ollama-intel-gpu

ollama-intel-gpu

Docker app from SpaceInvaderOne's Repository

Ollama for Intel Arc GPUs (B580, A770, A750, etc.) powered by Intel IPEX-LLM.

Requirements

Intel Arc GPU (B580, A770, A750, or other Arc series) with kernel driver loaded.
Overview
Ollama for Intel Arc GPUs (B580, A770, A750, etc.) powered by Intel IPEX-LLM. Drop-in replacement for the standard Ollama container — exposes the same API on port 11434. Requires an Intel Arc GPU

Ollama for Intel Arc GPUs (IPEX-LLM)

Docker image that runs Ollama on Intel Arc GPUs using Intel IPEX-LLM. Drop-in replacement for the standard Nvidia-based Ollama container — same API on port 11434.

Supported GPUs: Intel Arc B580, A770, A750, A380, and other Arc series.

Install on Unraid

Via Community Applications (Recommended)

  1. Open the Apps tab in Unraid
  2. Search for ollama-intel-gpu
  3. Click Install
  4. Configure the settings:
    • Model Storage — where models are saved on disk (default: /mnt/user/appdata/ollama-intel-gpu)
    • Ollama API Port — default 11434
    • GPU Device Selector — leave as level_zero:0 unless you have multiple Intel GPUs
  5. Click Apply

The container will start with /dev/dri passed through automatically for GPU access.

Verify It's Working

Go to the Unraid Docker tab, click the container icon, and select Logs. You should see:

oneAPI device name: Intel(R) Graphics [0xe212]
discovered 1 Level-Zero memory modules
inference compute  id=0 library=oneapi name="Intel(R) Graphics [0xe212]" total="15.9 GiB"

Pull a Model

Open the Unraid Docker tab, click the container icon, select Console, then run:

ollama run llama3.1:8b "Hello!"

Use with Open WebUI

If you're running Open WebUI on Unraid, point it at this container:

  • Set OLLAMA_BASE_URL in Open WebUI to http://<UNRAID_IP>:11434
  • The OLLAMA_ORIGINS variable is set to * by default in the template to allow cross-origin requests

Install on Linux

Requirements

  • Intel Arc GPU (B580, A770, A750, or other Arc series)
  • /dev/dri device nodes available on the host
  • Host kernel must have the i915 driver loaded

Docker Run

docker run -d \
  --name ollama-intel-gpu \
  --device=/dev/dri \
  -p 11434:11434 \
  -v ollama-data:/root/.ollama \
  spaceinvaderone/ollama-intel-gpu:latest

Then pull and run a model:

docker exec -it ollama-intel-gpu ollama run llama3.1:8b "Hello!"

Verify GPU Detection

docker logs ollama-intel-gpu

Look for oneAPI device name: Intel(R) Graphics and library=oneapi in the output.

Environment Variables

Variable Default Description
OLLAMA_HOST 0.0.0.0:11434 API listen address
ONEAPI_DEVICE_SELECTOR level_zero:0 Which Intel GPU to use
OLLAMA_NUM_PARALLEL 1 Parallel requests (keep at 1 for 12 GB cards)
OLLAMA_NUM_CTX 4096 Context window size in tokens
OLLAMA_KEEP_ALIVE 10m How long to keep model in VRAM (-1 = forever)
OLLAMA_ORIGINS (unset) Set to * for Open WebUI / CORS access

Testing the API

# Health check
curl http://localhost:11434/

# List models
curl http://localhost:11434/api/tags

# Generate
curl http://localhost:11434/api/generate -d '{
  "model": "llama3.1:8b",
  "prompt": "Hello!",
  "stream": false
}'

What's Inside

  • Base: Ubuntu 24.04
  • IPEX-LLM: v2.3.0 nightly (Ollama portable build optimised for Intel GPUs)
  • Intel GPU drivers: IGC v2.8.3, Compute Runtime 25.09.32961.7, Level-Zero Loader v1.21.9

Categories

Download Statistics

28,920
Total Downloads

Related apps

Explore more like this

Explore all

Links

Details

Repository
spaceinvaderone/ollama-intel-gpu
Last Updated2026-03-27
First Seen2026-04-03

Runtime arguments

Web UI
http://[IP]:[PORT:11434]/
Network
bridge
Shell
bash
Privileged
false
Extra Params
--device=/dev/dri

Template configuration

Model StoragePathrw

Path on the host for persistent model storage. Models are large (4-20 GB each).

Target
/root/.ollama
Default
/mnt/user/appdata/ollama-intel-gpu
Value
/mnt/user/appdata/ollama-intel-gpu
Ollama API PortPorttcp

Port for the Ollama API.

Target
11434
Default
11434
Value
11434
OLLAMA_ORIGINSVariable

Allowed origins for CORS. Set to * to allow Open WebUI and other frontends to connect.

Default
*
Value
*
ONEAPI_DEVICE_SELECTORVariable

Select which Intel GPU to use. Use level_zero:0 for the first GPU. Change only if you have multiple Intel GPUs.

Default
level_zero:0
Value
level_zero:0
OLLAMA_NUM_PARALLELVariable

Number of parallel inference requests. Set to 1 for 12 GB VRAM cards (B580). Increase only if you have more VRAM.

Default
1
Value
1
OLLAMA_NUM_CTXVariable

Context window size in tokens. Larger values use more VRAM. Default 4096 is a good balance for 12 GB cards.

Default
4096
Value
4096
OLLAMA_KEEP_ALIVEVariable

How long to keep a model loaded in VRAM after the last request. Use 5m for 5 minutes, -1 for forever, 0 to unload immediately.

Default
5m
Value
5m