Ollama Intel Arc GPU Needs Source Build

This is a textbook case of hardware marketing outrunning software support: the silicon can do the work, but the usable path is broken across runtime detection, version skew, and ABI mismatches.

• –Intel's standard oneAPI SYCL runtime does not detect Arc GPUs under WSL2, so the GPU never appears to Ollama in the default setup.
• –The patched ipex-llm runtime that does work is bundled with Ollama 0.9.3, which is too old for newer model support like Qwen 3.5.
• –Ollama's vendored ggml fork changes backend function signatures, so you cannot simply drop in a newer SYCL binary.
• –The fix requires source rebuilds, exact upstream commit matching, and runtime library grafting, which makes local GPU inference fragile for non-experts.
• –Even after the fix, the performance gain is modest, so the real value is compatibility and control rather than raw speed.