net: support krunkit network.mode: shared

[acelinkio]

This change should help addresses #1559. This change should allow for l2 announcements to work when using tools like metallb, cilium l2 announcements.

both lima & krunkit support socket_vmnet

• krunkit ≥ 1.0 supports virtio-net,type=unixstream,path=… (via libkrun krun_add_net_unixstream) Add --device virtio-net,type=unix{gram,stream} libkrun/krunkit#63
• lima ≥ 2.x's lima-driver-krunkit plumbs socket_vmnet Support for libkrun using krunkit lima-vm/lima#4137

Verified that the templating adds another interface.

🍎 ❯ grep -A10 "^networks:" ~/.colima/_lima/colima-mtest/lima.yaml
networks:
    - lima: user-v2
    - socket: /Users/ep/.colima/mtest/daemon/vmnet.sock
      interface: col0
      metric: 300
provision:
    - mode: system
      script: sysctl -w fs.inotify.max_user_watches=1048576
    - mode: system
      script: grep '127.0.0.1 colima-mtest' /etc/hosts || echo '127.0.0.1 colima-mtest' >> /etc/hosts
    - mode: system

Will test tomorrow with finishing up a Kubernetes cluster and seeing if the L2 announcements work as expected.

One step closer to having gpu accelerated pods running in Kubernetes!

[acelinkio]

Confirmed that network.mode: shared works following this change. I was able to successfully use L2Annoucements in my Kubernetes cluster without issue.

Thanks again for this project! Next item to start testing is if I can get a GPU accelerated pod running in Kubernetes.

[acelinkio]

I think this PR can be merged. Verified functionality of network and GPU passthru. Going to dump some of my progress here.

I was able to get the GPU Pass Thru following the same demo code here https://github.com/medyagh/ai-playground-minikube/tree/main/macos with one small change.

squat/generic-device-plugin image tag was not specified so its pulling a newer version to create that container. That version uses devic.es/dri: 1 instead of squat.ai/dri: 1 for devices.

🍎 ❯ kubectl exec pod/uvkc-gpu -it -- /uvkc/mad_throughput
2026-04-18T15:58:20+00:00
Running /uvkc/mad_throughput
Run on (4 X 48 MHz CPU s)
Load Average: 2.19, 0.78, 0.47
-----------------------------------------------------------------------------------------------------------------------------------------
Benchmark                                                                               Time             CPU   Iterations UserCounters...
-----------------------------------------------------------------------------------------------------------------------------------------
Virtio-GPU Venus (Apple M5 Pro)/mad_throughput_f32/1048576/100000/manual_time      511694 us         2334 us            1 FLOps=4.09845T/s
Virtio-GPU Venus (Apple M5 Pro)/mad_throughput_f32/1048576/200000/manual_time     1022130 us         4107 us            1 FLOps=4.10349T/s
Virtio-GPU Venus (Apple M5 Pro)/mad_throughput_f16/1048576/100000/manual_time      280612 us         1820 us            3 FLOps=7.4735T/s
Virtio-GPU Venus (Apple M5 Pro)/mad_throughput_f16/1048576/200000/manual_time      558286 us         2800 us            1 FLOps=7.51282T/s
llvmpipe (LLVM 20.1.7, 128 bits)/mad_throughput_f32/1048576/100000/manual_time   22940306 us         27.6 us            1 FLOps=91.4178G/s
llvmpipe (LLVM 20.1.7, 128 bits)/mad_throughput_f32/1048576/200000/manual_time   45849904 us         23.7 us            1 FLOps=91.479G/s
llvmpipe (LLVM 20.1.7, 128 bits)/mad_throughput_f16/1048576/100000/manual_time   72468475 us         23.0 us            1 FLOps=28.9388G/s
llvmpipe (LLVM 20.1.7, 128 bits)/mad_throughput_f16/1048576/200000/manual_time  368624292 us         87.0 us            1 FLOps=11.3783G/s

However when I go to use the ghcr.io/ggml-org/llama.cpp:light-vulkan image to run a model I am getting

root@llama-server-7988bcff84-p9l4t:/app# ./llama-cli -ngl 99
load_backend: loaded CPU backend from /app/libggml-cpu-armv8.2_2.so
warning: no usable GPU found, --gpu-layers option will be ignored
warning: one possible reason is that llama.cpp was compiled without GPU support
warning: consult docs/build.md for compilation instructions
error: --model is required

root@llama-server-7988bcff84-p9l4t:/app# ./llama-cli --list-devices
load_backend: loaded CPU backend from /app/libggml-cpu-armv8.2_2.so
Available devices:

root@llama-server-7988bcff84-p9l4t:/app# apt-get update && apt-get install -y vulkan-tools
...
...
...
root@llama-server-7988bcff84-p9l4t:/app# vulkaninfo
ERROR at ./vulkaninfo/./vulkaninfo.h:613:vkCreateInstance failed with ERROR_OUT_OF_HOST_MEMORY

still troubleshooting, but that is unrelated to this PR.

EDIT:
GPU PassThru working with
macos -> colima -> krunit -> k3s/containerd -> pod. issue was with the container image. Use quay.io/ramalama/ramalama:0.19

[abiosoft]

This was actually an oversight from me.
Thanks :)