2026-04-06-The-Docker-SYCL-Victory

April 6, 2026 - The Docker SYCL Victory: Taming the i9 Ultra

I have finally achieved the “Gold Standard” for my local AI stack. After the manual SYCL breakthroughs on April 3rd, I realized that maintaining the source builds and the manual oneAPI environment was a “manageable mess” that was quickly becoming unmanageable.

The Final Solution: Containerized SYCL (IPEX-LLM)

We have transitioned the Primary (P) layer of the PACE plan from a manual source build to a high-performance Docker container.

Why this is the “Victory”:

1. Isolation: All the finicky Intel Level Zero and oneAPI libraries are packaged into a single, stable image (intelanalytics/ipex-llm-inference-cpp-xpu).
2. 100% GPU Offload: Verified. My Qwen 2.5 Coder 7B model now runs with every single layer offloaded to the Intel Arc iGPU.
3. No More “Blind Metrics”: While intel_gpu_top is still “blind” to the new Xe driver’s activity, the Docker container’s logs and the instant response times confirm the hardware is being used.
4. Stability: Memory footprint is a rock-solid ~5GB VRAM (UMA), leaving the rest of my 32GB RAM free for VS Code and other tools. No more OOM kills.

Technical Reasoning (The AAR):

• Xe Driver vs. i915: Meteor Lake (MTL) uses the xe kernel driver. Traditional Intel tools expect i915 and report “No device found,” which led to initial confusion.
• Unified Memory (UMA): By setting the Docker shm_size to 16GB, we’ve given the Arc iGPU the room it needs to breathe.
• Port Discipline: I’ve disabled the host-level ollama service to free up Port 11434 for the optimized Docker container.

The Updated Stack

• Endpoint: http://localhost:11434
• Docker Compose: Located in Network/Services/Ollama/docker-compose.yml.
• Primary Model: qwen2.5:7b (High performance, perfect for coding).

Knowledge Check & Future SOP:

• Updating: Just docker pull the image. No more make -j$(nproc)!
• Monitoring: Trust the container logs (docker logs ollama-arc) over host-level tools until intel-gpu-tools catches up with the Xe driver.

Onward to more coding, less waiting! 🚀

---

Related: Ollama Container Setup, Ollama Setup Guide