diff --git a/src/command.cpp b/src/command.cpp
index 84a7cdfff..2fbeab67b 100644
--- a/src/command.cpp
+++ b/src/command.cpp
@@ -2507,7 +2507,7 @@ VkTransfer::~VkTransfer()
     }
 }
 
-void VkTransfer::record_upload(const Mat& src, VkMat& dst, const Option& opt)
+void VkTransfer::record_upload(const Mat& src, VkMat& dst, const Option& opt, bool flatten)
 {
 //     NCNN_LOGE("record_upload src = %d | %d %d %d @ %d", src.dims, src.w, src.h, src.c, src.elempack);
 
@@ -2525,7 +2525,7 @@ void VkTransfer::record_upload(const Mat& src, VkMat& dst, const Option& opt)
         }
     }
 
-    Mat src_flattened = src.reshape(src.w * src.h * src.c);
+    Mat src_flattened = flatten ? src.reshape(src.w * src.h * src.c) : src;
 
     // create dst
     dst.create_like(src_flattened, opt.blob_vkallocator);
diff --git a/src/command.h b/src/command.h
index c988cfa5a..ca1766194 100644
--- a/src/command.h
+++ b/src/command.h
@@ -181,7 +181,7 @@ public:
     ~VkTransfer();
 
 public:
-    void record_upload(const Mat& src, VkMat& dst, const Option& opt);
+    void record_upload(const Mat& src, VkMat& dst, const Option& opt, bool flatten = true);
 
     void record_upload(const Mat& src, VkImageMat& dst, const Option& opt);
 
diff --git a/src/layer/vulkan/memorydata_vulkan.cpp b/src/layer/vulkan/memorydata_vulkan.cpp
index 1d1bd82d9..9db5cedc5 100644
--- a/src/layer/vulkan/memorydata_vulkan.cpp
+++ b/src/layer/vulkan/memorydata_vulkan.cpp
@@ -63,23 +63,17 @@ int MemoryData_vulkan::create_pipeline(const Option& opt)
     return 0;
 }
 
-int MemoryData_vulkan::upload_model(VkTransfer& /*cmd*/, const Option& opt)
+int MemoryData_vulkan::upload_model(VkTransfer& cmd, const Option& opt)
 {
-    // VkTransfer will flatten weight data
-    // so we use VkCompute for uploading
-    VkCompute cmd2(vkdev);
-
     if (support_image_storage && opt.use_image_storage)
     {
-        cmd2.record_upload(data, data_gpu_image, opt);
+        cmd.record_upload(data, data_gpu_image, opt);
     }
     else
     {
-        cmd2.record_upload(data, data_gpu, opt);
+        cmd.record_upload(data, data_gpu, opt, /*bool flatten*/ false);
     }
 
-    cmd2.submit_and_wait();
-
     return 0;
 }