Repository navigation
|
vLLM moves fast and breaks things. How do you pin it so a rebuild is byte-reproducible enough? |
Answered by
iacker
Jul 10, 2026
Replies: 1 comment
|
Pin three layers: the CUDA base image by digest (not tag), vllm== plus its torch build in a locked requirements file (pip-compile or uv lock with hashes), and the model revision by commit SHA when pulling from the Hub. Tag-based pins drift because upstream re-tags; digests + hashes are what make a rebuild months later actually reproduce. Rebuild in CI on a schedule to catch upstream breakage early rather than at deploy time. |
0 replies
Answer selected by
iacker
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Pin three layers: the CUDA base image by digest (not tag), vllm== plus its torch build in a locked requirements file (pip-compile or uv lock with hashes), and the model revision by commit SHA when pulling from the Hub. Tag-based pins drift because upstream re-tags; digests + hashes are what make a rebuild months later actually reproduce. Rebuild in CI on a schedule to catch upstream breakage early rather than at deploy time.