From 3c3372cdb3a0d85e194620eafbb0904de8aa12a7 Mon Sep 17 00:00:00 2001
From: "Peter H. Hoeg" <peter@hoeg.com>
Date: Sat, 9 Aug 2025 22:54:10 +0200
Subject: [PATCH 1/2] llama-cpp: 5985 -> 6123

---
 pkgs/by-name/ll/llama-cpp/disable_bfloat16.patch | 15 +++++++++++++++
 pkgs/by-name/ll/llama-cpp/package.nix            | 12 +++++++-----
 2 files changed, 22 insertions(+), 5 deletions(-)
 create mode 100644 pkgs/by-name/ll/llama-cpp/disable_bfloat16.patch

diff --git a/pkgs/by-name/ll/llama-cpp/disable_bfloat16.patch b/pkgs/by-name/ll/llama-cpp/disable_bfloat16.patch
new file mode 100644
index 000000000000..38388e2b82dc
--- /dev/null
+++ b/pkgs/by-name/ll/llama-cpp/disable_bfloat16.patch
@@ -0,0 +1,15 @@
+index b97e7bf9..9cdc62a4 100644
+--- a/ggml/src/ggml-vulkan/CMakeLists.txt
++++ b/ggml/src/ggml-vulkan/CMakeLists.txt
+@@ -70,11 +70,5 @@ if (Vulkan_FOUND)
+         "GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT"
+     )
+
+-    test_shader_extension_support(
+-        "GL_EXT_bfloat16"
+-        "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp"
+-        "GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT"
+-    )
+-
+     target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
+     target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
diff --git a/pkgs/by-name/ll/llama-cpp/package.nix b/pkgs/by-name/ll/llama-cpp/package.nix
index c0fd0a168b75..8d04c1afd41c 100644
--- a/pkgs/by-name/ll/llama-cpp/package.nix
+++ b/pkgs/by-name/ll/llama-cpp/package.nix
@@ -72,13 +72,13 @@ let
 in
 effectiveStdenv.mkDerivation (finalAttrs: {
   pname = "llama-cpp";
-  version = "5985";
+  version = "6123";
 
   src = fetchFromGitHub {
     owner = "ggml-org";
     repo = "llama.cpp";
     tag = "b${finalAttrs.version}";
-    hash = "sha256-OoV/p4Es/X/xQW7PpDLq5YLVYjieIE5+1itvtJECH54=";
+    hash = "sha256-4kqbKGPPOkOkHXA4IeLuj/0P5jpqtGlGuVKeUD4UhZY=";
     leaveDotGit = true;
     postFetch = ''
       git -C "$out" rev-parse --short HEAD > $out/COMMIT
@@ -86,6 +86,8 @@ effectiveStdenv.mkDerivation (finalAttrs: {
     '';
   };
 
+  patches = lib.optionals vulkanSupport [ ./disable_bfloat16.patch ];
+
   postPatch = ''
     # Workaround for local-ai package which overrides this package to an older llama-cpp
     if [ -f ./ggml/src/ggml-metal.m ]; then
@@ -139,11 +141,11 @@ effectiveStdenv.mkDerivation (finalAttrs: {
   ++ optionals cudaSupport [
     (cmakeFeature "CMAKE_CUDA_ARCHITECTURES" cudaPackages.flags.cmakeCudaArchitecturesString)
   ]
-  ++ optionals rocmSupport ([
+  ++ optionals rocmSupport [
     (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.clr.hipClangPath}/clang++")
     # TODO: this should become `clr.gpuTargets` in the future.
     (cmakeFeature "CMAKE_HIP_ARCHITECTURES" rocmPackages.rocblas.amdgpu_targets)
-  ])
+  ]
   ++ optionals metalSupport [
     (cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
     (cmakeBool "LLAMA_METAL_EMBED_LIBRARY" true)
@@ -186,6 +188,6 @@ effectiveStdenv.mkDerivation (finalAttrs: {
     ];
     platforms = platforms.unix;
     badPlatforms = optionals (cudaSupport || openclSupport) lib.platforms.darwin;
-    broken = (metalSupport && !effectiveStdenv.hostPlatform.isDarwin);
+    broken = metalSupport && !effectiveStdenv.hostPlatform.isDarwin;
   };
 })

From 58d9978d685190d592bd835cd2c7941a85978128 Mon Sep 17 00:00:00 2001
From: "Peter H. Hoeg" <peter@hoeg.com>
Date: Sat, 9 Aug 2025 22:55:56 +0200
Subject: [PATCH 2/2] llama-cpp-vulkan: init at 6123

---
 pkgs/by-name/ll/llama-cpp-vulkan/package.nix | 3 +++
 1 file changed, 3 insertions(+)
 create mode 100644 pkgs/by-name/ll/llama-cpp-vulkan/package.nix

diff --git a/pkgs/by-name/ll/llama-cpp-vulkan/package.nix b/pkgs/by-name/ll/llama-cpp-vulkan/package.nix
new file mode 100644
index 000000000000..39ff8a5dfdfd
--- /dev/null
+++ b/pkgs/by-name/ll/llama-cpp-vulkan/package.nix
@@ -0,0 +1,3 @@
+{ llama-cpp }:
+
+llama-cpp.override { vulkanSupport = true; }