[intel-npu] Bugfix for total allocable memory property

src/plugins/intel_npu/src/backend/src/zero_device.cpp

@@ -12,6 +12,8 @@
 #include "zero_infer_request.hpp"
 #include "zero_remote_tensor.hpp"
 
+#define ZD_LEGACY_MAX_MEM_ALLOC_SIZE_BYTES (2147483648)  // 2GB in base-2
+
 using namespace intel_npu;
 
 ZeroDevice::ZeroDevice(const std::shared_ptr<ZeroInitStructsHolder>& initStructs)
@@ -167,7 +169,27 @@ uint64_t ZeroDevice::getTotalMemSize() const {
         _graph_ddi_table_ext.pfnQueryContextMemory(_initStructs->getContext(), ZE_GRAPH_QUERY_MEMORY_DDR, &query);
     THROW_ON_FAIL_FOR_LEVELZERO_EXT("pfnQueryContextMemory", result, _graph_ddi_table_ext);
 
-    return query.total;
+    // For drivers with graph_extension < 1.9 we report fixed 2GB max allocation size (old drivers don't support more)
+    // For drivers with graph_extension > 1.9 we report the value they return
+    // Extra case for windows only: for graph_extension 1.8 need to convert driver value to bytes
+    if (_initStructs->isExtensionSupported(std::string("ZE_extension_graph_1_9"), ZE_MAKE_VERSION(1, 9))) {
+        // we are safe here, can return the value directly from driver
+        return query.total;
+    }
+#if defined(_WIN32) || defined(__CYGWIN__)
+    else if (_initStructs->isExtensionSupported(std::string("ZE_extension_graph_1_8"), ZE_MAKE_VERSION(1, 8))) {
+        // special case for windows-only. It supports correct value from 1.8 but in incorrect unit
+        // need to convert driver value from KB to B
+        return query.total * 1024;  // KB to B -base2
+    }
+#endif  // _WIN32 || __CYGWIN__
+    else {
+        // for all older drivers we return fixed 2GB (in bytes)
+        return ZD_LEGACY_MAX_MEM_ALLOC_SIZE_BYTES;
+    }
+
+    // Default: return 2GB
+    return ZD_LEGACY_MAX_MEM_ALLOC_SIZE_BYTES;
 }
 
 ov::device::PCIInfo ZeroDevice::getPciInfo() const {