Nodes

We offer two kinds of nodes: Reserved and Auto Reserved Nodes. Once provisioned, you can run VMs on them.

Reserved vs Auto Reserved

Use Reserved nodes if you need guaranteed access to GPUs for a specific duration of time.

Auto Reserved nodes are a good fit if you have an interruptible workload and are flexible about the timing and duration assigned to your VMs.

Quickstart

To create a node, you'll need to install the sf CLI

curl -fsSL https://sfcompute.com/cli/install | bash

source ~/.bashrc # For Bash
source ~/.zshrc  # For Zsh

sf login

Run the following to generate a startup.sh that configures SSH access to your nodes. If you like, this can also be a cloud-init user-data file.

cat >startup.sh <<SCRIPT
#!/bin/bash

mkdir -p /root/.ssh
cat >>/root/.ssh/authorized_keys <<"EOF"
$(cat ~/.ssh/id_rsa.pub 2>/dev/null)
$(cat ~/.ssh/id_ecdsa.pub 2>/dev/null)
$(cat ~/.ssh/id_ecdsa_sk.pub 2>/dev/null)
$(cat ~/.ssh/id_ed25519.pub 2>/dev/null)
$(cat ~/.ssh/id_ed25519_sk.pub 2>/dev/null)
$(cat ~/.ssh/id_xmss.pub 2>/dev/null)
$(cat ~/.ssh/id_dsa.pub 2>/dev/null)
EOF
SCRIPT

Now, we can create a Reserved Node

# Create a reserved node for 1 hour with a max price of $20.00 per node hour and run statup.sh when the VM starts
sf nodes create cuda-crunch --zone landsend --duration 1h --max-price 20.00 --user-data-file ./startup.sh

To check the status of your node, run:

sf nodes list

NAME                    TYPE            STATUS       CURRENT VM                 GPU    ZONE          START/END
cuda-crunch             Reserved        Running      vm_mxCExUDERw8zvxrTf0e5W   H100   landsend      Oct 2, 3pm → 5pm

You can use the current node name or ID to get logs or SSH into the VM running on your node.

sf nodes logs cuda-crunch

sf nodes ssh root@cuda-crunch

Create Reserved Nodes

# Create a reserved node for 2 hours starting now
sf nodes create cuda-crunch --zone landsend --duration 2h --max-price 10.00

# Create a reserved node starting in 1 hour for 6 hours
sf nodes create cuda-crunch --zone landsend --start "+1h" --duration 6h --max-price 10.00

# Create a reserved node with specific start/end times
sf nodes create cuda-crunch --zone landsend --start "2024-01-15T10:00:00Z" --end "2024-01-15T12:00:00Z" --max-price 10.00

If you'd like to hold onto a reserved node for a longer period, you can extend it.

See Configure Nodes for details on using custom os images or configuring VMs on startup.

Create Auto Reserved Nodes

# Create an auto reserved node with a max price of $10 per node hour
sf nodes create cuda-crunch --auto --zone landsend --max-price 10

# Create 3 auto reserved nodes with a max price of $10 per node hour
sf nodes create -n 3 --auto --zone landsend --max-price 10 -U ./cloud_init.yaml

Once created, we will continually bid to rent Nodes at or below your max price. If successful, a VM will be provisioned on the node to run your workload.

If the market price exceeds your max price, any running VM will be terminated (at the end of your allotted time) but the node will remain active and continue to wait for the market price to drop and provision a new VM.

If you'd like to stop renewing an auto reserved node, you can release it.

See Configure Nodes for details on using custom os images or configuring VMs on startup

List Nodes

You can view all your nodes with:

sf nodes list --verbose

Node: cuda-crunch
ID: n_0d2f72f0b631b674
Type: Auto Reserved Node
Status: Released
GPU: H100
Zone: landsend
Owner: sfcompute-com

📅 Schedule:
   Start: 2025-08-04 19:03:09 UTC
   End: 2025-08-07 15:50:24 UTC
   Duration: 68 hours

💰 Pricing:
   Max Price: $10.00/hour

# One auto reserved node can have multiple VMs
# VMs are created when market price <= max price
💿 Virtual Machines         Status         Start/End
   ──────────────────────────────────────────────────────────────────────
   vm_47xHVvU6dewFvMXNJZJ9M Destroyed      2025-08-05 15:41 → 18:00
   vm_9imnBqbLvpJWToxmnYCJs Destroyed      2025-08-04 19:04 → 14:00

🎯 Actions:
   Logs: sf nodes logs cuda-crunch
   SSH: sf nodes ssh root@cuda-crunch

Configure Nodes

You can configure nodes to use a custom OS image or a cloud-init file.

Custom Images

We support any custom UEFI bootable x86_64 raw image.

# Upload a custom os.img to your account
sf nodes images upload ./path/to/os.img

# Create a node with the custom image
sf nodes create -n 1 --image <image-id> --zone landsend --max-price 30

Cloud-Init

With a cloud-init file, you can set up users, SSH keys, install packages, and run setup scripts.

# Create a new SSH key pair
ssh-keygen -t ed25519 -C "youremail@example.com"

# For MacOS
pbcopy < ~/.ssh/id_ed25519.pub

# For Linux
xclip -sel clip < ~/.ssh/id_ed25519.pub

#cloud-config
disable_root: false
ssh_pwauth: false
users:
  - name: root
    ssh_authorized_keys:
      - ssh-ed25519 [ssh_public_key] alice@example.com
      - ssh-ed25519 [ssh_public_key] bob@example.com # optionally add multiple keys

# Create a node with a custom cloud-init file
sf nodes create -n 1 --user-data-file ./startup.yaml --zone landsend --max-price 30

Update Node Configuration

You can redeploy a node to update its cloud-init file or OS image.

# Redeploy a node with a new cloud-init file
sf nodes redeploy cuda-crunch --user-data-file ./new_startup.yaml

# Redeploy a node with a new os image
sf nodes redeploy cuda-crunch --image <new-image-id>

# Redeploy a node with both a new cloud-init file and os image
sf nodes redeploy cuda-crunch --image <new-image-id> --user-data-file ./new_startup.yaml

# Redeploy a node with the same configuration (re-runs cloud-init)
sf nodes redeploy cuda-crunch

# Redeploy a node with a new os and remove the existing cloud-init file
sf nodes redeploy cuda-crunch --image <new-image-id> --override-empty

Extend Nodes

To hold onto a reserved node for a longer period, you can extend it by buying more time.

# Extend a reserved node by 1 hour for $10.00 per node hour
sf nodes extend cuda-crunch --duration 1h --max-price 10.00

Release Nodes

To stop renewing an auto reserved node, you can release it. Any running VMs will be terminated at the end of your allotted time.

# Release an auto reserved Node
sf nodes release cuda-crunch

Node Observability

You can view logs or SSH into any VM running on your node.

Run sf nodes list to get the vm_id of the VM running on your node

sf nodes list

NAME                    TYPE            STATUS       CURRENT VM                 GPU    ZONE          START/END
cuda-crunch             Reserved        Running      vm_mxCExUDERw8zvxrTf0e5W   H100   landsend      Oct 2, 3pm → 5pm

Get logs for the VM running on your node

sf nodes logs cuda-crunch

# You can also get logs from a specific VM
sf nodes logs --instance vm_mxCExUDERw8zvxrTf0e5W

SSH into the VM running on your node

sf nodes ssh root@cuda-crunch

Delete Nodes

sf nodes delete cuda-crunch

You cannot delete a node that is currently running a VM. You must first release (for auto reserved nodes) or wait for the node to finish (for reserved nodes).

Deleted nodes do not appear in the output of sf nodes list

Limitations

• Nodes only support Virtual Machines

• You cannot extend or redeploy nodes less than five minutes before their scheduled end time.

• There are no public IPs.

  • To host an inference server, you’ll need to setup a proxy, like nginx on another cloud, connect that cloud to your VPN, and then point your proxy at your nodes.

• Infiniband is not supported.

• Nodes do not currently share a VPC or virtual LAN. If you want this to be the case, configure a VPN.

• Nodes take about 5 minutes to spin up. If you reboot the machines, you will lose access for a while.

• There is no persistent storage. If the underlying machine dies due to a hardware issue, you’ll be given another one, but that one will not share the same disk as your previous one.

• There is no GPU monitoring, such as active or passive checks on VMs. Some GPUs may “fall off the bus” and not appear occasionally, and you can either reboot the node, or tell us.

• If all nodes are taken, and yours dies, you may not get another node. We’ll refund you for the time, but it’s not automatic and you’ll need to tell us.

• VMs created using the nodes command do not show up in the sf vm list command.