Repository navigation
feat(ansible): download simpl_prediction, ptv3, lidar_frnet, calibration_status_classifier from Hugging Face - #7224
Conversation
…ion_status_classifier from Hugging Face Replace the per-file get_url tasks of the four batch A models with single pinned hf download tasks, following the lidar_centerpoint pattern. Each Hugging Face model repo was verified byte-identical to the previously pinned files, and each task pins a version tag. Part of #7223. Signed-off-by: Mete Fatih Cırıt <mfc@autoware.org>
|
Thank you for contributing to the Autoware project! 🚧 If your pull request is in progress, switch it to draft mode. Please ensure:
|
|
The migration task was mostly automated but I verified at many intermediary stages to ensure the task is completed correctly. I did a self-review and confirmed all the replaced files have the correct sha signatures. And this PR should not change the behavior of anything since it just pulls in the same files. Also ansible role is tested by the CI: https://github.com/autowarefoundation/autoware/actions/runs/29767437404/job/88441050627?pr=7224#step:5:313 @YoshiRi @mitsudome-r @ktro2828 @amadeuszsz please review 🙏 |
Address review feedback: the version string was duplicated between the task name and the --revision flag, which is error prone to update. Define it as a task-level revision var and template both places. Applied to all five hf download tasks, including the pre-existing lidar_centerpoint one, to keep the file uniform. Signed-off-by: Mete Fatih Cırıt <mfc@autoware.org>
…acts Templating the injected top-level ansible_env fact emits an INJECT_FACTS_AS_VARS deprecation warning on every hf download task (once for the PATH environment and once through the data_dir default), and the injection will be removed in ansible-core 2.24. Switch the artifacts role to the recommended ansible_facts.env form. Signed-off-by: Mete Fatih Cırıt <mfc@autoware.org>
|
Also here d208d7e I applied this pattern:
To conform to this migration guide: https://docs.ansible.com/projects/ansible/latest/porting_guides/porting_guide_13.html#inject-facts-as-vars (This is not an issue for the ansible 10 that we use but on my machine I have latest ansible 14 installed) |
youtalk
left a comment
There was a problem hiding this comment.
I've reproduced the ansible scripts and it works correctly. There is no difference from the gdown version.
LGTM
|
Tick the box to add this pull request to the merge queue (same as
|
…ion_status_classifier from Hugging Face (autowarefoundation#7224) * feat(ansible): download simpl_prediction, ptv3, lidar_frnet, calibration_status_classifier from Hugging Face Replace the per-file get_url tasks of the four batch A models with single pinned hf download tasks, following the lidar_centerpoint pattern. Each Hugging Face model repo was verified byte-identical to the previously pinned files, and each task pins a version tag. Part of autowarefoundation#7223. Signed-off-by: Mete Fatih Cırıt <mfc@autoware.org>
Description
Batch A of the ML artifact hosting migration to Hugging Face: replaces the per-file
get_urltasks ofsimpl_prediction,ptv3,lidar_frnet, andcalibration_status_classifierwith single pinnedhf download --revision <tag>tasks, following the lidar_centerpoint pattern. Also updates the role README's checksum sentence to describe both download mechanisms.v0.1perception/models/simpl/v0.1.0v4.0perception/models/ptv3/v4v2.0perception/models/frnet/v2v2.0sensing/models/calibration_status_classifier/v2Verification
checksum:entries, both before upload and via round-trip download after upload.hf downloadcommands in this diff were run locally; all 13 downloaded model and param files hash-match the recorded values.run:health-checklabel runs the full artifacts playbook in CI withHF_TOKEN.Notes for reviewers: like the lidar_centerpoint task, the hf tasks run without
become(files are user-owned instead of root-owned; consumers only read them), andhf downloadadditionally places the model card anddeploy_metadata.yamlin each model directory.