Common Issues

Frequently encountered problems and their solutions.

Installation Issues

"Ollama command not found"

Problem: PowerShell doesn't recognize ollama command.

Solutions:

1. Restart PowerShell after installation
2. Check PATH:

   $env:PATH -split ';' | Select-String ollama

3. Reinstall Ollama from ollama.ai

"CUDA not available"

Problem: Ollama uses CPU instead of GPU.

Check:

nvidia-smi

Solutions:

1. Update NVIDIA drivers (525.60+ for CUDA 12)
2. Verify GPU is detected:

   nvidia-smi --query-gpu=name --format=csv

3. Reinstall CUDA toolkit if needed

"Access denied" during setup

Problem: Scripts can't create files or modify settings.

Solution: Run PowerShell as Administrator:

Start-Process powershell -Verb RunAs

Model Issues

"Model not found"

Problem: ollama run modelname fails.

Solutions:

1. Check exact model name:

   ollama list

2. Pull the model first:

   ollama pull qwen3:32b

3. Check for typos (e.g., qwen3:32b not qwen-3:32b)

"Out of memory"

Problem: CUDA out of memory error during inference.

Solutions:

1. Use smaller quantization:

   ollama pull qwen3:32b-q4_K_M  # Instead of default

2. Reduce context window:

   ollama run qwen3:32b --num-ctx 2048

3. Unload other models:

   ollama ps  # Check loaded models
   ollama stop other-model

4. Use a smaller model

"Model loads slowly"

Problem: First response takes 10+ seconds.

Causes:

• Model loading from disk
• Cold start after OLLAMA_KEEP_ALIVE timeout

Solutions:

1. Increase keep-alive:

   $env:OLLAMA_KEEP_ALIVE = "30m"

2. Use faster storage (NVMe SSD)
3. Preload model:

   ollama run qwen3:32b ""  # Empty prompt just loads

"Model gives wrong answers"

Problem: Output is incorrect or nonsensical.

Solutions:

1. Lower temperature for factual tasks:

   ollama run qwen3:32b --temperature 0.3

2. Use appropriate model (coding model for code, etc.)
3. Check if model is corrupted:

   ollama rm qwen3:32b
   ollama pull qwen3:32b

Container Issues

"Container won't start"

Problem: docker/podman run fails.

Check:

# Docker
docker info

# Podman
podman info

Solutions:

1. Start container runtime:
   • Docker: Open Docker Desktop
   • Podman: podman machine start
2. Check for port conflicts:

   netstat -an | Select-String "3000"

"Open WebUI shows no models"

Problem: WebUI loads but no Ollama models appear.

Causes:

• Container can't reach Ollama
• Wrong OLLAMA_BASE_URL

Solutions:

1. Check Ollama is running:

   curl http://localhost:11434/api/tags

2. For Podman, see Podman Connectivity
3. Verify environment variable:

   podman inspect open-webui --format '{{range .Config.Env}}{{println .}}{{end}}'

"Permission denied" in container

Problem: Container logs show permission errors.

Solutions:

1. Check volume ownership:

   docker volume inspect open-webui

2. Run as root (temporary fix):

   docker run --user root ...

Network Issues

"Connection refused"

Problem: Can't connect to services.

Checklist:

1. Is the service running?

   ollama ps        # Ollama
   docker ps        # Containers

2. Is the port open?

   netstat -an | Select-String "11434"

3. Is firewall blocking?

   Get-NetFirewallRule | Where-Object {$_.DisplayName -like "*Ollama*"}

"Timeout" errors

Problem: Requests take too long and fail.

Solutions:

1. Check if Ollama is under load:

   ollama ps

2. Increase timeout in client
3. Check network connectivity:

   Test-NetConnection localhost -Port 11434

"Slow responses"

Problem: Inference is slower than expected.

Check:

# GPU utilization
nvidia-smi -l 1

Solutions:

1. Ensure GPU is being used (not CPU)
2. Check for thermal throttling (keep GPU < 80°C)
3. Close other GPU-intensive applications
4. Use quantized models for speed

Performance Issues

"GPU not fully utilized"

Problem: nvidia-smi shows low GPU usage.

Causes:

• Small batch size
• CPU bottleneck during prompt processing

Solutions:

1. Increase batch size:

   $env:OLLAMA_BATCH_SIZE = 512

2. Use longer prompts to keep GPU busy
3. Enable parallel requests:

   $env:OLLAMA_NUM_PARALLEL = 4

"System becomes unresponsive"

Problem: Computer freezes during inference.

Causes:

• All RAM consumed
• GPU driver crash

Solutions:

1. Limit CPU layers:

   $env:OLLAMA_NUM_CPU = 0  # GPU only

2. Use smaller model
3. Update GPU drivers
4. Add more system RAM

Web Search Issues

"Search returns no results"

Problem: Web search feature doesn't work.

For Open WebUI:

1. Check search is enabled (Settings → Web Search)
2. Verify API key (if required by provider)
3. Try different provider (DuckDuckGo needs no key)

For Perplexica:

1. Check SearXNG is running:

   curl http://localhost:4000

2. Verify engines are enabled in searxng/settings.yml

"SearXNG returns empty"

Problem: SearXNG queries return nothing.

Solutions:

1. Some engines may be rate-limited
2. Try direct search:

   curl "http://localhost:4000/search?q=test&format=json"

3. Enable more engines in settings

Still Stuck?

Collect Debug Information

# System info
nvidia-smi
ollama --version
docker --version 2>$null
podman --version 2>$null

# Ollama status
ollama ps
ollama list

# Container status
docker ps -a 2>$null
podman ps -a 2>$null

# Network
netstat -an | Select-String "11434|3000|3002|4000"

Run Test Suite

.\test-ollama-stack.ps1 -Full

Check Logs

# Ollama logs (if running as service)
Get-Content "$env:USERPROFILE\.ollama\logs\server.log" -Tail 50

# Container logs
docker logs open-webui --tail 50