diff --git a/example/mnist/linear.go b/example/mnist/linear.go
index 3e2cf97..e690fdc 100644
--- a/example/mnist/linear.go
+++ b/example/mnist/linear.go
@@ -66,9 +66,6 @@ func runLinear() {
 		 *       bs.ZeroGrad()
 		 *       loss.MustBackward()
 		 *
-		 *       // TODO: why `loss` need to print out to get updated?
-		 *       fmt.Printf("loss (epoch %v): %v\n", epoch, loss.MustToString(0))
-		 *
 		 *       ts.NoGrad(func() {
 		 *         ws.MustAdd_(ws.MustGrad().MustMul1(ts.FloatScalar(-1.0)))
 		 *         bs.MustAdd_(bs.MustGrad().MustMul1(ts.FloatScalar(-1.0)))
diff --git a/libtch/tensor.go b/libtch/tensor.go
index 10731e9..d48678e 100644
--- a/libtch/tensor.go
+++ b/libtch/tensor.go
@@ -221,6 +221,7 @@ func AtSaveMulti(tensors []Ctensor, tensor_names []string, ntensors int, filenam
 
 	cpointerSize := 4
 	cnamesPtr := (*[1 << 30]**C.char)(C.malloc(C.size_t(cpointerSize * len(tensor_names))))
+	defer C.free(unsafe.Pointer(cnamesPtr))
 	for i := 0; i < len(tensor_names); i++ {
 		cname := C.CString(tensor_names[i])
 		cnamesPtr[i] = &cname
diff --git a/tensor/tensor-generated-sample.go b/tensor/tensor-generated-sample.go
index 1ae247e..7574c03 100644
--- a/tensor/tensor-generated-sample.go
+++ b/tensor/tensor-generated-sample.go
@@ -583,6 +583,8 @@ func (ts Tensor) MustLogSoftmax(dim int64, dtype int32) (retVal Tensor) {
 
 func (ts Tensor) NllLoss(target Tensor) (retVal Tensor, err error) {
 	ptr := (*lib.Ctensor)(unsafe.Pointer(C.malloc(0)))
+	// NOTE: uncomment this causes panic
+	// defer C.free(unsafe.Pointer(ptr))
 
 	weight := NewTensor()
 
diff --git a/tensor/tensor.go b/tensor/tensor.go
index 271cbfb..e833a81 100644
--- a/tensor/tensor.go
+++ b/tensor/tensor.go
@@ -466,7 +466,7 @@ func RunBackward(tensors []Tensor, inputs []Tensor, keepGraphB bool, createGraph
 	// TODO: Are they allocated continouslly???
 	for i := 0; i < len(inputs); i++ {
 		outputPtr := (*lib.Ctensor)(unsafe.Pointer(C.malloc(0)))
-		// defer C.free(unsafe.Pointer(outputPtr))
+		defer C.free(unsafe.Pointer(outputPtr))
 		outputsPtr = append(outputsPtr, outputPtr)
 	}
 
diff --git a/tensor/util.go b/tensor/util.go
index 74a5be5..760c469 100644
--- a/tensor/util.go
+++ b/tensor/util.go
@@ -53,6 +53,8 @@ func init() {
 func CMalloc(nbytes int) (dataPtr unsafe.Pointer, buf *bytes.Buffer) {
 
 	dataPtr = C.malloc(C.size_t(nbytes))
+	// NOTE: uncomment this cause panic!
+	// defer C.free(unsafe.Pointer(dataPtr))
 
 	// Recall: 1 << 30 = 1 * 2 * 30
 	// Ref. See more at https://stackoverflow.com/questions/48756732