diff --git a/backends/vulkan/runtime/api/containers/StagingBuffer.cpp b/backends/vulkan/runtime/api/containers/StagingBuffer.cpp
index ca020ab6572..499f0b43d05 100644
--- a/backends/vulkan/runtime/api/containers/StagingBuffer.cpp
+++ b/backends/vulkan/runtime/api/containers/StagingBuffer.cpp
@@ -159,6 +159,11 @@ void StagingBuffer::cast_half_to_float_and_copy_from(
   for (size_t i = 0; i < numel; ++i) {
     dst[i] = half_to_float(src[i]);
   }
+  vmaFlushAllocation(
+      vulkan_buffer_.vma_allocator(),
+      vulkan_buffer_.allocation(),
+      0u,
+      VK_WHOLE_SIZE);
 }
 
 void StagingBuffer::cast_float_to_half_and_copy_to(
diff --git a/backends/vulkan/runtime/api/containers/StagingBuffer.h b/backends/vulkan/runtime/api/containers/StagingBuffer.h
index d786d030b39..19060804693 100644
--- a/backends/vulkan/runtime/api/containers/StagingBuffer.h
+++ b/backends/vulkan/runtime/api/containers/StagingBuffer.h
@@ -88,6 +88,11 @@ class StagingBuffer final {
     for (size_t i = 0; i < numel; ++i) {
       dst[i] = static_cast<DST_T>(src[i]);
     }
+    vmaFlushAllocation(
+        vulkan_buffer_.vma_allocator(),
+        vulkan_buffer_.allocation(),
+        0u,
+        VK_WHOLE_SIZE);
   }
 
   void cast_half_to_float_and_copy_from(
@@ -109,6 +114,11 @@ class StagingBuffer final {
   template <typename SRC_T, typename DST_T>
   void cast_and_copy_to(DST_T* dst, const size_t numel) {
     VK_CHECK_COND(numel <= this->numel());
+    vmaInvalidateAllocation(
+        vulkan_buffer_.vma_allocator(),
+        vulkan_buffer_.allocation(),
+        0u,
+        VK_WHOLE_SIZE);
     const SRC_T* src = reinterpret_cast<const SRC_T*>(data());
     for (size_t i = 0; i < numel; ++i) {
       dst[i] = static_cast<DST_T>(src[i]);