1925282 - models and organizations allow/deny lists

Status: RESOLVED FIXED

(Core :: Machine Learning: General, enhancement)

Description

[Tarek Ziadé (:tarek)]

We will store on Remote Settings a list of Hugging Face models and organizations that are blessed to run in our inference engine without needing the MOZ_ALLOW_EXTERNAL_ML_HUB flag

Comment 1

[Tarek Ziadé (:tarek)]

Proposed mechanism: simple case insensitive pattern with URLs - we just match that every file we download match the allow list and is not in the deny list. This will be enough to handle all cases, from the rogue model or specific version, to blessing a full domain or organization

examples of patterns:

• ALL models ALL VERSIONS from the mozilla organization on hugging face : https://huggingface.co/Mozilla/
• ALL models ALL VERSIONS from our hub: https://model-hub.mozilla.org/
• A specific model ALL VERSIONS https://huggingface.co/typeform/distilbert-base-uncased-mnli/
• A specific model and a specific version https://huggingface.co/Mozilla/distilvit/blob/v0.5.0/

Each URL is tested and needs to be included in the allowlist and not in the denylist

The remote settings records have 3 fields:

• filter: ALLOW or DENY
• match: the url to match
• description

Comment 2

[Tarek Ziadé (:tarek)]

Attachment: WIP: Bug 1925282 - models and organizations allow/deny lists

Depends on D227327

Comment 3

[Tarek Ziadé (:tarek)]

Attachment: Bug 1925282 - models and organizations allow/deny lists r?rrando

Depends on D221606

Comment 4

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/9b85469b7aaa
models and organizations allow/deny lists r=rrando

Comment 5

[Norisz Fay [:noriszfay]]

https://hg.mozilla.org/mozilla-central/rev/9b85469b7aaa