diff --git a/CMakeLists.txt b/CMakeLists.txt
index 24ae538a..2f93ed48 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -114,6 +114,10 @@ set(ONEBIT_SERVE_TEST_GGUF "" CACHE FILEPATH "GGUF the serve_e2e tests load")
 # --device cpu the same build with no GPU layers. The build has no Vulkan or ROCm backend
 # (RFC #213 stage 3): what HRX does not run, Lemonade serves with its own llamacpp backends.
 option(ONEBIT_HRX "Build the HRX llama-server and wire it into Lemonade" OFF)
+set(ONEBIT_GPU_PRIVATE "" CACHE PATH "A gpu-kernels checkout whose addons/hip-hrx HIP kernels are built into the HRX llama.cpp (needs ONEBIT_HRX)")
+if(ONEBIT_GPU_PRIVATE AND NOT ONEBIT_HRX)
+    message(FATAL_ERROR "ONEBIT_GPU_PRIVATE needs ONEBIT_HRX=ON")
+endif()
 if(ONEBIT_HRX)
     include(cmake/hrx.cmake)
     add_dependencies(onebit llama_hrx)
diff --git a/cmake/hrx.cmake b/cmake/hrx.cmake
index b330658c..06c48891 100644
--- a/cmake/hrx.cmake
+++ b/cmake/hrx.cmake
@@ -51,6 +51,25 @@ endif()
 set(ONEBIT_HRX_LIBHSA "${_hrx_libhsa_first}" CACHE FILEPATH "TheRock HSA runtime HRX loads (IREE_HAL_AMDGPU_LIBHSA_PATH)")
 message(STATUS "ONEBIT_HRX: HSA runtime ${ONEBIT_HRX_LIBHSA}")
 
+# Private GPU kernels (docs/hrx.md, "Private HIP kernels"): a checkout of the private gpu-kernels
+# repository whose addons/hip-hrx is built into the HRX llama.cpp (its GGML_HRX_HIP_ADDON_DIR hook).
+set(_hrx_gpu_private_args)
+if(ONEBIT_GPU_PRIVATE)
+    if(NOT IS_DIRECTORY "${ONEBIT_GPU_PRIVATE}/addons/hip-hrx/kernels")
+        message(FATAL_ERROR "ONEBIT_GPU_PRIVATE=${ONEBIT_GPU_PRIVATE} has no addons/hip-hrx/kernels")
+    endif()
+    set(_hrx_hip_cmake "${CMAKE_SOURCE_DIR}/third_party/llama.cpp/ggml/src/ggml-hrx/hip/ggml-hrx-hip.cmake")
+    set(_hrx_addon_hook)
+    if(EXISTS "${_hrx_hip_cmake}")
+        file(STRINGS "${_hrx_hip_cmake}" _hrx_addon_hook REGEX "GGML_HRX_HIP_ADDON_DIR")
+    endif()
+    if(NOT _hrx_addon_hook)
+        message(FATAL_ERROR "ONEBIT_GPU_PRIVATE: third_party/llama.cpp has no GGML_HRX_HIP_ADDON_DIR hook (needs the HIP plumbing)")
+    endif()
+    list(APPEND _hrx_gpu_private_args "-DGGML_HRX_HIP_ADDON_DIR=${ONEBIT_GPU_PRIVATE}/addons/hip-hrx")
+    message(STATUS "ONEBIT_HRX: private HIP kernels from ${ONEBIT_GPU_PRIVATE}/addons/hip-hrx")
+endif()
+
 set(ONEBIT_HRX_SERVER "${CMAKE_BINARY_DIR}/hrx/llama/bin/llama-server")
 ExternalProject_Add(llama_hrx
     SOURCE_DIR ${CMAKE_SOURCE_DIR}/third_party/llama.cpp
@@ -63,6 +82,7 @@ ExternalProject_Add(llama_hrx
         -DHRX_SOURCE_DIR=${CMAKE_SOURCE_DIR}/third_party/hrx-system
         -DGGML_HRX=ON -DGGML_VULKAN=OFF -DGGML_CUDA=OFF -DGGML_HIP=OFF -DGGML_NATIVE=ON -DGGML_CPU=ON
         -DLLAMA_BUILD_SERVER=ON -DLLAMA_CURL=OFF
+        ${_hrx_gpu_private_args}
     BUILD_COMMAND ${CMAKE_COMMAND} --build <BINARY_DIR> --target llama-server llama-bench
     INSTALL_COMMAND ""
     BUILD_ALWAYS ON
diff --git a/docs/hrx.md b/docs/hrx.md
index a623dee0..5f170aa2 100644
--- a/docs/hrx.md
+++ b/docs/hrx.md
@@ -507,6 +507,16 @@ cmake --build build --target onebit
   build path it searches `/opt/rocm-therock`. To run
   llama-server or llama-bench by hand, export that variable yourself.
 
+### Private HIP kernels
+
+Our own HIP kernels for HRX are closed source, like the NPU kernels: they live in the private
+`1bit-MONSTER/gpu-kernels` repository (`addons/hip-hrx`). The llama.cpp fork keeps only the
+public HIP-through-HRX plumbing (`ggml/src/ggml-hrx/hip/`). `-DONEBIT_GPU_PRIVATE=<gpu-kernels checkout>`
+(with `-DONEBIT_HRX=ON`) passes `GGML_HRX_HIP_ADDON_DIR=<checkout>/addons/hip-hrx` to the HRX
+llama.cpp build, which compiles those kernels and matchers in; it needs a llama.cpp pin with
+that hook. Without the option the build is unchanged. To drop the add-on from an existing
+build directory, delete `build/hrx/llama` (its CMake cache keeps the setting).
+
 ## Verified (2026-09-23, llama.cpp `f1a0aca`, hrx-system `51b1739`)
 
 `ctest` passed, re-run after the first daily bump (#13). At the time the check