diff --git a/CMakeLists.txt b/CMakeLists.txt
index 143b6894f..fa5f28f18 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -438,6 +438,7 @@ target_include_directories(ggml PUBLIC . ./ggml/include ./ggml/src ./ggml/src/gg
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
 set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
+target_compile_options(ggml PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-use_fast_math -extended-lambda>)
 
 add_library(ggml_v1
             otherarch/ggml_v1.c
diff --git a/Makefile b/Makefile
index bccd9a95b..c152634f5 100644
--- a/Makefile
+++ b/Makefile
@@ -183,7 +183,7 @@ ifdef LLAMA_CUBLAS
 	CUBLAS_OBJS += $(patsubst %.cu,%.o,$(filter-out ggml/src/ggml-cuda/ggml-cuda.cu, $(wildcard ggml/src/ggml-cuda/*.cu)))
 	CUBLAS_OBJS += $(OBJS_CUDA_TEMP_INST)
 	NVCC      = nvcc
-	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
+	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math -extended-lambda
 
 ifdef LLAMA_ADD_CONDA_PATHS
 	CUBLASLD_FLAGS += -Lconda/envs/linux/lib -Lconda/envs/linux/lib/stubs