TL;DR

Quick diagnostic commands for the most common Ollama problems:

# Check if Ollama is running
systemctl status ollama
curl http://localhost:11434/api/version

# Check GPU detection
ollama ps
nvidia-smi   # NVIDIA
rocm-smi     # AMD

# Check disk space for model downloads
df -h ~/.ollama

# Check memory available
free -h

# View Ollama logs
journalctl -u ollama -n 50 --no-pager

# Force CPU-only mode if GPU is broken
OLLAMA_NUM_GPU=0 ollama serve

If you are running into an issue not covered here, the Ollama logs are almost always the fastest path to a diagnosis. Start there.

Error: Model Requires More System Memory

This is the single most common Ollama error. It means the model you are trying to load exceeds available RAM or VRAM.

Symptoms

Error: model requires more system memory (X GiB) than is available (Y GiB)

Root Cause

Large language models must fit entirely in memory during inference. A 7B parameter model at Q4 quantization needs roughly 4-5 GB. A 70B model needs 35-40 GB. If your system does not have enough combined RAM and VRAM, Ollama refuses to load the model.

Fixes

ApproachCommandWhen to Use
Use a smaller modelollama run llama3.2:3bUnder 8 GB RAM
Use higher quantization compressionollama run llama3.1:8b-q2_KTight on memory
Force CPU-only (frees VRAM)OLLAMA_NUM_GPU=0 ollama run llama3.1:8bGPU VRAM too small
Close other applications–Free up system RAM
Add swap spacesudo fallocate -l 8G /swapfileLast resort, slow

Check what is consuming memory before blaming the model:

free -h
# Look at the "available" column, not "free"

# Check GPU VRAM usage
nvidia-smi --query-gpu=memory.used,memory.total --format=csv

Caution: Running models from swap is technically possible but results in extremely slow inference. It is not a practical solution for interactive use.

Model Size Reference

ModelParametersQ4_K_M SizeMinimum RAM
Phi-3 Mini3.8B~2.2 GB4 GB
Llama 3.23B~2.0 GB4 GB
Llama 3.18B~4.7 GB8 GB
Mistral7B~4.1 GB8 GB
Llama 3.170B~40 GB48 GB
Llama 3.1405B~230 GB256 GB

Connection Refused on Port 11434

Symptoms

Error: could not connect to ollama app, is it running?
dial tcp 127.0.0.1:11434: connect: connection refused

Root Cause

The Ollama server process is not running, or it is bound to a different address.

Fixes

Step 1: Start the service.

# If installed via the official install script
sudo systemctl start ollama
sudo systemctl enable ollama   # persist across reboots

# If running manually
ollama serve

Step 2: Check the service status.

systemctl status ollama
# Look for "Active: active (running)"

Step 3: Verify the listen address.

By default, Ollama binds to 127.0.0.1:11434. If you need it accessible from other machines (such as for Open WebUI on another host), set the OLLAMA_HOST environment variable:

# Edit the systemd service override
sudo systemctl edit ollama

# Add these lines:
[Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"

Then reload and restart:

sudo systemctl daemon-reload
sudo systemctl restart ollama

Security note: Binding to 0.0.0.0 exposes Ollama to your entire network. The Ollama API has no authentication. Only do this on trusted networks or behind a reverse proxy with authentication.

Step 4: Check for port conflicts.

sudo ss -tlnp | grep 11434

If another process is using the port, either stop it or change the Ollama port via OLLAMA_HOST=127.0.0.1:11435.

Slow Inference

Symptoms

You are getting responses, but at 1-3 tokens per second instead of the expected 20-60+ tokens per second.

Root Cause

Ollama has fallen back to CPU inference, or the model is too large for your GPU VRAM and is being partially offloaded.

Diagnostic Steps

# Check if the model is using GPU layers
ollama ps
# Look at the "processor" column -- it should show "gpu" or a percentage

# Check GPU utilization during inference
watch -n 1 nvidia-smi    # NVIDIA
watch -n 1 rocm-smi      # AMD

# Check Ollama logs for GPU layer offloading
journalctl -u ollama -n 100 --no-pager | grep -i "gpu\|cuda\|layer"

Fixes

GPU not being used at all:

# Verify CUDA toolkit is installed (NVIDIA)
nvcc --version
nvidia-smi

# Verify ROCm is installed (AMD)
rocm-smi
rocminfo | grep "Name:" | head -5

If nvidia-smi shows your GPU but Ollama is not using it, reinstall Ollama. The install script detects GPU drivers at install time:

curl -fsSL https://ollama.com/install.sh | sh

Partial GPU offload (model too big for VRAM):

When a model exceeds VRAM, Ollama splits layers between GPU and CPU. This is slower than full GPU but faster than pure CPU. You can control this:

# Force all layers to GPU (will fail if VRAM is insufficient)
OLLAMA_NUM_GPU=999 ollama run llama3.1:8b

# Force CPU only (useful for debugging)
OLLAMA_NUM_GPU=0 ollama run llama3.1:8b

Disk I/O bottleneck:

Models are memory-mapped from disk on first load. If your models are on a spinning HDD, initial load and cold inference will be slow.

# Check where models are stored
ls -la ~/.ollama/models/

# Check disk type
lsblk -d -o NAME,ROTA
# ROTA=1 means rotational (HDD), ROTA=0 means SSD/NVMe

Move your model storage to an NVMe drive if possible:

# Move models to faster storage
sudo systemctl stop ollama
mv ~/.ollama /fast-nvme-drive/ollama
ln -s /fast-nvme-drive/ollama ~/.ollama
sudo systemctl start ollama

Model Download Failures

Symptoms

Error: pull model manifest: Get "https://registry.ollama.com/...": dial tcp: lookup registry.ollama.com: no such host
Error: unexpected EOF

Fixes

Network issues:

# Test connectivity to the Ollama registry
curl -I https://registry.ollama.com/v2/

# Check DNS resolution
nslookup registry.ollama.com

Disk space:

# Model downloads go to ~/.ollama
df -h ~/.ollama

# Check current model storage usage
du -sh ~/.ollama/models/

Models require roughly 2x their final size during download (compressed download plus extracted files). A 4.7 GB model may need 10 GB of free space.

Partial downloads and cache corruption:

If a download was interrupted, Ollama may leave partial files. Clear the cache and retry:

# Remove incomplete downloads
rm -rf ~/.ollama/models/blobs/sha256-*-partial

# Or remove a specific model entirely and re-pull
ollama rm llama3.1:8b
ollama pull llama3.1:8b

Proxy/firewall issues:

# If behind a proxy
export HTTPS_PROXY=http://proxy.corp:8080
export HTTP_PROXY=http://proxy.corp:8080

# For the systemd service
sudo systemctl edit ollama
# Add:
[Service]
Environment="HTTPS_PROXY=http://proxy.corp:8080"

Context Length Exceeded

Symptoms

Error: context length exceeded

Or the model silently truncates input and produces nonsensical output.

Root Cause

Every model has a maximum context window. If your prompt plus conversation history exceeds this limit, inference fails or degrades.

Fixes

Increase the context size at runtime:

# Set context length to 8192 tokens
ollama run llama3.1:8b --num-ctx 8192

# Via API
curl http://localhost:11434/api/generate -d '{
  "model": "llama3.1:8b",
  "prompt": "your prompt here",
  "options": { "num_ctx": 8192 }
}'

Caution: Larger context windows consume more memory. Doubling num_ctx roughly doubles the KV cache memory requirement. On a GPU with 8 GB VRAM, setting num_ctx to 32768 on a 7B model may push you into CPU offloading territory.

ModelDefault ContextMax ContextVRAM at Max Context
Llama 3.1 8B2048131072~50 GB
Mistral 7B204832768~12 GB
Phi-3 Mini2048128000~45 GB

Create a Modelfile for persistent configuration:

FROM llama3.1:8b
PARAMETER num_ctx 16384

ollama create llama3.1-16k -f Modelfile
ollama run llama3.1-16k

GPU Not Detected

Symptoms

Ollama runs but uses CPU only. ollama ps shows no GPU layers.

Diagnostic Commands

# NVIDIA: verify driver is loaded
nvidia-smi
cat /proc/driver/nvidia/version

# AMD: verify ROCm
rocm-smi
rocminfo

# Check if Ollama sees the GPU
journalctl -u ollama | grep -i "gpu\|cuda\|rocm\|metal"

Fixes for NVIDIA

# Install/update NVIDIA drivers
sudo apt update
sudo apt install nvidia-driver-550   # or latest available

# Verify CUDA compatibility
nvidia-smi  # shows CUDA version supported by driver

# If running in Docker, install NVIDIA Container Toolkit
sudo apt install nvidia-container-toolkit
sudo systemctl restart docker

After installing drivers, reinstall Ollama so it detects the GPU at install time:

curl -fsSL https://ollama.com/install.sh | sh
sudo systemctl restart ollama

Fixes for AMD

# Install ROCm (Debian/Ubuntu)
sudo apt install rocm-hip-runtime rocm-opencl-runtime

# Add user to required groups
sudo usermod -aG video $USER
sudo usermod -aG render $USER
# Log out and back in

# Verify
rocm-smi

Permission Errors

Symptoms

Error: listen tcp 127.0.0.1:11434: bind: permission denied
Error: open /home/user/.ollama/models: permission denied

Fixes

Socket/port permissions:

Ports below 1024 require root. Ollama’s default port (11434) does not require root, but if another process has locked the port or the user lacks permissions:

# Check who owns the Ollama data directory
ls -la ~/.ollama/

# Fix ownership if wrong
sudo chown -R $USER:$USER ~/.ollama/

# If running as systemd service, check the service user
grep -i "user" /etc/systemd/system/ollama.service

Systemd service user:

The default Ollama systemd service runs as the ollama user. If you installed models as your own user, there may be a permissions mismatch:

# Option 1: Fix permissions for the ollama user
sudo chown -R ollama:ollama /usr/share/ollama/.ollama/

# Option 2: Change the service to run as your user
sudo systemctl edit ollama
# Add:
[Service]
User=your-username

Group membership (GPU access):

# NVIDIA
sudo usermod -aG video $USER

# AMD ROCm
sudo usermod -aG video $USER
sudo usermod -aG render $USER

# Apply without logout
newgrp video

Models Loading Slowly

Symptoms

The first prompt after starting Ollama or switching models takes 30+ seconds before any tokens appear.

Root Cause

Ollama memory-maps model files from disk. On the first load, the OS reads the model into memory. Subsequent loads use the page cache and are faster.

Fixes

# Check your disk speed
sudo hdparm -t /dev/sda    # HDD/SSD
sudo hdparm -t /dev/nvme0n1  # NVMe

# Typical speeds:
# HDD:    ~150 MB/s  (30s to load a 4.7 GB model)
# SATA SSD: ~550 MB/s (9s)
# NVMe:  ~3500 MB/s  (1.3s)

Keep models loaded in memory:

# Set keep-alive to prevent model unloading (default is 5 minutes)
curl http://localhost:11434/api/generate -d '{
  "model": "llama3.1:8b",
  "keep_alive": -1
}'

Setting keep_alive to -1 keeps the model in memory indefinitely. This eliminates reload latency but means that memory is not freed for other models.

Preload a model at boot:

# Add to a startup script or cron @reboot
curl -s http://localhost:11434/api/generate -d '{"model":"llama3.1:8b","keep_alive":-1}' > /dev/null

General Diagnostic Workflow

When none of the above sections match your issue, follow this sequence:

# 1. Check service status
systemctl status ollama

# 2. Read recent logs
journalctl -u ollama -n 100 --no-pager

# 3. Check system resources
free -h
df -h ~/.ollama
nvidia-smi  # if NVIDIA GPU

# 4. Test the API directly
curl http://localhost:11434/api/version
curl http://localhost:11434/api/tags

# 5. Try a minimal model to isolate the issue
ollama run tinyllama "hello"

# 6. Check Ollama version (update if outdated)
ollama --version
curl -fsSL https://ollama.com/install.sh | sh  # update to latest

Most Ollama issues fall into three buckets: insufficient resources (memory/disk), missing drivers (GPU), or service configuration (systemd/networking). Start with the logs, check your resources, and work from there.