Problem Description: Ollama runs exclusively in CPU mode despite detecting AMD GPU information. When executing `ollama ps`, the GPU is displayed as the processor. However, system monitoring shows no GPU usage. It is expected that Ollama can use an AMD GPU to accelerate running LLMs, which it cannot. Installation Details: The Ollama package was installed with the following ``` USE="rocm -blas -cuda -mkl" AMDGPU_TARGETS="gfx1100 -gfx90a -gfx803 -gfx900 -gfx906 -gfx908 -gfx940 -gfx941 -gfx942 -gfx1010 -gfx1011 -gfx1012 -gfx1030 -gfx1031 -gfx1101 -gfx1102" CPU_FLAGS_X86="avx avx2 avx512_bf16 avx512_vnni avx512f avx512vbmi f16c fma3 -amx_int8 -amx_tile -avx_vnni" ``` Cause of the Issue: According to the ebuild file at https://gitweb.gentoo.org/repo/proj/guru.git/tree/sci-ml/ollama/ollama-0.6.5-r1.ebuild, Ollama is compiled with: ``` CMAKE_HIP_ARCHITECTURES="$(get_amdgpu_flags)" CMAKE_HIP_PLATFORM="amd" ``` However, these settings are not recognized during the configuration and compilation process. As a result, the file `/usr/lib64/ollama/rocm/libggml-hip.so` is missing after installing the Ollama package. Solution Investigated and Tested: (in existing ebuild file) 1. Remove -DCMAKE_HIP_ARCHITECTURES and -DCMAKE_HIP_PLATFORM (line 234 and line 345), 2. Replace them with `-DAMDGPU_TARGETS="$(get_amdgpu_flags)"` (as referenced in https://gitlab.archlinux.org/archlinux/packaging/packages/ollama/-/blob/main/PKGBUILD) This modification successfully generates `/usr/lib64/ollama/rocm/libggml-hip.so` with rocblas. Ollama can now utilize the GPU when running (e.g., start the server with `ollama serve` and run LLM with `ollama run qwen2.5:32b`).
Created attachment 926180 [details, diff] Patch file for ollama-0.6.5-r1.ebuild
The bug has been closed via the following commit(s): https://gitweb.gentoo.org/repo/proj/guru.git/commit/?id=e92cad7ec294d72b3435e5f3aaa52a48da3c57a5 commit e92cad7ec294d72b3435e5f3aaa52a48da3c57a5 Author: Paul Zander <negril.nx+gentoo@gmail.com> AuthorDate: 2025-04-27 14:20:05 +0000 Commit: Paul Zander <negril.nx+gentoo@gmail.com> CommitDate: 2025-04-28 18:48:24 +0000 sci-ml/ollama: add 0.6.6 Closes: https://bugs.gentoo.org/954399 Signed-off-by: Paul Zander <negril.nx+gentoo@gmail.com> sci-ml/ollama/Manifest | 2 + sci-ml/ollama/ollama-0.6.6.ebuild | 287 ++++++++++++++++++++++++++++++++++++++ sci-ml/ollama/ollama-9999.ebuild | 3 + 3 files changed, 292 insertions(+)
