Run Ollama, Stable Diffusion and Automatic Speech Recognition with your Intel Arc GPU

[Blog]

Effortlessly deploy a Docker-based solution that uses Open WebUI as your user-friendly AI Interface and Ollama for integrating Large Language Models (LLM).

Additionally, you can run ComfyUI or SD.Next docker containers to streamline Stable Diffusion capabilities.

You can also run an optional docker container with OpenAI Whisper to perform Automatic Speech Recognition (ASR) tasks.

The Ollama container runs a native SYCL (or Vulkan) llama.cpp backend built directly from upstream Ollama — no IPEX-LLM. The Stable Diffusion and Whisper containers are still optimized for Intel Arc GPUs using Intel® Extension for PyTorch.

Services

1. Ollama

   • Runs Ollama with a native llama.cpp ggml-sycl backend, compiled from Ollama source against Intel® oneAPI (icpx / Level Zero). No IPEX-LLM dependency.
   • The image is built locally in two stages (see ollama-sycl/Dockerfile): stage 1 builds libggml-sycl.so with oneAPI; stage 2 drops it next to the official Ollama binary on a slim Ubuntu runtime with the Intel GPU user-space drivers (Level Zero, compute-runtime, IGC, GMM).
   • A Vulkan alternative is also provided (docker-compose.ollama-vulkan.yml) using the stock ollama/ollama image. On Meteor Lake / Xe-LPG iGPUs the Vulkan backend is often competitive with SYCL while requiring no custom build — worth benchmarking on your hardware.
   • Runtime behavior is tuned via a .env file (see .env.example) — context length, KV-cache type, flash attention, GPU offload, etc.
   • Exposes port 11434 for connecting other tools to your Ollama service.

2. Open WebUI

   • Uses the official distribution of Open WebUI.
   • WEBUI_AUTH is turned off for authentication-free usage.
   • ENABLE_OPENAI_API and ENABLE_OLLAMA_API flags are set to off and on, respectively, allowing interactions via Ollama only.
   • ENABLE_IMAGE_GENERATION is set to true, allowing you to generate images from the UI.
   • IMAGE_GENERATION_ENGINE is set to automatic1111 (SD.Next is compatible).

3. ComfyUI

   • The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.
   • Uses as the base container the official Intel® Extension for PyTorch

4. SD.Next

   • All-in-one for AI generative image based on Automatic1111
   • Uses as the base container the official Intel® Extension for PyTorch
   • Uses a customized version of the SD.Next docker file, making it compatible with the Intel Extension for Pytorch image.

5. OpenAI Whisper

   • Robust Speech Recognition via Large-Scale Weak Supervision
   • Uses as the base container the official Intel® Extension for PyTorch

Setup

First, create your .env file from the example and adjust it to your hardware if needed:

$ git clone https://github.com/eleiton/ollama-intel-arc.git
$ cd ollama-intel-arc
$ cp .env.example .env

Then start Ollama + Open WebUI with the native SYCL backend (this builds the Ollama image locally the first time):

$ podman compose -f docker-compose.ollama-sycl.yml up -d --build

Alternatively, use the Vulkan backend (no local build — pulls the stock Ollama image):

$ podman compose -f docker-compose.ollama-vulkan.yml up -d

The repository also ships a legacy docker-compose.yml based on the now-outdated intelanalytics/ipex-llm-inference-cpp-xpu image. It is kept for reference only; the native SYCL/Vulkan composes above are the recommended path.

Additionally, if you want to run one or more of the image generation tools, run these command in a different terminal:

For ComfyUI

$ podman compose -f docker-compose.comfyui.yml up -d

For SD.Next

$ podman compose -f docker-compose.sdnext.yml up -d

If you want to run Whisper for automatic speech recognition, run this command in a different terminal:

$ podman compose -f docker-compose.whisper.yml up -d

Configuration

Ollama runtime behavior is controlled through environment variables in your .env file (passed through by the compose files). The defaults in .env.example are tuned for an Intel Arc Graphics (Meteor Lake-P) integrated GPU with shared/UMA memory. Key settings:

Variable	Default	Notes
OLLAMA_CONTEXT_LENGTH	8192	Larger contexts grow the KV cache and reduce the model size that fits.
OLLAMA_KV_CACHE_TYPE	q4_0	Quantized KV cache saves memory. On a UMA iGPU with plenty of RAM, try f16 or q8_0 — it can be faster and higher quality at no real memory cost.
OLLAMA_FLASH_ATTENTION	true	Works on both the SYCL and Vulkan paths on this iGPU.
OLLAMA_NUM_GPU	999	Offload all transformer layers to the GPU.
OLLAMA_NUM_PARALLEL	1	One request at a time — UMA iGPUs are bandwidth-bound, so parallelism gives no throughput gain.
OLLAMA_KEEP_ALIVE	2h	Keep models resident to avoid reload latency.
GGML_SYCL_F16	1	Enable fp16 math in the SYCL backend.

See .env.example for the full annotated list.

Validate

Run the following command to verify your Ollama instance is up and running

$ curl http://localhost:11434/
Ollama is running

When using Open WebUI, you should see this partial output in your console, indicating your arc gpu was detected

[ollama-sycl] | Found 1 SYCL devices:
[ollama-sycl] | |  |                   |                                       |       |Max    |        |Max  |Global |                     |
[ollama-sycl] | |  |                   |                                       |       |compute|Max work|sub  |mem    |                     |
[ollama-sycl] | |ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|
[ollama-sycl] | |--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
[ollama-sycl] | | 0| [level_zero:gpu:0]|                     Intel Arc Graphics|  12.71|    128|    1024|   32| 62400M|         1.6.32224+14|

(The Vulkan backend logs ggml_vulkan: Found ... Intel(R) Arc(TM) Graphics instead.)

Using Image Generation

• Open your web browser to http://localhost:7860 to access the SD.Next web page.
• For the purposes of this demonstration, we'll use the DreamShaper model.
• Follow these steps:
• Download the dreamshaper_8 model by clicking on its image (1).
• Wait for it to download (~2GB in size) and then select it in the dropbox (2).
• (Optional) If you want to stay in the SD.Next UI, feel free to explore (3).
• For more information on using SD.Next, refer to the official documentation.
• Open your web browser to http://localhost:4040 to access the Open WebUI web page.
• Go to the administrator settings page.
• Go to the Image section (1)
• Make sure all settings look good, and validate them pressing the refresh button (2)
• (Optional) Save any changes if you made them. (3)
• For more information on using Open WebUI, refer to the official documentation
• That's it, go back to Open WebUI main page and start chatting. Make sure to select the Image button to indicate you want to generate Images.

Using Automatic Speech Recognition

• This is an example of a command to transcribe audio files:

  podman exec -it  whisper-ipex whisper https://www.lightbulblanguages.co.uk/resources/ge-audio/hobbies-ge.mp3 --device xpu --model small --language German --task transcribe

• Response:

  [00:00.000 --> 00:08.000]  Ich habe viele Hobbys. In meiner Freizeit mache ich sehr gerne Sport, wie zum Beispiel Wasserball oder Radfahren.
  [00:08.000 --> 00:13.000]  Außerdem lese ich gerne und lerne auch gerne Fremdsprachen.
  [00:13.000 --> 00:19.000]  Ich gehe gerne ins Kino, höre gerne Musik und treffe mich mit meinen Freunden.
  [00:19.000 --> 00:22.000]  Früher habe ich auch viel Basketball gespielt.
  [00:22.000 --> 00:26.000]  Im Frühling und im Sommer werde ich viele Radtouren machen.
  [00:26.000 --> 00:29.000]  Außerdem werde ich viel schwimmen gehen.
  [00:29.000 --> 00:33.000]  Am liebsten würde ich das natürlich im Meer machen.

• This is an example of a command to translate audio files:

  podman exec -it  whisper-ipex whisper https://www.lightbulblanguages.co.uk/resources/ge-audio/hobbies-ge.mp3 --device xpu --model small --language German --task translate

• Response:

  [00:00.000 --> 00:02.000]  I have a lot of hobbies.
  [00:02.000 --> 00:05.000]  In my free time I like to do sports,
  [00:05.000 --> 00:08.000]  such as water ball or cycling.
  [00:08.000 --> 00:10.000]  Besides, I like to read
  [00:10.000 --> 00:13.000]  and also like to learn foreign languages.
  [00:13.000 --> 00:15.000]  I like to go to the cinema,
  [00:15.000 --> 00:16.000]  like to listen to music
  [00:16.000 --> 00:19.000]  and meet my friends.
  [00:19.000 --> 00:22.000]  I used to play a lot of basketball.
  [00:22.000 --> 00:26.000]  In spring and summer I will do a lot of cycling tours.
  [00:26.000 --> 00:29.000]  Besides, I will go swimming a lot.
  [00:29.000 --> 00:33.000]  Of course, I would prefer to do this in the sea.

• To use your own audio files instead of web files, place them in the ~/whisper-files folder and access them like this:

  podman exec -it  whisper-ipex whisper YOUR_FILE_NAME.mp3 --device xpu --model small --task translate

Updating the containers

For the native SYCL Ollama image, updates come from rebuilding against a newer Ollama release. Bump OLLAMA_VERSION (and, if needed, the Intel GPU driver pins) at the top of ollama-sycl/Dockerfile, then rebuild:

$ podman compose -f docker-compose.ollama-sycl.yml build --no-cache
$ podman compose -f docker-compose.ollama-sycl.yml up -d

For the Vulkan image, bump the ollama/ollama tag in docker-compose.ollama-vulkan.yml and pull:

$ podman compose -f docker-compose.ollama-vulkan.yml pull
$ podman compose -f docker-compose.ollama-vulkan.yml up -d

For Open WebUI and the other latest-tagged images, stop the stack and pull:

$ podman compose -f docker-compose.ollama-sycl.yml down
$ podman compose -f docker-compose.ollama-sycl.yml pull open-webui
$ podman compose -f docker-compose.ollama-sycl.yml up -d

Manually connecting to your Ollama container

You can connect directly to your Ollama container by running these commands:

$ podman exec -it ollama-sycl /bin/bash
$ ollama -v

(Use ollama-vulkan as the container name if you are running the Vulkan compose.)

My development environment:

• Core Ultra 7 155H
• Intel® Arc™ Graphics (Meteor Lake-P)
• Fedora 43

References

• Open WebUI documentation
• Ollama
• llama.cpp SYCL backend
• llama.cpp Vulkan backend
• Intel® oneAPI Base Toolkit
• Intel compute-runtime (Level Zero / OpenCL driver)
• Docker - Intel extension for pytorch
• GitHub - Intel extension for pytorch