Add opencl release memory

ggml-opencl.cpp

@@ -662,6 +662,15 @@ static void ggml_cl_pool_free(cl_mem mem, size_t size) {
     clReleaseMemObject(mem);
 }
 
+void ggml_cl_data_free(const struct ggml_tensor* tensor) {
+    if (tensor->backend != GGML_BACKEND_GPU) {
+        return;
+    }
+
+    cl_mem mem = (cl_mem)tensor->data;
+    clReleaseMemObject(mem);
+}
+
 static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t offset, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cl_event* ev) {
     cl_int err;
     const uint64_t ne0 = src->ne[0];

ggml-opencl.h

@@ -16,6 +16,8 @@ void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor
 void * ggml_cl_host_malloc(size_t size);
 void   ggml_cl_host_free(void * ptr);
 
+void ggml_cl_data_free(const struct ggml_tensor* tensor);
+
 void ggml_cl_transform_tensor(struct ggml_tensor * tensor);
 void ggml_cl_load_data(const char * fname, struct ggml_tensor * tensor, size_t offset);
 

llama.cpp

@@ -211,6 +211,11 @@ struct llama_model {
             ggml_cuda_free_data(tensors_by_name[i].second);
         }
 #endif // GGML_USE_CUBLAS
+#if defined(GGML_USE_CLBLAST)
+        for (size_t i = 0; i < tensors_by_name.size(); ++i) {
+            ggml_cl_data_free(tensors_by_name[i].second);
+        }
+#endif // GGML_USE_CLBLAST
     }
 };