diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 69b081d6..ff44bb99 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -115,6 +115,7 @@ option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"
 option(GGML_IQK_MUL_MAT                     "ggml: use optimized iqk matrix multiplications"  ON)
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
+option(GGML_CUDA_LTO                        "ggml: use CUDA LTO"                              OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
 option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index aab44d7c..78da50f4 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -315,6 +315,10 @@ if (GGML_CUDA)
         find_package(MUSAToolkit)
         set(CUDAToolkit_FOUND ${MUSAToolkit_FOUND})
     else()
+        if (GGML_CUDA_LTO)
+            set(CMAKE_CUDA_SEPARABLE_COMPILATION ON)
+            set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)
+        endif()
         find_package(CUDAToolkit)
         set(CMAKE_CUDA_USE_RESPONSE_FILE_FOR_INCLUDES 0)
         set(CMAKE_CUDA_USE_RESPONSE_FILE_FOR_LIBRARIES 0)
@@ -344,6 +348,9 @@ if (GGML_CUDA)
             set(CMAKE_CUDA_COMPILER ${MUSAToolkit_MCC_EXECUTABLE})
         else()
             enable_language(CUDA)
+            if (GGML_CUDA_LTO)
+                set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -dlto")
+            endif()
         endif()
 
         file(GLOB   GGML_HEADERS_CUDA "ggml-cuda/*.cuh")