [coverage](dockerhub-e2e/coverage.json), and [final command review](dockerhub-e2e/final-review.json).

The 47 completed ladders produced 42 strict passes and five strict failures; failed-strict speed figures remain diagnostic. The capacity-blocked row has no strict verdict or TPS. Three V620 TP8 placeholders remain unmeasured and have no executable command. Original recorded results are preserved separately from these fresh validation results.

The 20 original recorded image IDs and five derived MI50 host-portability images are available from [joe2gaan/localaiservers](https://hub.docker.com/r/joe2gaan/localaiservers). The current portable commands reference 21 distinct runtime images across 48 executable configurations. Each row pins an immutable registry manifest digest. Original image bytes and all new layers/configs were checked against fresh registry downloads. Publication and byte verification do not by themselves establish a completed GPU test.

Start with [PORTABLE-LAUNCH.md](PORTABLE-LAUNCH.md). The portable commands take explicit model paths, fresh work paths and GPU render nodes. The original measurements, recovered launch sources and retained lab commands remain available as separate evidence. The local Docker archives are optional offline backups; the new launch commands pull from Docker Hub and do not need them.

[Original credential and portability audit](credential-portability-audit.json) · [Original registry byte verification](registry-content-verification.json) · [MI50 compatibility change](mi50-portable-compatibility.json) · [New-image credential scan](mi50-portable-credential-scan.json) · [New-image registry verification](mi50-portable-registry-verification.json).

[Completed-result and final-asset credential review](dockerhub-e2e/asset-credential-review-r3.json) accounts for every scanner finding in the added evidence and final chart snapshots. No project credentials were found.

The original two MI50 services were restored with unchanged identities/configurations and verified with first and warm inference requests: [restoration verification](dockerhub-e2e/restoration-review.json). [File and command checks](chart-dockerhub-validation.json) and [browser checks](chart-browser-check-dockerhub-final.json) cover the final local chart. The owner authorized website publication on 2026-11-10; the retained validation reports describe their original local test dates.

The [HTML chart](qwen-mi50-v620-release-chart.html), [Excel workbook](qwen-mi50-v620-release-chart.xlsx), [CSV](qwen-mi50-v620-release-chart.csv), and [exact JSON](qwen-mi50-v620-release-chart.json) retain the same 51 result rows. Historical measurements and their fresh reproductions stay separate. Failed-strict and capacity-limited entries are labeled; no new successful result is implied by publication.

Each executable row includes an immutable image pull reference, a portable serving command, a benchmark command, pinned model acquisition, and retained lab launch commands and source bindings as references. The current portable launch records its actual fully resolved Docker argument array; the historical sources do not supply a complete original Docker inspection for every old measurement. Download the package named in that exact command: [r2](launch-tools-r2.tar.gz) / [checksum](launch-tools-r2.tar.gz.sha256), [MI50 r3](launch-tools-mi50-r3.tar.gz) / [checksum](launch-tools-mi50-r3.tar.gz.sha256), [V620 benchmark r4](launch-tools-v620-r4.tar.gz) / [checksum](launch-tools-v620-r4.tar.gz.sha256), [V620 serving r5](launch-tools-v620-serve-r5.tar.gz) / [checksum](launch-tools-v620-serve-r5.tar.gz.sha256), or [historical r6](launch-tools-historical-r6.tar.gz) / [checksum](launch-tools-historical-r6.tar.gz.sha256). Do not interchange packages. They write the full resolved Docker argument array and command into the chosen fresh run directory. All 19 historical commands were tested individually in this campaign.

Twenty detailed records exceed Excel's 32,767-character cell limit. They are preserved without truncation in the workbook's **Full JSON records** sheet, split into ordered parts, and in downloadable JSON assets. The Exact data sheet links internally to those complete records, so the workbook works independently of a local asset folder. CSV and JSON also contain each complete value.

Model-download availability was checked for all 129 pinned public assets: [verification](public-model-url-verification.json). Model weights are not embedded in the Docker images.

Raw benchmark evidence, original launch captures, source identities, and all previous failure records remain linked from the chart. The optional offline Docker archives remain in the local export; they are not required for Docker Hub launches.
