feat(packages): add llama-cpp b9645 for Strix Halo (ROCm + Vulkan)

Pin upstream llama.cpp b9645 and build it against the nixpkgs-latest
input (the flake's pinned nixpkgs only carries the pre-tools/ui b8983).
ROCm deps come from rocmPackages.gfx1151, so clr/rocBLAS/hipBLAS are
built for gfx1151 only; clr's own build is arch-independent, so its
store path stays substitutable. Vulkan on, CPU variants with znver5 /
AVX-512, web UI and npm toolchain dropped.

Ported from nixos-config's llm-engine.nix with two fixes: npmConfigHook
hard-fails when npmDeps is null, so nodejs and the hook are stripped
from nativeBuildInputs, and CMAKE_HIP_FLAGS is appended to
cmakeFlagsArray because the cmake hook word-splits plain cmakeFlags.

Verified on the Radeon 8060S (gfx1151): llama-bench loads both ROCm and
Vulkan backends and completes pp16/tg8 runs.

packages/llama-cpp/default.nix

@@ -0,0 +1,17 @@
+{
+  pkgs,
+  inputs,
+  ...
+}:
+# llama.cpp b9645 is newer than the flake's pinned nixpkgs (llama-cpp b8983,
+# pre-tools/ui) and needs its UI/ROCm layout, so build against the
+# nixpkgs-latest input — same pattern as mcp-gateway/kubernetes-mcp-server.
+let
+  latestPkgs = inputs.nixpkgs-latest.legacyPackages.${pkgs.stdenv.hostPlatform.system};
+in
+latestPkgs.callPackage ./package.nix {
+  # Strix Halo: build the ROCm dependencies (clr, rocBLAS, hipBLAS) for gfx1151
+  # only instead of every supported arch. clr's own build is arch-independent,
+  # so its store path is unchanged and stays substitutable.
+  rocmPackages = latestPkgs.rocmPackages.gfx1151;
+}

packages/llama-cpp/package.nix

@@ -0,0 +1,117 @@
+{
+  lib,
+  # Build skeleton: nixpkgs' own llama-cpp derivation (b-tag layout, cmake/npm
+  # plumbing), retargeted at the pinned tag below.
+  llama-cpp,
+  fetchFromGitHub,
+  rocmPackages,
+  # nixpkgs-latest's nodejs/npm hooks: referenced only to strip them from
+  # nativeBuildInputs (the npm/web UI build is disabled below).
+  nodejs_latest,
+  npmHooks,
+  # Strix Halo (Radeon 8060S). clr/rocBLAS/hipBLAS come prefixed for this arch
+  # by the scoped rocmPackages in ./default.nix.
+  rocmGpuTargets ? [ "gfx1151" ],
+}:
+
+let
+  # Upstream release tag. Bump with:
+  #   nix flake prefetch github:ggml-org/llama.cpp/b<NNNN>
+  # then update buildNumber and the src hash.
+  buildNumber = "9645";
+
+  # HIP flags from the llm-engine.nix bring-up config. The -I paths for
+  # hipBLASLt/rocWMMA are dropped because this tag's HIP backend links
+  # hipBLAS/rocBLAS only; GGML_HIPBLASLT and GGML_CUDA_ENABLE_UNIFIED_MEMORY
+  # are also unused by b9645 (cmake warns and ignores them).
+  hipFlags = lib.concatStringsSep " " [
+    "-mllvm"
+    "-amdgpu-early-inline-all=true"
+    "-mllvm"
+    "-amdgpu-function-calls=false"
+    "-mprefer-vector-width=512"
+    "-famd-opt"
+    "-mllvm"
+    "-inline-threshold=600"
+    "-mllvm"
+    "-unroll-threshold=150"
+  ];
+in
+(llama-cpp.override {
+  inherit rocmPackages rocmGpuTargets;
+  rocmSupport = true;
+  vulkanSupport = true;
+  cudaSupport = false;
+}).overrideAttrs
+  (old: {
+    version = buildNumber;
+
+    src = fetchFromGitHub {
+      owner = "ggml-org";
+      repo = "llama.cpp";
+      tag = "b${buildNumber}";
+      # Tarball hash (this tag has no submodules).
+      hash = "sha256-ntRwfI2/yVqi3QjQfO0ZajMFSD8r/6XPiuy0X2BhwVk=";
+    };
+
+    # The embedded web UI is disabled via cmakeFlags, so none of the npm
+    # machinery is needed: no dependency tree, and npmConfigHook hard-fails
+    # when npmDeps is null, so it (and nodejs) must come off the inputs.
+    npmDeps = null;
+    npmRoot = null;
+    npmDepsHash = null;
+    nativeBuildInputs = builtins.filter (x: x != nodejs_latest && x != npmHooks.npmConfigHook) (
+      old.nativeBuildInputs or [ ]
+    );
+
+    # The release tarball has no .git; upstream wants the build id in COMMIT.
+    postPatch = ''
+      echo ${buildNumber} > COMMIT
+    '';
+
+    # Replaces the base derivation's npm build step. CMAKE_HIP_FLAGS holds
+    # several space-separated flags, so it must be appended as an array element:
+    # plain cmakeFlags scalars are word-split by the cmake hook.
+    preConfigure = ''
+      prependToVar cmakeFlags "-DLLAMA_BUILD_COMMIT:STRING=$(cat COMMIT)"
+      cmakeFlagsArray+=("-DCMAKE_HIP_FLAGS=${hipFlags}")
+    '';
+
+    cmakeFlags =
+      builtins.filter (f: !(builtins.isString f && builtins.match ".*LLAMA_BUILD_NUMBER.*" f != null)) (
+        old.cmakeFlags or [ ]
+      )
+      ++ [
+        "-DCMAKE_C_FLAGS=-march=znver5"
+        "-DCMAKE_CXX_FLAGS=-march=znver5"
+        "-DGGML_CPU=ON"
+        "-DGGML_AVX=ON"
+        "-DGGML_AVX_VNNI=ON"
+        "-DGGML_AVX2=ON"
+        "-DGGML_BMI2=ON"
+        "-DGGML_AVX512=ON"
+        "-DGGML_AVX512_VNNI=ON"
+        "-DGGML_AVX512_VBMI=ON"
+        "-DGGML_AVX512_BF16=ON"
+        "-DGGML_FMA=ON"
+        "-DGGML_F16C=ON"
+        "-DHIP_PLATFORM=amd"
+        "-DGGML_CUDA_FORCE_CUBLAS=OFF"
+        "-DGGML_HIPBLASLT=ON"
+        "-DGGML_HIP_NO_VMM=ON"
+        "-DGGML_HIP_GRAPHS=ON"
+        "-DGGML_CUDA_ENABLE_UNIFIED_MEMORY=ON"
+        "-DGGML_HIP_ROCWMMA_FATTN=OFF"
+        "-DGGML_HIP_MMQ_MFMA=ON"
+        "-DGGML_CUDA_FA_ALL_QUANTS=ON"
+        "-DCMAKE_BUILD_TYPE=Release"
+        "-DLLAMA_BUILD_NUMBER:STRING=${buildNumber}"
+        (lib.cmakeBool "LLAMA_BUILD_UI" false)
+        (lib.cmakeBool "LLAMA_USE_PREBUILT_UI" false)
+      ];
+
+    # The base's update script tracks nixpkgs' llama-cpp attr, not this pin.
+    passthru = builtins.removeAttrs (old.passthru or { }) [ "updateScript" ] // {
+      category = "AI Inference";
+    };
+  })