add free for context

examples/embd_input/embd_input.h

@@ -23,6 +23,7 @@ bool eval_id(struct MyModel* mymodel, int id);
 bool eval_string(struct MyModel* mymodel, const char* str);
 const char* sampling(struct MyModel* mymodel);
 llama_token sampling_id(struct MyModel* mymodel);
+void free_mymodel(struct MyModel* mymodel);
 
 }
 

examples/embd_input/embd_input.py

@@ -17,6 +17,8 @@ class MyModel:
         args_c = (c_char_p * argc)(*c_str)
         self.model = c_void_p(libc.create_mymodel(argc, args_c))
 #         print("self.model", self.model)
+    def __del__(self):
+        libc.free_mymodel(self.model)
 
     def eval_float(self, x):
         libc.eval_float(self.model, x.astype(np.float32).ctypes.data_as(POINTER(c_float)), x.shape[0])

examples/embd_input/embd_input_lib.cpp

@@ -101,6 +101,13 @@ struct MyModel* create_mymodel(int argc, char ** argv) {
     return ret;
 }
 
+void free_mymodel(struct MyModel* mymodel) {
+    llama_context* ctx = mymodel->ctx;
+    llama_print_timings(ctx);
+    llama_free(ctx);
+    delete mymodel;
+}
+
 
 bool eval_float(void* model, float* input, int N){
     MyModel* mymodel = (MyModel* )model;

examples/embd_input/embd_input_test.cpp

@@ -27,5 +27,6 @@ int main(int argc, char** argv) {
         // eval_id(mymodel, id);
     }
     printf("\n");
+    free_mymodel(mymodel);
     return 0;
 }