Reducing VMA large pool allocation to 4 MB from 32 MB to reduce memory wastage.

Differential Revision: D65864032

Pull Request resolved: pytorch#6816

Author: trivedivivek

backends/vulkan/runtime/vk_api/memory/vma_api.h

@@ -21,7 +21,7 @@
 #define VMA_DYNAMIC_VULKAN_FUNCTIONS 0
 #endif /* USE_VULKAN_WRAPPER */
 
-#define VMA_DEFAULT_LARGE_HEAP_BLOCK_SIZE (32ull * 1024 * 1024)
+#define VMA_DEFAULT_LARGE_HEAP_BLOCK_SIZE (4ull * 1024 * 1024)
 #define VMA_SMALL_HEAP_MAX_SIZE (256ull * 1024 * 1024)
 
 #define VMA_STATS_STRING_ENABLED 0