0.2.6-rocm and above cannot be pulled with containerd on fedora
What is the issue?
Pulling the image results in
Error: failed to extract layer sha256:00d2c36d84f963d50ac6a568b0be71eea96f3579770ef47c2ac3f94d4d3c346a: exit status 1: unpigz: skipping: <stdin>: corrupted -- crc32 mismatch
This happens for 0.2.6-rocm and further vesions Not sure why it fails and what kind of issue pigz has here. Im using k0s which comes packaged with its own containerd so its not using fedora installed.
Really not sure where to look for a potential cause
OS
Linux
GPU
No response
CPU
Intel, AMD
Ollama version
0.3.0
fwiw, I'm able to pull and run with Docker Desktop. (confirmed with ollama/ollama:0.2.6-rocm as well as ollama/ollama:rocm) so I think the images we pushed up to Docker Hub are valid.
Im also able to pull with docker on my PC.
Need it on k8s and cannot pull it nor understand what is the issue, im here looking for ideas
Having the same problem here. When I start pulling an image, either via the command line or via Portainer, it works normall at first. Only after the download and extraction finished, the pulling exited with the error: „layers from manifest don't match image configuration“.
The „normal“ :latest image is working fine, just my homeserver cant take ithe load without the GPU.
Are there any alternative downloads available?
I'm also having this issue on Fedora 40 with docker engine (non-desktop)
I'm having the same issue.
I have the same problem. Normal image works fine, rocm build (layers from manifest don't match image configuration)
Seems that the issue is with the package pigz which in my case when detected on system means that containerd will use unpigz to exctract the image. Removing the package and reboot works. In my case pigz was not a dependency of anything to it was easy to remove
Containerd on start detects if there is unpigz and should it use it
https://github.com/containerd/containerd/issues/3824#issuecomment-553802643