Forky Pig: The Demo

Ephemeral dev boxes fit for AI Workloads is now here!

Turn one GPU server into an isolated executeion platform for agents, testing, anything your heart desires!

This is a single-instance demonstration of Forky Pig. Instead of Firecracker, we use Porker VMM. This allows for GPU passthrough to vm environments. This means it's NOT a container trick, it's NOT some namespace hack; it's a real isolated environment!

Pre-reqs

You'll need:

• A bare-metal Ubuntu 24.04 host with an NVIDIA GPU
• Root (SSH in as root, or a user with sudo)
• VT-d / AMD-Vi (IOMMU) turned on in the BIOS — most server boards have it, some consumer ones, too!
• ~20 GB free on /var

Install on a target host:

You will need a baremetal machine somewhere.

sudo ./install/preflight.sh    # sanity check — does this box even have what we need?
sudo ./install/host-prep.sh    # IOMMU + GPU rebinding. edits boot config, then reboots
# ... box reboots ...
cd forky-pig-0.1.0
sudo ./install/install.sh      # the actual install

Running Forky Pig server

I've been running it like so:

sudo -E env     DATABASE_URL="postgres://forky:forky@localhost:5432/forky?sslmode=disable"     PORT=8080     EXTERNAL_IFACE=wlp0s20f3     EXTERNAL_IP=10.0.0.59     GPU_BDFS=0000:01:00.0     PORKER_BIN_PATH=/home/user/code/firecracker/build/cargo_target/release/firecracker     KERNEL_IMAGE_PATH=/home/user/code/firecracker/gpu-kernel-build/out/vmlinux     ROOTFS_DIR=/var/lib/forky/rootfs     LOG_DIR=./tmp/forky/logs     OVERLAY_INITRD_PATH=/home/user/code/vm_hypercomputer/forky-pig-v2/overlay-initramfs.cpio.gz     CATALOG_PATH=./catalog.yaml  SECRETS_KEY=LosSWgk12uVoPX7tF0+9svYso98gSws8noS7PnveYK8= GPU_DRIVER_BUNDLE_PATH=/home/user/code/firecracker/driver-6.1.155-580.159.03.ext4 ./bin/forky-daemon

You can also do:

sudo systemctl start forky-daemon
sudo systemctl status forky-daemon
sudo journalctl -u forky-daemon -f     # follow logs

Your GPU_BDFS will probably be different than mine!

Using Forky Pig

fpig will be dropped into /usr/local/bin/fpig

# use the token install.sh printed
fpig auth login --api-url http://<your-box-ip>:8080

# register your ssh key so you can get into sandboxes
fpig ssh-key add --name laptop --key-file ~/.ssh/id_ed25519.pub

# boot a sandbox with the GPU attached
fpig create --shape gpu.whole --image ubuntu-server --ssh

Forking Images

# 1. Register an image from a Docker reference.
#    This pulls + converts to a bootable rootfs asynchronously; --wait blocks
#    until it's ready instead of handing you back a "pending" row.
fpig image create \
  --name my-nginx \
  --reference docker.io/library/nginx:1.27 \
  --wait

# 2. Boot a sandbox from that image.
fpig create --shape gpu.whole --image my-nginx --ssh
#   → prints a sandbox id (sb_...) and an ssh line

# 3. Change something inside it (this is the state you want to capture).
ssh -p <port> root@<host-ip>
#   ... apt install stuff, drop in weights/config, whatever ...  then exit

# 4. Fork it — snapshot the running sandbox's disk into a NEW private image.
fpig snapshot sb_xxxxxxxx --name my-nginx-tweaked

# 5. Boot the fork. It carries all your changes from step 3.
fpig create --shape gpu.whole --image my-nginx-tweaked --ssh