Eval bug: It is not possible to run docker image

[andrew-aladev]

Name and Version

Recent llama version from the master branch in docker.

Operating systems

Linux

GGML backends

CPU

Hardware

Any CPU

Models

Any model

Problem description & steps to reproduce

docker build \
  --tag 'local/llama.cpp:light-cpu' \
  --target 'light' \
  --file '.devops/cpu.Dockerfile' \
  --progress 'plain' \
  '.'

docker run --rm 'local/llama.cpp:light-cpu' --version

First Bad Commit

No response

Relevant log output

/app/llama-cli: error while loading shared libraries: libllama.so.0: cannot open shared object file: No such file or directory

[andrew-aladev]

Bad commit: 4a5b8af 4a5b8af

[Tehhs]

Yep, this is an issue for me too and I cant use "HF Inference Endpoints" deploy options via hugging face anymore as I'm getting this error. Was working yesterday before changes.

[furrysalamander]

I'm going to put any further discussion in this bug ticket instead of the PR.
Here's a link to what I've found so far:
#17091 (comment)

[furrysalamander]

I'm assuming we want to link against the top level version of the so files (rather than the unversioned ones), so I'm looking at just updating the dockerfiles to include the symlinks.

[furrysalamander]

I have a commit that should fix this. It's the same approach for cuda, cpu, etc, but the s390x Dockerfile is structured differently so I'm double checking the build instructions for that so I can validate it.

[furrysalamander]

Opened a PR.