Fix CVE's for caikit-nlp #439
Conversation
rpancham
requested review from
evaline-ju,
gabe-l-hart and
gkumbhat
as code owners
Signed-off-by: rpancham <[email protected]>
gabe-l-hart
left a comment
gabe-l-hart
left a comment
There was a problem hiding this comment.
We'd like to keep support for older torch versions possible while also removing the upper bound. Assuming tests pass with the latest torch, this should be merge ready once we restore the known-working lower bound of 2.3.1.
| "sentence-transformers>=3.4.0,<3.5.0", | ||
| "tokenizers>=0.20.0", | ||
| "torch>=2.3.1,<2.9.0", | ||
| "torch>=2.8.0", |
There was a problem hiding this comment.
In general, we try to avoid requiring a lower bound like this since it disallows users that are stuck on older versions for one reason or another. I think the correct thing to do here is to simply remove the upper bound:
| "torch>=2.8.0", | |
| "torch>=2.3.1", |
There was a problem hiding this comment.
@rpancham @gabe-l-hart is this change actually needed, or do we just need to explicitly build an image somewhere with torch 2.8?
I followed the CVEs, and it seems to be patched in 2.8+, but that should already be compatible with things as they are, right?
GHSA-887c-mr87-cxwp
GHSA-3749-ghw9-m3mg
There was a problem hiding this comment.
As per my understanding, the CVE fix for PyTorch (GHSA-887c-mr87-cxwp) is included in torch 2.8.0+, so ensuring our image/tests use a patched version is sufficient from a security perspective. Because of that, we don’t need to keep the upper bound in pyproject.toml.
If the goal is to maintain compatibility with older torch versions, I can update the dependency to:
"torch>=2.3.1",
and remove the upper bound as suggested.
Please let me know if you'd like me to make this change — I can update the PR and sign off accordingly.
There was a problem hiding this comment.
Yes, let's go ahead and simply remove the upper bound. If CI passes, this will be a good additional piece of flexibility for users and will enable forward-compatible CVE fixes.
There was a problem hiding this comment.
@rpancham Have you had a chance to circle back on this change?
6 participants
Addresses: https://issues.redhat.com/browse/RHOAIENG-37431