From b4443ea1317506ba795dff739815404a9fc23580 Mon Sep 17 00:00:00 2001
From: emmatherock <emanuelxoda@gmail.com>
Date: Mon, 31 Aug 2026 03:59:06 -0300
Subject: [PATCH] feat(homelab): add Ollama service with ROCm

---
 hosts/miku-homelab/default.nix |  8 ++++++++
 modules/services/ollama.nix    | 35 ++++++++++++++++++++++++++++++++++
 2 files changed, 43 insertions(+)
 create mode 100644 modules/services/ollama.nix

diff --git a/hosts/miku-homelab/default.nix b/hosts/miku-homelab/default.nix
index 4ee7778..135d8a9 100644
--- a/hosts/miku-homelab/default.nix
+++ b/hosts/miku-homelab/default.nix
@@ -5,6 +5,7 @@
     ./hardware-configuration.nix
     ../../modules/profiles/desktop.nix
     ../../modules/profiles/server.nix
+    ../../modules/services/ollama.nix
     ../../modules/services/samba.nix
     ../../modules/system/home-profile.nix
     ../../modules/system/nix-ld.nix
@@ -157,5 +158,12 @@
     experimental-features = [ "nix-command" "flakes" ];
   };
 
+  zramSwap = {
+    enable = true;
+    algorithm = "zstd";
+    memoryPercent = 50;
+    priority = 100;
+  };
+
   system.stateVersion = "25.11";
 }
diff --git a/modules/services/ollama.nix b/modules/services/ollama.nix
new file mode 100644
index 0000000..0f272b4
--- /dev/null
+++ b/modules/services/ollama.nix
@@ -0,0 +1,35 @@
+{ pkgs, ... }:
+
+{
+  services.ollama = {
+    enable = true;
+    # `services.ollama.acceleration` does not exist (that option only lives on
+    # services.tabby); GPU backend selection for ollama is done by package choice.
+    package = pkgs.ollama-rocm;
+
+    # Same "don't expose services directly" pattern as the rest of the homelab:
+    # bind to localhost only. If this needs to be reachable from other devices
+    # later, put it behind Traefik + Headscale like the other services instead
+    # of opening it up directly.
+    host = "127.0.0.1";
+    port = 11434;
+
+    environmentVariables = {
+      # Ollama is on-demand by design: it loads a model into VRAM on the first
+      # request and unloads it after this much idle time. 5m is already the
+      # ollama default, set explicitly so it's not an implicit behavior.
+      OLLAMA_KEEP_ALIVE = "5m";
+    };
+
+    # RDNA4 (gfx1200/gfx1201, incl. the RX 9060 XT) has official ROCm support
+    # since ROCm 6.4.4. Verified on this host (2026-08-30) with rocmPackages.clr
+    # 7.2.3: `rocminfo` correctly reports "Name: gfx1200" and `rocm-smi
+    # --showmeminfo vram` reports the full 16GiB, so the "detected with 0 VRAM,
+    # falls back to CPU" issue does not reproduce here and rocmOverrideGfx is
+    # not needed. If a future nixpkgs bump regresses this, re-check with:
+    #   nix shell nixpkgs#rocmPackages.rocminfo -c rocminfo | grep -i gfx
+    #   nix shell nixpkgs#rocmPackages.rocm-smi -c rocm-smi --showmeminfo vram
+    # and only then set rocmOverrideGfx = "12.0.0"; (gfx1200) if VRAM reports 0.
+    # rocmOverrideGfx = "12.0.0";
+  };
+}