ollama: model store to /mnt/data, gfx1151 rocBLAS workaround, load tuning

modules/ollama.nix

@@ -7,10 +7,25 @@
     rocmOverrideGfx = "11.5.1";  # gfx1151, Krackan Point RDNA 3.5
     host = "127.0.0.1";
     port = 11434;
+    # Model store on plain-ext4 /mnt/data: model blobs are public data, so
+    # they don't need to occupy the 512G LUKS root or pay dm-crypt on reads.
+    # Static user instead of the default DynamicUser — a random dynamic UID
+    # can't own a persistent store outside /var/lib.
+    user = "ollama";
+    group = "ollama";
+    home = "/mnt/data/ollama";
     environmentVariables = {
       HSA_OVERRIDE_GFX_VERSION = "11.5.1";
+      # gfx1151 (RDNA 3.5) lacks tuned hipBLASLt kernels in current ROCm; force
+      # the stable rocBLAS GEMM path to avoid crashes/garbage output on this iGPU.
+      ROCBLAS_USE_HIPBLASLT = "0";
       OLLAMA_FLASH_ATTENTION = "1";
       OLLAMA_KV_CACHE_TYPE = "q8_0";
+      # Keep the loaded model resident: every eviction re-pays a multi-minute
+      # 20+ GiB disk read (client keep_alive can still override per-request).
+      OLLAMA_KEEP_ALIVE = "-1";
+      # Default 5m load timeout kills large-model loads under memory pressure.
+      OLLAMA_LOAD_TIMEOUT = "15m";
     };
     loadModels = [ "qwen3:14b" ];
   };