Skip to main content

Troubleshooting

"Your ollama is too old"​

The worker needs ollama 0.32.15 or newer and checks the version at startup. Upgrade:

curl -fsSL https://ollama.com/install.sh | sh    # Linux / WSL
brew upgrade ollama                              # macOS

Or download the current build from ollama.com/download (Windows). Check it with ollama --version, then start the worker again.

HTTP 500 from ollama, or the model never loads​

Same cause: an ollama older than 0.32.15. It can't build or run this model and reports it as a generic HTTP 500 rather than anything useful. Upgrade ollama (above) and retry — there is no worker-side workaround.

Single-digit tok/s on an RTX 30xx​

The CUDA build fell back to CPU. This was seen on ollama 0.32.14, which silently ran RTX 30xx cards on CPU while looking otherwise healthy — a 3090 that benchmarks at 5 tok/s instead of ~40-87 is this.

  1. Upgrade ollama to 0.32.15+
  2. Start a job and run ollama ps — the model should show 100% GPU. Anything with a CPU share means it isn't fully loaded on the card.

"Your device is too slow (X tok/s). Minimum required: 5 tok/s."​

Your GPU is not being used. ollama is running on CPU, which is too slow for the network. Check your ollama version first (see above), then:

NVIDIA (Linux/Windows):

# Both of these should work:
nvcc --version
nvidia-smi

If either fails, install the CUDA toolkit. On Ubuntu: sudo apt install nvidia-cuda-toolkit. On Windows, download from developer.nvidia.com.

AMD (Linux/Windows): Vulkan should be auto-detected. If not, install Vulkan drivers:

# Ubuntu
sudo apt install mesa-vulkan-drivers
# Verify
vulkaninfo | head

Apple Silicon: Metal auto-detects on Apple Silicon. If performance is unexpectedly low:

  • Check Activity Monitor → GPU tab for GPU usage
  • Make sure you're running native arm64 Node.js: node -p "process.arch" should output arm64
  • Free up RAM — close other apps

"Connection error: Invalid authentication token"​

  • Your worker token may be expired or invalid
  • Generate a new one from c0mpute.ai/earn → Native Worker → Get Worker Token
  • Make sure you're logged in to the same account that generated the token
  • Tokens start with cwt_ — make sure you copied the full string

Model download fails​

On first run the weights are pulled from a pinned HuggingFace revision into ~/.config/compute-worker/models, then built into ollama.

  • Check disk space: you need ~36GB free — the kept download plus ollama's copy of the built model (~/.ollama by default), on macOS too.
  • Check internet: try curl -I https://huggingface.co to verify connectivity
  • Retry: HuggingFace occasionally has temporary issues. Just run the command again.
  • Behind a proxy? Set HTTPS_PROXY environment variable

Worker disconnects frequently​

  • Check your internet stability — packet loss or high latency causes disconnects
  • The worker auto-reconnects after a disconnect, but you lose any in-progress jobs
  • If using WiFi, try a wired connection
  • Check if your firewall is blocking WebSocket connections

Low tok/s on Windows​

This is the most common issue. Native Windows CUDA support is flaky with ollama.

Solution: use WSL.

  1. Install WSL2: wsl --install
  2. Install Node.js and CUDA toolkit inside WSL
  3. Run the worker from WSL terminal

See the Windows setup guide for full instructions.

Key point: nvidia-smi should work inside WSL, not just in PowerShell. CUDA needs to be installed in the WSL environment.

Multi-GPU rig: only one card is working​

The worker starts one child per GPU on its own, but they all need their own ollama. If a box-wide ollama serve was already running on port 11434, GPU 0's worker adopts it instead of starting a pinned one — and that daemon sees every card, so the rig behaves like a single unpinned worker.

Stop the pre-existing daemon, then start the worker again:

pkill -f "ollama serve"           # macOS/Linux (also: sudo systemctl stop ollama)
taskkill /F /IM ollama.exe        # Windows

Then check the startup output. You should see N GPUs detected — starting one worker per GPU, and each child's lines prefixed [gpu 0], [gpu 1], … If you only want some cards, pass --gpu 3 or --gpu 0,2,5. See Multi-GPU rigs.

"Too many workers from this network (max 10 per IP)"​

The network caps concurrent workers at 10 per IP and 10 per account. On a rig with more than 10 GPUs the extra workers are refused at registration — run a subset with --gpu 0,1,2,... or split the rig across networks/accounts.

Worker starts but gets no jobs​

  • Check that your worker benchmarks above 5 tok/s (minimum threshold)
  • The network matches jobs based on availability — if many workers are online, jobs are distributed
  • There is only one text model, so there is no wrong model to be on: every native worker is eligible for every 27B job
  • Check the worker page at c0mpute.ai/earn for network status