diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h
index 3a2b2d24ab..f58f914e70 100644
--- a/ggml/src/ggml-backend-impl.h
+++ b/ggml/src/ggml-backend-impl.h
@@ -221,6 +221,9 @@ extern "C" {
         // (optional) get a pointer to a function in the backend
         // backends can add custom functions that are not part of the standard ggml-backend interface
         void * (*get_proc_address)(ggml_backend_reg_t reg, const char * name);
+
+        // (optional) free backend reg (before unloading the dynamic library)
+        void (*free)(ggml_backend_reg_t reg);
     };
 
     struct ggml_backend_reg {
@@ -236,8 +239,6 @@ extern "C" {
     // Optional: obtain a score for the backend based on the system configuration
     // Higher scores are preferred, 0 means the backend is not supported in the current system
     typedef int                (*ggml_backend_score_t)(void);
-    // Optional: release all resources held by the backend (before unloading the dynamic library)
-    typedef void               (*ggml_backend_reg_free_t)(ggml_backend_reg_t reg);
 
 #ifdef GGML_BACKEND_DL
 #    ifdef __cplusplus
@@ -255,13 +256,6 @@ extern "C" {
             int ggml_backend_score(void) {                 \
                 return score_fn();                         \
             }
-#        define GGML_BACKEND_DL_FREE_IMPL(free_fn)                                \
-            extern "C" {                                                          \
-            GGML_BACKEND_API void ggml_backend_reg_free(ggml_backend_reg_t reg); \
-            }                                                                     \
-            void ggml_backend_reg_free(ggml_backend_reg_t reg) {                  \
-                free_fn(reg);                                                     \
-            }
 #    else
 #        define GGML_BACKEND_DL_IMPL(reg_fn)                              \
             GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void);  \
@@ -273,16 +267,10 @@ extern "C" {
             int                  ggml_backend_score(void) { \
                 return score_fn();                          \
             }
-#        define GGML_BACKEND_DL_FREE_IMPL(free_fn)                                \
-            GGML_BACKEND_API void ggml_backend_reg_free(ggml_backend_reg_t reg); \
-            void                  ggml_backend_reg_free(ggml_backend_reg_t reg) { \
-                free_fn(reg);                                                     \
-            }
 #    endif
 #else
 #    define GGML_BACKEND_DL_IMPL(reg_fn)
 #    define GGML_BACKEND_DL_SCORE_IMPL(score_fn)
-#    define GGML_BACKEND_DL_FREE_IMPL(free_fn)
 #endif
 
 #ifdef  __cplusplus
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index 15ba6c8a7b..416bd95eed 100644
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -278,11 +278,8 @@ struct ggml_backend_registry {
             devices.end());
 
         // if backend supports free(), use it
-        if (it->handle) {
-            auto * free_fn = (ggml_backend_reg_free_t) dl_get_sym(it->handle.get(), "ggml_backend_reg_free");
-            if (free_fn) {
-                free_fn(reg);
-            }
+        if (reg->iface.free) {
+            reg->iface.free(reg);
         }
 
         // remove backend
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 152cc16cdd..9e70fcb75c 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -5722,17 +5722,40 @@ static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, con
     return nullptr;
 }
 
+static bool ggml_backend_cuda_reg_initialized = false;
+
+static void ggml_backend_cuda_reg_free(ggml_backend_reg_t reg) {
+    ggml_backend_cuda_reg_context * ctx = (ggml_backend_cuda_reg_context *) reg->context;
+
+    for (ggml_backend_dev_t dev : ctx->devices) {
+        int device = ((ggml_backend_cuda_device_context *) dev->context)->device;
+        delete (ggml_backend_cuda_device_context *) dev->context;
+        delete dev;
+
+        cudaError_t err = cudaSetDevice(device);
+        if (err == cudaSuccess) {
+            cudaDeviceReset();
+        }
+    }
+
+    delete ctx;
+    reg->context = nullptr;
+
+    ggml_backend_cuda_reg_initialized = false;
+}
+
 static const ggml_backend_reg_i ggml_backend_cuda_reg_interface = {
     /* .get_name          = */ ggml_backend_cuda_reg_get_name,
     /* .get_device_count  = */ ggml_backend_cuda_reg_get_device_count,
     /* .get_device        = */ ggml_backend_cuda_reg_get_device,
     /* .get_proc_address  = */ ggml_backend_cuda_reg_get_proc_address,
+    /* .free              = */ ggml_backend_cuda_reg_free,
 };
 
 // backend registry
 ggml_backend_reg_t ggml_backend_cuda_reg() {
     static ggml_backend_reg reg;
-    static bool initialized = false;
+    bool & initialized = ggml_backend_cuda_reg_initialized;
 
     {
         static std::mutex mutex;
@@ -5779,23 +5802,6 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
     return ®
 }
 
-static void ggml_backend_cuda_reg_free(ggml_backend_reg_t reg) {
-    ggml_backend_cuda_reg_context * ctx = (ggml_backend_cuda_reg_context *) reg->context;
-
-    for (ggml_backend_dev_t dev : ctx->devices) {
-        int device = ((ggml_backend_cuda_device_context *) dev->context)->device;
-        delete (ggml_backend_cuda_device_context *) dev->context;
-        delete dev;
-
-        cudaError_t err = cudaSetDevice(device);
-        if (err == cudaSuccess) {
-            cudaDeviceReset();
-        }
-    }
-
-    delete ctx;
-}
-
 ggml_backend_t ggml_backend_cuda_init(int device) {
     if (device < 0 || device >= ggml_backend_cuda_get_device_count()) {
         GGML_LOG_ERROR("%s: invalid device %d\n", __func__, device);
@@ -5819,4 +5825,3 @@ ggml_backend_t ggml_backend_cuda_init(int device) {
 }
 
 GGML_BACKEND_DL_IMPL(ggml_backend_cuda_reg)
-GGML_BACKEND_DL_FREE_IMPL(ggml_backend_cuda_reg_free)