diff --git a/hardware/gpu/cuda-common.nix b/hardware/gpu/cuda-common.nix
index 4aa9df4..075745f 100644
--- a/hardware/gpu/cuda-common.nix
+++ b/hardware/gpu/cuda-common.nix
@@ -72,7 +72,7 @@ lib.mkMerge [
   # this works, but you have to use `--device nvidia.com/gpu=all` rather than --gpus
   hardware.nvidia-container-toolkit.enable = lib.mkDefault true;
 
-  services.ollama.acceleration = lib.mkDefault "cuda";
+  # services.ollama.acceleration = lib.mkDefault "cuda";
 
   # only do these per-host
 
diff --git a/hardware/gpu/rocm.nix b/hardware/gpu/rocm.nix
index f0454ee..6ccacda 100644
--- a/hardware/gpu/rocm.nix
+++ b/hardware/gpu/rocm.nix
@@ -21,8 +21,7 @@
 
       nixpkgs.config.rocmSupport = true;
 
-      services.ollama.acceleration = lib.mkDefault "rocm";
-
+      # services.ollama.acceleration = lib.mkDefault "rocm";
       hardware.amdgpu.opencl.enable = lib.mkDefault true;
 
       # https://libreddit.noximilien.pbsds.net/r/archlinux/comments/nih9c9/amdgpu_vs_modesetting_in_current_510_kernels_xorg/