FAQ & Troubleshooting

Configuration

Context Window Size

Default is 4096 tokens. Override per request or globally:

# Global (env var)
OLLAMA_CONTEXT_LENGTH=8192 ollama serve

Per request (API)
curl http://localhost:11434/api/generate -d '{
  "model": "qwen3:1.7b",
  "prompt": "...",
  "options": {"num_ctx": 8192}
}'

Interactive session
/set parameter num_ctx 4096

Keep Models Loaded

Default: 5 minutes after last request. Control via:

# Keep forever
curl http://localhost:11434/api/generate -d '{"model": "qwen3:1.7b", "keep_alive": -1}'

Unload immediately
curl http://localhost:11434/api/generate -d '{"model": "qwen3:1.7b", "keep_alive": 0}'
ollama stop qwen3:1.7b

Global default
OLLAMA_KEEP_ALIVE=10m ollama serve

Preload Models

Send empty request to load model into memory:

curl http://localhost:11434/api/generate -d '{"model": "qwen3:1.7b"}'
curl http://localhost:11434/api/chat -d '{"model": "qwen3:1.7b"}'
ollama run qwen3:1.7b ""

GPU vs CPU Status

ollama ps
NAME          SIZE    PROCESSOR   UNTIL
qwen3:1.7b   1.4 GB  100% CPU    4 minutes from now

Expose on Network

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Allow CORS Origins

OLLAMA_ORIGINS=chrome-extension://,http://localhost:3000 ollama serve

// ~/.ollama/server.json
{"disable_ollama_cloud": true}

[Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"
Environment="OLLAMA_MODELS=/data/models"
Environment="OLLAMA_KEEP_ALIVE=10m"
Environment="OLLAMA_CONTEXT_LENGTH=8192"
Environment="OLLAMA_FLASH_ATTENTION=1"
Environment="OLLAMA_KV_CACHE_TYPE=q8_0"
Environment="OLLAMA_MAX_LOADED_MODELS=3"
Environment="OLLAMA_NUM_PARALLEL=1"
Environment="OLLAMA_MAX_QUEUE=512"

RAM scales by: OLLAMA_NUM_PARALLEL * OLLAMA_CONTEXT_LENGTH

Performance Tuning

Flash Attention

Reduces memory with large contexts:

OLLAMA_FLASH_ATTENTION=1 ollama serve

KV Cache Quantization

Requires Flash Attention enabled:

OLLAMA_KV_CACHE_TYPE=q8_0 ollama serve

GPU Selection

# NVIDIA — select specific GPUs
CUDA_VISIBLE_DEVICES=0,1 ollama serve

AMD
ROCR_VISIBLE_DEVICES=0 ollama serve

Force CPU only
CUDA_VISIBLE_DEVICES=-1 ollama serve

Troubleshooting

Finding Logs

# Linux (systemd)
journalctl -u ollama --no-pager --follow

macOS
cat ~/.ollama/logs/server.log

Docker
docker logs <container-name>

Linux GPU Not Detected After Suspend

sudo rmmod nvidia_uvm && sudo modprobe nvidia_uvm

AMD Driver Version Mismatch

If GPU discovery times out (30s), upgrade to ROCm v7:

# Install amdgpu-install from AMD docs, then:
sudo amdgpu-install
sudo reboot

Override LLM Library

OLLAMA_LLM_LIBRARY="cpu_avx2" ollama serve
Options: cpu, cpu_avx, cpu_avx2, cuda_v11, rocm_v5, rocm_v6

503 Server Overloaded

Increase queue: OLLAMA_MAX_QUEUE=1024

Upgrade Ollama

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

Specific version
curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.5.7 sh

mindX Production Notes (from deployment at mindx.pythai.net)

VPS Constraints (4GB RAM, No GPU)

Dual-URL Failover (from ollama_url.py)

Primary:  MINDX_LLM__OLLAMA__BASE_URL (10.0.0.155:18080 when GPU available)
Fallback: localhost:11434 (always available on VPS)

Timeout:  120s total, 10s connect, 60s sock_read

Health Check Endpoint

# Simple ping
curl http://localhost:11434/api/tags

mindX admin route
curl http://localhost:8000/api/admin/ollama/status