ggml : disable CUDA graphs for non-llama.cpp projects

ggerganov · Jun 26, 2024 · ceb7736 · ceb7736
1 parent dc8cc2d
commit ceb7736
Showing 1 changed file with 10 additions and 1 deletion.
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
@@ -295,7 +295,16 @@ if (GGML_CUDA)
 
  list(APPEND GGML_CDEF_PUBLIC GGML_USE_CUDA)
 
- add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+ # TODO: for now CUDA graphs should be used only with llama.cpp
+ # https://github.com/ggerganov/whisper.cpp/issues/2258
+ message(STATUS "CMAKE_PROJECT_NAME: ${CMAKE_PROJECT_NAME}")
+ if (CMAKE_PROJECT_NAME STREQUAL "llama.cpp")
+ add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+ message(STATUS "GGML_CUDA_USE_GRAPHS enabled")
+ else()
+ message(STATUS "GGML_CUDA_USE_GRAPHS disabled")
+ endif()
+
  add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
  add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
  add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})