Skip to main content

​
GPU Targets

Targets let you run Wafer commands on remote GPUs—your own machines, cloud instances, or managed services. Configure once, then use for evaluation, profiling, and development.

​
Quick Start

# Initialize local GPU target
wafer config targets init local

# Add an SSH target
wafer config targets init ssh

# List configured targets
wafer config targets list

# Test connectivity
wafer config targets probe my-target

​
Target Types

Wafer supports several target types:
Use GPUs on your local machine:
wafer config targets init local
This detects NVIDIA or AMD GPUs automatically.

​
Commands

​
wafer config targets list

List all configured targets: 
wafer config targets list
Output: 
NAME          TYPE    GPU       STATUS
local         local   RTX 4090  ready
h100-box      ssh     H100      ready
runpod-a100   runpod  A100      stopped

​
wafer config targets init

Initialize a new target interactively: 
wafer config targets init <type>
Where <type> is one of: local, ssh, runpod, digitalocean

​
wafer config targets add

Add a target with explicit configuration: 
wafer config targets add <name> \
  --type ssh \
  --host 192.168.1.100 \
  --user ubuntu \
  --gpu H100

​
wafer config targets show

Show target configuration: 
wafer config targets show <name>

​
wafer config targets probe

Test connectivity and GPU detection: 
wafer config targets probe <name>
This verifies:
  • SSH connection (for remote targets)
  • GPU availability
  • Required tools installed

​
wafer config targets remove

Remove a target: 
wafer config targets remove <name>

​
wafer config targets default

Set the default target: 
wafer config targets default <name>
The default target is used when no --target flag is specified.

​
wafer config targets install

Install Wafer dependencies on a target: 
wafer config targets install <name>
This installs Python packages and tools needed for remote execution.

​
Using Targets

Once configured, use targets with other commands: 
# Evaluate on a specific target
wafer evaluate gpumode --impl kernel.py --target h100-box

# Run baseline on remote GPU
wafer baseline run "torch.matmul(A, B)" --target my-target

# Profile remotely
wafer nvidia ncu analyze ./profile.ncu-rep --target local

​
SSH Target Configuration

For SSH targets, ensure:
  1. SSH key authentication is set up (password auth not recommended)
  2. NVIDIA/AMD drivers are installed on the remote machine
  3. Python 3.8+ is available
  4. CUDA/ROCm toolkit is installed (for profiling)
Example manual target config: 
wafer config targets add gpu-server \
  --type ssh \
  --host gpu.example.com \
  --user ubuntu \
  --key ~/.ssh/gpu_key \
  --gpu "NVIDIA H100" \
  --cuda-path /usr/local/cuda

​
Target Pools

Group multiple targets for load balancing: 
# List pools
wafer config targets pool-list

# Create a pool
wafer config targets pool-create training-pool --targets h100-1,h100-2,h100-3

# Check pool status
wafer config targets pool-status training-pool

​
Troubleshooting

Check:
  • SSH connectivity: ssh user@host
  • GPU drivers: nvidia-smi or rocm-smi
  • Python availability: python --version
Increase timeout or check network connectivity:
wafer config targets probe my-target --timeout 60
Ensure drivers are installed and GPU is visible:
# For NVIDIA
ssh user@host nvidia-smi

# For AMD
ssh user@host rocm-smi

​
Next Steps

Workspaces

Use managed cloud workspaces.

Specs

Configure GPU specifications.

Evaluation

Run evaluations on targets.

Profiling

Profile on remote GPUs.