diff --git a/src/command.cpp b/src/command.cpp
index 043d7d359..6939a3273 100644
--- a/src/command.cpp
+++ b/src/command.cpp
@@ -20,7 +20,7 @@
 
 namespace ncnn {
 
-Command::Command(VulkanDevice* _vkdev, uint32_t _queue_index) : vkdev(_vkdev), queue_index(_queue_index)
+Command::Command(const VulkanDevice* _vkdev, uint32_t _queue_index) : vkdev(_vkdev), queue_index(_queue_index)
 {
     // get queue
     vkGetDeviceQueue(vkdev->vkdevice(), queue_index, 0, &queue);
@@ -161,7 +161,7 @@ int Command::wait_fence()
     return 0;
 }
 
-VkCompute::VkCompute(VulkanDevice* _vkdev) : Command(_vkdev, _vkdev->info.compute_queue_index)
+VkCompute::VkCompute(const VulkanDevice* _vkdev) : Command(_vkdev, _vkdev->info.compute_queue_index)
 {
     if (vkdev->info.support_VK_KHR_push_descriptor)
     {
@@ -701,7 +701,7 @@ void VkCompute::transfer_transfer_barrier(VkBuffer buffer, size_t offset, size_t
     vkCmdPipelineBarrier(command_buffer, srcStageMask, dstStageMask, 0, 0, 0, 1, &bufferBarrier, 0, 0);
 }
 
-VkTransfer::VkTransfer(VulkanDevice* _vkdev) : Command(_vkdev, _vkdev->info.transfer_queue_index)
+VkTransfer::VkTransfer(const VulkanDevice* _vkdev) : Command(_vkdev, _vkdev->info.transfer_queue_index)
 {
     buffer_offset_alignment = vkdev->info.buffer_offset_alignment;
     staging_data = 0;
diff --git a/src/command.h b/src/command.h
index b39c6ee8a..c6a6cd79a 100644
--- a/src/command.h
+++ b/src/command.h
@@ -29,7 +29,7 @@ namespace ncnn {
 class Command
 {
 public:
-    Command(VulkanDevice* vkdev, uint32_t queue_index);
+    Command(const VulkanDevice* vkdev, uint32_t queue_index);
     ~Command();
 
 protected:
@@ -43,7 +43,7 @@ protected:
     int wait_fence();
 
 protected:
-    VulkanDevice* vkdev;
+    const VulkanDevice* vkdev;
     uint32_t queue_index;
 
     VkQueue queue;
@@ -57,7 +57,7 @@ protected:
 class VkCompute : public Command
 {
 public:
-    VkCompute(VulkanDevice* vkdev);
+    VkCompute(const VulkanDevice* vkdev);
     ~VkCompute();
 
     void record_upload(const VkMat& m);
@@ -153,7 +153,7 @@ protected:
 class VkTransfer : public Command
 {
 public:
-    VkTransfer(VulkanDevice* vkdev);
+    VkTransfer(const VulkanDevice* vkdev);
     ~VkTransfer();
 
     void record_upload(const Mat& src, VkMat& dst);
diff --git a/src/layer.cpp b/src/layer.cpp
index e2a02ecbd..0f363ebbb 100644
--- a/src/layer.cpp
+++ b/src/layer.cpp
@@ -30,7 +30,7 @@ Option::Option()
     workspace_allocator = 0;
 
 #if NCNN_VULKAN
-    vulkan_compute = true;
+    vulkan_compute = false;
     blob_vkallocator = 0;
     workspace_vkallocator = 0;
     staging_vkallocator = 0;
diff --git a/src/layer/concat.cpp b/src/layer/concat.cpp
index fb3a040b3..b6f74f796 100644
--- a/src/layer/concat.cpp
+++ b/src/layer/concat.cpp
@@ -13,6 +13,7 @@
 // specific language governing permissions and limitations under the License.
 
 #include "concat.h"
+#include <algorithm>
 
 namespace ncnn {
 
diff --git a/src/net.cpp b/src/net.cpp
index 319fbe2c3..c9e089031 100644
--- a/src/net.cpp
+++ b/src/net.cpp
@@ -44,6 +44,7 @@ Net::Net()
 
 #if NCNN_VULKAN
     vkdev = 0;
+    vkdev_local = 0;
     weight_vkallocator = 0;
     weight_staging_vkallocator = 0;
 #endif // NCNN_VULKAN
@@ -52,6 +53,10 @@ Net::Net()
 Net::~Net()
 {
     clear();
+
+#if NCNN_VULKAN
+    delete vkdev_local;
+#endif // NCNN_VULKAN
 }
 
 #if NCNN_STRING
@@ -142,8 +147,10 @@ int Net::load_param(FILE* fp)
 #if NCNN_VULKAN
     if (use_vulkan_compute && !vkdev)
     {
-        fprintf(stderr, "vulkan device not set, vulkan compute disabled\n");
-        use_vulkan_compute = false;
+        // use default vulkan device
+        if (!vkdev_local)
+            vkdev_local = new VulkanDevice;
+        vkdev = vkdev_local;
     }
 #endif // NCNN_VULKAN
 
@@ -311,8 +318,10 @@ int Net::load_param_mem(const char* _mem)
 #if NCNN_VULKAN
     if (use_vulkan_compute && !vkdev)
     {
-        fprintf(stderr, "vulkan device not set, vulkan compute disabled\n");
-        use_vulkan_compute = false;
+        // use default vulkan device
+        if (!vkdev_local)
+            vkdev_local = new VulkanDevice;
+        vkdev = vkdev_local;
     }
 #endif // NCNN_VULKAN
 
@@ -483,8 +492,10 @@ int Net::load_param_bin(FILE* fp)
 #if NCNN_VULKAN
     if (use_vulkan_compute && !vkdev)
     {
-        fprintf(stderr, "vulkan device not set, vulkan compute disabled\n");
-        use_vulkan_compute = false;
+        // use default vulkan device
+        if (!vkdev_local)
+            vkdev_local = new VulkanDevice;
+        vkdev = vkdev_local;
     }
 #endif // NCNN_VULKAN
 
@@ -720,8 +731,10 @@ int Net::load_param(const unsigned char* _mem)
 #if NCNN_VULKAN
     if (use_vulkan_compute && !vkdev)
     {
-        fprintf(stderr, "vulkan device not set, vulkan compute disabled\n");
-        use_vulkan_compute = false;
+        // use default vulkan device
+        if (!vkdev_local)
+            vkdev_local = new VulkanDevice;
+        vkdev = vkdev_local;
     }
 #endif // NCNN_VULKAN
 
@@ -918,7 +931,7 @@ Extractor Net::create_extractor() const
 }
 
 #if NCNN_VULKAN
-void Net::set_vulkan_device(VulkanDevice* _vkdev)
+void Net::set_vulkan_device(const VulkanDevice* _vkdev)
 {
     vkdev = _vkdev;
 }
@@ -1592,6 +1605,8 @@ Extractor::Extractor(const Net* _net, int blob_count) : net(_net)
     opt = get_default_option();
 
 #if NCNN_VULKAN
+    opt.vulkan_compute = net->use_vulkan_compute;
+
     if (net->use_vulkan_compute)
     {
         blob_mats_gpu.resize(blob_count);
diff --git a/src/net.h b/src/net.h
index c4b9d1966..6c877e4e9 100644
--- a/src/net.h
+++ b/src/net.h
@@ -106,7 +106,7 @@ public:
 
 #if NCNN_VULKAN
 
-    void set_vulkan_device(VulkanDevice* vkdev);
+    void set_vulkan_device(const VulkanDevice* vkdev);
 
 #endif // NCNN_VULKAN
 
@@ -132,7 +132,8 @@ protected:
     std::vector<layer_registry_entry> custom_layer_registry;
 
 #if NCNN_VULKAN
-    VulkanDevice* vkdev;
+    const VulkanDevice* vkdev;
+    const VulkanDevice* vkdev_local;
 
     VkAllocator* weight_vkallocator;
     VkAllocator* weight_staging_vkallocator;