Skip to content

ci: probe experimental backends (windows-rocm, sycl, openvino)#50

Open
worthant wants to merge 5 commits into
devfrom
ci/experimental-backends
Open

ci: probe experimental backends (windows-rocm, sycl, openvino)#50
worthant wants to merge 5 commits into
devfrom
ci/experimental-backends

Conversation

@worthant

Copy link
Copy Markdown
Member

Build probes to answer "do these even compile in this fork?" — not wired into any release. Triggered only on dispatch and on changes to the probe file itself, so unrelated PRs don't pay the cost. Opening this PR runs all three once.

What each would actually deliver (read before celebrating a green check)

Backend Turbo? Verdict
windows-x64-rocm ✅ real Kernels are RDNA-aware, so this ships turbo on AMD Windows — if MSVC+hipcc cooperate. This is the one worth having.
linux-x64-sycl (fp16/fp32) ❌ none No turbo kernels for SYCL → plain llama.cpp on Intel GPUs. The core feature is absent.
linux-x64-openvino ❌ none Graph-offload backend; custom KV quant types don't map. Compatibility only.

Known sharp edges (the probe may fail at these — that's the finding)

  • Windows ROCm: AMD's Windows HIP SDK is versioned as a PRO-Edition driver, not ROCm semver. HIP_SDK_VERSION: 6.2.4 is a guess; if the installer 404s or HIP_PATH is wrong, we bump to a currently-published "For HIP SDK" build.
  • SYCL/OpenVINO: heavy toolchains (oneAPI ~10min install). If they build, we decide whether a turbo-less Intel backend is worth publishing at all.

Decision this PR informs

Whether to invest in Windows ROCm (likely yes — real feature) vs. SYCL/OpenVINO (only if a turbo-less Intel path has product value, or after we port turbo kernels to SYCL — a multi-week effort). Do not merge to keep permanently until we've read the results and decided per-backend.

worthant added 3 commits July 24, 2026 17:32
Promote dev to master: sched split-input fix for Gemma 4 E-series
Build probes to see whether these backends compile in this fork. Not
wired into any release — triggered on dispatch and on changes to this
file only, so unrelated PRs don't pay the cost.

Honesty about what each delivers:
- windows-x64-rocm: REAL turbo (kernels are RDNA-aware); probes whether
  MSVC + hipcc cooperate. HIP SDK version is a guess and may need a bump.
- linux-x64-sycl (fp16/fp32): COMPATIBILITY ONLY — no turbo kernels for
  SYCL, so plain llama.cpp on Intel GPUs.
- linux-x64-openvino: COMPATIBILITY ONLY — graph-offload backend, custom
  KV quant types don't map onto it.
worthant added 2 commits July 24, 2026 19:34
…heRock

- openvino: install ocl-icd-opencl-dev + opencl-headers (build failed with
  'Could NOT find OpenCL')
- windows-rocm: the EULA-gated PRO-Edition installer URL is dead (404);
  use TheRock's portable ROCm dist (own clang toolchain) instead, building
  the RDNA3 gfx110X family as the probe representative
TheRock clang toolchain is found and configuring now; GGML_BACKEND_DL
guards on BUILD_SHARED_LIBS, which defaults ON on Linux but not in this
Windows+Ninja setup. Explicit flag. (Linux ROCm job unaffected — same as
the passing CUDA/Vulkan jobs.)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant