desko: ollama: fix/re-enable GPU acceleration

hosts/by-name/desko/default.nix

@@ -53,6 +53,5 @@
   # needed to use libimobiledevice/ifuse, for iphone sync
   services.usbmuxd.enable = true;
 
-  # TODO(2025-01-01): re-enable once rocm build is fixed: <https://github.com/NixOS/nixpkgs/pull/367695>
-  # hardware.amdgpu.opencl.enable = true;  # desktop (AMD's opencl implementation AKA "ROCM"); probably required for ollama
+  hardware.amdgpu.opencl.enable = true;  # desktop (AMD's opencl implementation AKA "ROCM"); probably required for ollama
 }

modules/services/ollama.nix

@@ -11,8 +11,13 @@
 #
 # models are defined e.g. here: <https://ollama.com/library/llama3.2:3b/blobs/dde5aa3fc5ff>
 #
-# TODO: enable ShellGPT integration?
-# - <https://github.com/TheR1D/shell_gpt?tab=readme-ov-file#docker--ollama>
+### to confirm GPU acceleration
+# grep `journalctl -u ollama` for:
+# - `looking for compatible GPUs`
+# and see if it says anything bad afterward like:
+# - `no compatible GPUs were discovered`
+# then run a model and check for:
+# `offloading <n> repeating layers to GPU`
 { config, lib, pkgs, ... }:
 let
   cfg = config.sane.services.ollama;
@@ -89,7 +94,7 @@ in
     # these acceleration settings are relevant to `desko`.
     services.ollama.acceleration = lib.mkIf config.hardware.amdgpu.opencl.enable "rocm";  # AMD GPU acceleration (achieves the same as `nixpkgs.config.rocmSupport = true` but just for ollama (the global toggle rebuilds the world))
     services.ollama.rocmOverrideGfx = "10.1.0";  #< `nix-shell -p "rocmPackages.rocminfo" --run "rocminfo" | grep "gfx"`  (e.g. gfx1010)
-    # services.ollama.environmentVariables.HCC_AMDGPU_TARGET = "gfx1010";  # seems to be unnecessary
+    services.ollama.environmentVariables.HCC_AMDGPU_TARGET = "gfx1010";
 
     users.groups.ollama = {};