While removing the constraint for modelscope==1.15.0 is a good cleanup, leaving modelscope completely unpinned in a production Dockerfile can lead to non-reproducible builds and potential breakages in the future if a new version introduces breaking changes.

To improve stability, it's best practice to pin dependencies. Could you please consider pinning modelscope to a known good version range? For example, if versions greater than 1.15.0 are known to be stable, you could use a lower bound like >=1.15.1. This would prevent the problematic version while still allowing updates, but in a more controlled way. An even safer approach would be to pin to a specific known-good version, e.g., ==1.16.0.

    uv pip install --system accelerate hf_transfer 'modelscope>=1.15.1' "bitsandbytes>=${BITSANDBYTES_VERSION}" 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]

Similar to the other Dockerfile, leaving modelscope unpinned here could affect the stability and reproducibility of the build. A future release of modelscope with breaking changes could cause this build to fail.

To ensure the Docker image remains stable over time, could you pin modelscope to a known good version or at least a safe version range, for example, >=1.15.1?

    pip install accelerate hf_transfer pytest 'modelscope>=1.15.1'