diff --git a/Makefile b/Makefile
index 284b0f1..4a676f1 100644
--- a/Makefile
+++ b/Makefile
@@ -215,9 +215,9 @@ endif
 
 ifdef WHISPER_CUBLAS
 	ifeq ($(shell expr $(NVCC_VERSION) \>= 11.6), 1)
-		CUDA_ARCH_FLAG=native
+		CUDA_ARCH_FLAG ?= native
 	else
-		CUDA_ARCH_FLAG=all
+		CUDA_ARCH_FLAG ?= all
 	endif
 
 	CFLAGS      += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
diff --git a/whisper.cpp b/whisper.cpp
index ba867b0..59d5cff 100644
--- a/whisper.cpp
+++ b/whisper.cpp
@@ -1060,7 +1060,7 @@ static ggml_backend_t whisper_backend_init(const whisper_context_params & params
 #ifdef GGML_USE_CUBLAS
     if (params.use_gpu && ggml_cublas_loaded()) {
         WHISPER_LOG_INFO("%s: using CUDA backend\n", __func__);
-        backend_gpu = ggml_backend_cuda_init(0);
+        backend_gpu = ggml_backend_cuda_init(params.gpu_device);
         if (!backend_gpu) {
             WHISPER_LOG_ERROR("%s: ggml_backend_cuda_init() failed\n", __func__);
         }
@@ -3213,6 +3213,7 @@ int whisper_ctx_init_openvino_encoder(
 struct whisper_context_params whisper_context_default_params() {
     struct whisper_context_params result = {
         /*.use_gpu    =*/ true,
+        /*.gpu_device =*/ 0,
     };
     return result;
 }
diff --git a/whisper.h b/whisper.h
index 3143cea..d571a12 100644
--- a/whisper.h
+++ b/whisper.h
@@ -86,6 +86,7 @@ extern "C" {
 
     struct whisper_context_params {
         bool  use_gpu;
+        int   gpu_device;  // CUDA device
     };
 
     typedef struct whisper_token_data {