diff --git a/backends/cpu/cpu_ops.cpp b/backends/cpu/cpu_ops.cpp
index a9c8908..463acde 100644
--- a/backends/cpu/cpu_ops.cpp
+++ b/backends/cpu/cpu_ops.cpp
@@ -725,7 +725,6 @@ void CPUOps::build_dropout_mask(
     Tensor*/*shape*/, Tensor*/*strides*/
 ) {
     assert(mask != nullptr);
-    // assert(mask->get_dim() == 1);
     auto length = mask->length();
     for (int i = 0; i < length; ++i) {
         int index = 0;
diff --git a/backends/gpu/metal/metal_ops.cpp b/backends/gpu/metal/metal_ops.cpp
index cbb0e2d..10ea995 100644
--- a/backends/gpu/metal/metal_ops.cpp
+++ b/backends/gpu/metal/metal_ops.cpp
@@ -876,22 +876,6 @@ void MetalOps::fill(Tensor* tensor, float value) {
     for (int i = 0; i < tensor->length(); ++i) {
         data[i] = value;
     }
-
-    // fillOps->prepare(device, commandQueue, commandBuffer, encoder);
-    // int* argsInt = (int*)bufferIntArgs->contents();
-    // float* argsFloat = (float*)bufferFloatArgs->contents();
-    // auto length = tensor->length();
-    // argsInt[0] = length;
-    // argsFloat[0] = value;
-    // auto offset_tensor = calc_offset(tensor);
-
-    // assert(encoder != nullptr);
-    // encoder->setBuffer(reinterpret_cast<MTL::Buffer*>(tensor->get_storage()->ctx), offset_tensor, 0);
-    // encoder->setBuffer(bufferIntArgs, 0, 1);
-    // encoder->setBuffer(bufferFloatArgs, 0, 2);
-    // MTL::Size gridDim = MTL::Size((length + TILE_WIDTH - 1) / TILE_WIDTH, 1, 1);
-    // MTL::Size blockDim = MTL::Size(TILE_WIDTH, 1, 1);
-    // encoder->dispatchThreadgroups(gridDim, blockDim);
 }
 
 void MetalOps::reshape_deep_cp(
diff --git a/checkpoint.cpp b/checkpoint.cpp
index dee36a1..1f4b3a1 100644
--- a/checkpoint.cpp
+++ b/checkpoint.cpp
@@ -70,7 +70,6 @@ void diff_tensor_buffer(Tensor* tensor, char* buffer) {
                 << ", got: " << buffer_f[i] << std::endl;
             std::cerr << "tensor meta : " << tensor->get_meta_info() << std::endl;
             break;
-            // abort();
         }
     }
     ::free(tensor_buffer);
diff --git a/graph/actions.cpp b/graph/actions.cpp
index 62010fb..30b1496 100644
--- a/graph/actions.cpp
+++ b/graph/actions.cpp
@@ -638,12 +638,6 @@ void LazyDivAction::execute() {
     assert(lhs != nullptr);
     assert(res != nullptr);
     float fvalue = 0;
-    // g_backend_ops->cp_from_device(
-    //     reinterpret_cast<char*>(&fvalue),
-    //     value,
-    //     value->size()
-    // );
-    // fvalue += 1e-20;
     g_backend_ops->div(res, lhs, value);
 }
 
diff --git a/graph/node.cpp b/graph/node.cpp
index 3fd052e..05ab46a 100644
--- a/graph/node.cpp
+++ b/graph/node.cpp
@@ -543,11 +543,6 @@ namespace graph {
                 value
             )
         );
-        // g_backend_ops->cp_to_device(
-        //     value_tensor,
-        //     (char*)&value,
-        //     sizeof(float)
-        // );
         gCreateAction(
             new LazyDivAction(
                 l_tensor,
diff --git a/tensor/tensor.cpp b/tensor/tensor.cpp
index 0387929..2a880f4 100644
--- a/tensor/tensor.cpp
+++ b/tensor/tensor.cpp
@@ -559,13 +559,18 @@ void allocMemAndInitTensors() {
     void* tensors_data_ctx = nullptr;
     void* c_tensors_data_ctx = nullptr;
     void* grad_tensors_data_ctx = nullptr;
-    tensors_data = g_backend_ops->alloc(tensors_data_capacity, &tensors_data_ctx);
-    c_tensors_data = g_backend_ops->alloc(c_tensors_data_capacity, &c_tensors_data_ctx);
-    grad_tensors_data = g_backend_ops->alloc(grad_tensors_data_capacity, &grad_tensors_data_ctx);
-
-    g_backend_ops->memset(tensors_data, 0, tensors_data_capacity);
-    g_backend_ops->memset(c_tensors_data, 0, c_tensors_data_capacity);
-    g_backend_ops->memset(grad_tensors_data, 0, grad_tensors_data_capacity);
+    if (tensors_data_capacity > 0) {
+        tensors_data = g_backend_ops->alloc(tensors_data_capacity, &tensors_data_ctx);
+        g_backend_ops->memset(tensors_data, 0, tensors_data_capacity);
+    }
+    if (c_tensors_data_capacity > 0) {
+        c_tensors_data = g_backend_ops->alloc(c_tensors_data_capacity, &c_tensors_data_ctx);
+        g_backend_ops->memset(c_tensors_data, 0, c_tensors_data_capacity);
+    }
+    if (grad_tensors_data_capacity > 0) {
+        grad_tensors_data = g_backend_ops->alloc(grad_tensors_data_capacity, &grad_tensors_data_ctx);
+        g_backend_ops->memset(grad_tensors_data, 0, grad_tensors_data_capacity);
+    }
 
     int64_t offset = 0;
     for (Tensor* tensor : g_tensors) {