Nix fixes

.devops/nix/apps.nix

@@ -11,6 +11,9 @@
             "llama-server"
             "quantize"
             "train-text-from-scratch"
+            "llama-bench"
+            "imatrix"
+            "perplexity"
           ];
           mkApp = name: {
             type = "app";

.devops/nix/package.nix

@@ -30,8 +30,15 @@
   useMpi ? false, # Increases the runtime closure size by ~700M
   useOpenCL ? false,
   useRocm ? config.rocmSupport,
+  rocmTargets ? "gfx1030",
   useVulkan ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
+  llamaNative ? true,
+  llamaLTO ? true,
+  llamaAVX ? true,
+  llamaAVX2 ? true,
+  llamaFMA ? true,
+  llamaF16C ? true,
 
   # It's necessary to consistently use backendStdenv when building with CUDA support,
   # otherwise we get libstdc++ errors downstream.
@@ -205,7 +212,12 @@ effectiveStdenv.mkDerivation (
 
     cmakeFlags =
       [
-        (cmakeBool "LLAMA_NATIVE" false)
+        (cmakeBool "LLAMA_NATIVE" llamaNative)
+        (cmakeBool "LLAMA_LTO"    llamaLTO)
+        (cmakeBool "LLAMA_AVX"    llamaAVX)
+        (cmakeBool "LLAMA_AVX2"   llamaAVX2)
+        (cmakeBool "LLAMA_FMA"    llamaFMA)
+        (cmakeBool "LLAMA_F16C"   llamaF16C)
         (cmakeBool "LLAMA_BUILD_SERVER" true)
         (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
@@ -234,7 +246,7 @@ effectiveStdenv.mkDerivation (
         # in https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
         # and select the line that matches the current nixpkgs version of rocBLAS.
         # Should likely use `rocmPackages.clr.gpuTargets`.
-        "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
+        "-DAMDGPU_TARGETS=${rocmTargets}"
       ]
       ++ optionals useMetalKit [
         (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")