Ollama 0.15.6 with SYCL backend for Intel GPU (~2x faster than Vulkan)

[eSlider]

Summary

• Upgrade ollama from 0.9.3 (IPEX-LLM bundle) to official v0.15.6
• Replace Vulkan backend with SYCL — custom-built ggml-sycl from upstream llama.cpp using Intel oneAPI 2025.1.1, delivering ~2x
inference speed on Intel GPUs
• Add GitHub Actions CI to build and push the Docker image to GHCR automatically
• Update Intel GPU driver stack to latest releases (Level-Zero 1.28.0, IGC 2.28.4, compute-runtime 26.05)

Why SYCL over Vulkan

Ollama's official release only ships a Vulkan backend for Intel GPUs. SYCL with oneAPI unlocks oneMKL, oneDNN, and Level-Zero
direct access — benchmarks show +45-100% token/s on integrated Arc (MTL) and +57-83% on discrete Arc.

Intel GPU	Vulkan	SYCL	Gain
MTL iGPU (155H)	~8-11 tok/s	~16 tok/s	+45-100%
Arc A770	~30-35 tok/s	~55 tok/s	+57-83%

Tested on Intel Core Ultra 7 155H (Meteor Lake): gemma3:1b — 27/27 layers on GPU, 10.2 tok/s generation, 65.3 tok/s prompt eval.

How the SYCL build works

Ollama intentionally excludes ggml-sycl source from its vendored ggml. This PR rebuilds it in a multi-stage Dockerfile:

1. Clones ollama v0.15.6 source (for ggml build system + headers)
2. Fetches ggml-sycl from the exact llama.cpp commit (a5bb8ba4) that ollama vendors — critical for ABI compatibility
3. Applies two patches via patch-sycl.py:
   • graph_compute signature: ollama adds int batch_size parameter
   • GGML_TENSOR_FLAG_COMPUTE: ollama removes this flag; without the patch, all compute nodes get skipped producing garbage output
4. Compiles with icpx + oneAPI, bundles runtime libs into the image

Changes

File	What changed
Dockerfile	Multi-stage build: oneAPI SYCL compile -> minimal Ubuntu runtime
patch-sycl.py	New — patches ggml-sycl for ollama API compatibility
docker-compose.yml	SYCL env vars, full /dev/dri access, debug logging
README.md	Full rewrite with architecture diagram, benchmarks, troubleshooting
CHANGELOG.md	Documents full migration history
.github/workflows/build-push.yml	New — CI to build and push image to GHCR
start-ollama.sh	Custom entrypoint fixing hardcoded OLLAMA_HOST (from IPEX-LLM era)
.gitignore	IDE and swap file exclusions

Test plan

• docker compose build completes without errors
• curl http://localhost:11434/api/tags returns model list
• SYCL backend detected in logs (SYCL0, Intel(R) Arc(TM) Graphics)
• All model layers offloaded to GPU (27/27)
• Inference produces correct output (gemma3:1b, gemma3:4b)
• CI workflow builds successfully on GitHub Actions
• Pre-built image pullable from GHCR

---

Note

Medium Risk
Reworks the container build/run stack (multi-stage oneAPI build, custom SYCL runner, updated Intel GPU runtimes), which could break runtime compatibility or device detection if versions/patching drift. No direct auth/data-path changes, but it modifies the primary deployment artifact and defaults (e.g., WebUI auth disabled).

Overview
Switches the Docker build to a SYCL-first Intel GPU stack. The Dockerfile is rewritten into a multi-stage build that compiles libggml-sycl.so with Intel oneAPI, bundles required oneAPI runtime libraries, installs the official Ollama v0.15.6 binary, and removes unused runners (CUDA/MLX/Vulkan) while updating Intel GPU user-space drivers.

Adds SYCL compatibility patching and updates runtime defaults. Introduces patch-sycl.py to patch upstream ggml-sycl for Ollama API differences, and updates docker-compose.yml to pass the Ollama version build arg, expose 11434, increase shm_size, set SYCL/Level-Zero env vars, and disable WebUI auth while switching WebUI to :latest with local persisted data.

Adds automation and repo hygiene/docs. Adds a GitHub Actions workflow to build and push the image to GHCR with versioned tagging, plus new/updated README.md/CHANGELOG.md, .gitignore, and .github/FUNDING.yml, and removes legacy IPEX-LLM init/run scripts and the WSL2 compose file.

^{Written by Cursor Bugbot for commit e397010. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

This PR is being reviewed by Cursor Bugbot

Details

You are on the Bugbot Free tier. On this plan, Bugbot will review limited PRs each billing cycle.

To receive Bugbot reviews on all of your PRs, visit the Cursor dashboard to activate Pro and start your 14-day free trial.