CLI update misidentifies DD as CE and installs an older version

[f2bo]

Today, Docker Desktop informed of an update to the Model Runner CLI v0.1.46, so I carried out the following steps:

1. Before updating, I checked the current version and noticed it was v0.1.45.

2. Following the update, I checked again and was surprised to see that the version hadn't changed, so I restarted Docker Desktop, even though the update stated that a restart was not required.

3. I checked the version again and saw that it had reverted to v0.1.23 and that the Docker Engine Kind displayed Moby.

   

Moreover, I executed docker model run, specifying a model I'd already been using previously and it began to pull the model again, so I canceled it.

This has already happened in the past and was told that it was probably caused by a bug in the CLI misidentifying DD as CE (see #121 (comment)). I'm not sure why this resurfaced but something that may be significant is that yesterday I deleted several docker/model-runner unused images that I was told at the time were not needed (with Docker Desktop). Today, I see that one of these images is back.

Also, I don't know if it's related, but I reported an additional problem where DMR stopped using Vulkan (see #305). However, this had already happened before this update.

@p1-0tr

[f2bo]

Interesting discovery. What I reported above was the result of running the DMR CLI from a WSL bash prompt. Running the same commands from a Windows command prompt shows a different result (v0.1.46). I'm also able to run an existing model without pulling it again.

However, switching back to the WSL prompt, it still shows v0.1.23 as the version and if I run the same model, it attempts to pull it again.

I see a PR has been opened. In the meantime, any suggestion on how to proceed? I'm hesitant to attempt changes that might make it more difficult to pinpoint the cause.

[ericcurtin]

Interesting discovery. What I reported above was the result of running the DMR CLI from a WSL bash prompt. Running the same commands from a Windows command prompt shows a different result (v0.1.46). I'm also able to run an existing model without pulling it again.

However, switching back to the WSL prompt, it still shows v0.1.23 as the version and if I run the same model, it attempts to pull it again. I see a PR has been opened. In the meantime, any suggestion on how to proceed? I'm hesitant to attempt changes that might make it more difficult to pinpoint the cause.

Yeah I'm not sure if that PR is correct though... If you have installed Docker Engine (Moby) in WSL2 and also use via Docker Desktop this may be expected behaviour @p1-0tr WDYT?

[f2bo]

If you have installed Docker Engine (Moby) in WSL2 and also use via Docker Desktop this may be expected behaviour.

I have not.

[f2bo]

I've tried a few things at this point: docker model uninstall-runner, deleting model-runner images, disabling and re-enabling Model Runner, disabling and re-enabling WSL integration in Docker Desktop, but always end up in the same place, with an (outdated) model-runner image being downloaded again.

Another problem, which perhaps can provide a hint of what could be wrong, at least for someone more familiar with the internals of DMR, is that running a command appears to always attempt to re-create the docker-model-runner container and fails. Stopping and deleting the container allows the command to succeed one more time, but then it fails again with the same error. For example, listing models, twice in succession:

Also, why does it always download an outdated image (Moby v0.1.23)? Is it hard-coded or is it querying a distribution server and perhaps getting the wrong answer?

[doringeman]

@f2bo, you can use docker ps --filter label=com.docker.desktop.service to see the docker/model-runner container.
Do you see the model plugin in docker info?

[f2bo]

Hi @doringeman.

Thanks. The container ID is also included in the error message so I could still delete it but yes, that allows me to see the docker/model-runner container. Nevertheless, it's re-created every time I execute a model runner command.

docker info does show the model plugin, but it shows different versions for the WSL and Windows "sides".

WSL

Windows

[f2bo]

Found the following error in the docker-model-runner container logs:

"Failed to start sandboxed llama.cpp process to probe GPU support: unable to start process: exec: "--list-devices": executable file not found in $PATH"

Could that be significant?

Log:

~$docker logs "/docker-model-runner" time="2025-11-04T13:52:17Z" level=warning msg="Could not read VRAM size: could not get nvidia VRAM size" time="2025-11-04T13:52:17Z" level=info msg="Running on system with 23462 MB RAM" time="2025-11-04T13:52:17Z" level=info msg="Successfully initialized store" component=model-manager time="2025-11-04T13:52:17Z" level=info msg="LLAMA_SERVER_PATH: /app/bin" time="2025-11-04T13:52:17Z" level=info msg="Metrics endpoint enabled at /metrics" time="2025-11-04T13:52:17Z" level=info msg="Listening on TCP port 12434" time="2025-11-04T13:52:17Z" level=info msg="failed to ensure latest llama.cpp: llama.cpp auto-updated is disabled\n" time="2025-11-04T13:52:17Z" level=warning msg="Failed to start sandboxed llama.cpp process to probe GPU support: unable to start process: exec: \"--list-devices\": executable file not found in $PATH" time="2025-11-04T13:52:17Z" level=info msg="installed llama-server with gpuSupport=false" time="2025-11-04T13:52:18Z" level=info msg="Listing available models" component=model-manager time="2025-11-04T13:52:18Z" level=info msg="Successfully listed models, count: 1" component=model-manager

I also found another report from someone else having the same problem:

https://superuser.com/questions/1904384/docker-model-ls-keeps-trying-to-create-a-new-model-runner-container-and-that

[f2bo]

So I made some progress.

In ~/.docker/cli-plugins/, I've a docker-model symlink pointing to .docker/modules/cli-plugins-wsl/docker-model in my Windows home directory. When I execute it, it displays:

Docker Model Runner version v0.1.23
Docker Engine Kind: Moby

On the other hand, /usr/local/lib/docker/cli-plugins/ also contains a docker-model symlink pointing to /mnt/wsl/docker-desktop/cli-tools/usr/local/lib/docker/cli-plugins/docker-model. When I execute it, it displays:

Docker Model Runner version v0.1.45
Docker Engine Kind: Docker Desktop

I renamed the docker-model symlink in my home directory and when I run docker model version, it now displays v0.1.45. I'm also able to run a command multiple times without downloading the docker/model-runner image or creating a docker-model-runner container. Moreover, it also makes use of Vulkan, which is something that had suddenly stopped working and I'd reported previously (#305).

However, while things are better now, it still isn't quite right as there's still a discrepancy with the version reported by model runner from a Windows command prompt.

Docker Model Runner version v0.1.46
Docker Engine Kind: Docker Desktop 

Any suggestion on how to properly restore everything so that both WSL and Windows use the latest release?

Finally, any theories on what caused these out-of-date plugins to be installed? Was it Docker Desktop's latest update (v.4.49.0)? Presumably not, otherwise it would have broken everyone and I only found just the one report I mentioned above. I've been testing DMR since it was experimental and perhaps these are remnants of some earlier installation. Did the location of the plugin change recently?

[ericcurtin]

@f2bo thanks for making progress on this, personally I don't have a Windows machine. There's two ways of running on Windows, natively (by default it Docker Desktop)... Or in a Linux VM via WSL2, it sounds like there's some sort of mismatch here... I don't know if they are stepping on each other's ports also.