diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3cd58f5c59..c1aebe3b10 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -4,7 +4,7 @@
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
 cmake_minimum_required(VERSION 3.20.0 FATAL_ERROR)
-project(unified-runtime VERSION 0.10.3)
+project(unified-runtime VERSION 0.10.4)
 
 # Check if unified runtime is built as a standalone project.
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR OR UR_STANDALONE_BUILD)
diff --git a/source/adapters/opencl/usm.cpp b/source/adapters/opencl/usm.cpp
index 77a2941e0f..58ce4c0fc6 100644
--- a/source/adapters/opencl/usm.cpp
+++ b/source/adapters/opencl/usm.cpp
@@ -87,6 +87,11 @@ urUSMHostAlloc(ur_context_handle_t hContext, const ur_usm_desc_t *pUSMDesc,
   void *Ptr = nullptr;
   uint32_t Alignment = pUSMDesc ? pUSMDesc->align : 0;
 
+  if (pUSMDesc && pUSMDesc->align != 0 &&
+      ((pUSMDesc->align & (pUSMDesc->align - 1)) != 0)) {
+    return UR_RESULT_ERROR_INVALID_VALUE;
+  }
+
   std::vector<cl_mem_properties_intel> AllocProperties;
   if (pUSMDesc && pUSMDesc->pNext) {
     UR_RETURN_ON_FAILURE(usmDescToCLMemProperties(
@@ -130,6 +135,11 @@ urUSMDeviceAlloc(ur_context_handle_t hContext, ur_device_handle_t hDevice,
   void *Ptr = nullptr;
   uint32_t Alignment = pUSMDesc ? pUSMDesc->align : 0;
 
+  if (pUSMDesc && pUSMDesc->align != 0 &&
+      ((pUSMDesc->align & (pUSMDesc->align - 1)) != 0)) {
+    return UR_RESULT_ERROR_INVALID_VALUE;
+  }
+
   std::vector<cl_mem_properties_intel> AllocProperties;
   if (pUSMDesc && pUSMDesc->pNext) {
     UR_RETURN_ON_FAILURE(usmDescToCLMemProperties(
@@ -173,6 +183,11 @@ urUSMSharedAlloc(ur_context_handle_t hContext, ur_device_handle_t hDevice,
   void *Ptr = nullptr;
   uint32_t Alignment = pUSMDesc ? pUSMDesc->align : 0;
 
+  if (pUSMDesc && pUSMDesc->align != 0 &&
+      ((pUSMDesc->align & (pUSMDesc->align - 1)) != 0)) {
+    return UR_RESULT_ERROR_INVALID_VALUE;
+  }
+
   std::vector<cl_mem_properties_intel> AllocProperties;
   if (pUSMDesc && pUSMDesc->pNext) {
     UR_RETURN_ON_FAILURE(usmDescToCLMemProperties(
diff --git a/source/common/CMakeLists.txt b/source/common/CMakeLists.txt
index e327d7672b..025c3a3808 100644
--- a/source/common/CMakeLists.txt
+++ b/source/common/CMakeLists.txt
@@ -23,8 +23,8 @@ if (NOT DEFINED UMF_REPO)
 endif()
 
 if (NOT DEFINED UMF_TAG)
-    # v0.9.x 19.08.2024: Merge pull request #688 ...
-    set(UMF_TAG 59c4150b7120a7af5b3c8eb2d9b8bbb5d2e96aa3)
+    # v0.9.x 12.09.2024: 0.9.0 release
+    set(UMF_TAG v0.9.0)
 endif()
 
 message(STATUS "Will fetch Unified Memory Framework from ${UMF_REPO}")
diff --git a/source/loader/layers/tracing/ur_tracing_layer.cpp b/source/loader/layers/tracing/ur_tracing_layer.cpp
index 722ee77faa..c6fd4ca40d 100644
--- a/source/loader/layers/tracing/ur_tracing_layer.cpp
+++ b/source/loader/layers/tracing/ur_tracing_layer.cpp
@@ -34,14 +34,15 @@ struct XptiContextManager {
     ~XptiContextManager() { xptiFrameworkFinalize(); }
 };
 
-static std::shared_ptr<XptiContextManager> xptiContextManagerGlobal = [] {
-    return std::make_shared<XptiContextManager>();
-}();
+static std::shared_ptr<XptiContextManager> xptiContextManagerGet() {
+    static auto contextManager = std::make_shared<XptiContextManager>();
+    return contextManager;
+};
 static thread_local xpti_td *activeEvent;
 
 ///////////////////////////////////////////////////////////////////////////////
 context_t::context_t() : logger(logger::create_logger("tracing", true, true)) {
-    this->xptiContextManager = xptiContextManagerGlobal;
+    this->xptiContextManager = xptiContextManagerGet();
 
     call_stream_id = xptiRegisterStream(CALL_STREAM_NAME);
     std::ostringstream streamv;
diff --git a/source/loader/ur_loader.cpp b/source/loader/ur_loader.cpp
index bfc9da3e50..7609747622 100644
--- a/source/loader/ur_loader.cpp
+++ b/source/loader/ur_loader.cpp
@@ -15,6 +15,27 @@ context_t *getContext() { return context_t::get_direct(); }
 
 ///////////////////////////////////////////////////////////////////////////////
 ur_result_t context_t::init() {
+#ifdef _WIN32
+    // Suppress system errors.
+    // Tells the system to not display the critical-error-handler message box.
+    // Instead, the system sends the error to the calling process.
+    // This is crucial for graceful handling of adapters that couldn't be
+    // loaded, e.g. due to missing native run-times.
+    // TODO: add reporting in case of an error.
+    // NOTE: we restore the old mode to not affect user app behavior.
+    // See https://github.com/intel/llvm/blob/sycl/sycl/ur_win_proxy_loader/ur_win_proxy_loader.cpp (preloadLibraries())
+    UINT SavedMode = SetErrorMode(SEM_FAILCRITICALERRORS);
+#endif
+
+#ifdef UR_STATIC_ADAPTER_LEVEL_ZERO
+    // If the adapters were force loaded, it means the user wants to use
+    // a specific adapter library. Don't load any static adapters.
+    if (!adapter_registry.adaptersForceLoaded()) {
+        auto &level_zero = platforms.emplace_back(nullptr);
+        ur::level_zero::urAdapterGetDdiTables(&level_zero.dditable.ur);
+    }
+#endif
+
     for (const auto &adapterPaths : adapter_registry) {
         for (const auto &path : adapterPaths) {
             auto handle = LibLoader::loadAdapterLibrary(path.string().c_str());
@@ -24,6 +45,10 @@ ur_result_t context_t::init() {
             }
         }
     }
+#ifdef _WIN32
+    // Restore system error handling.
+    (void)SetErrorMode(SavedMode);
+#endif
 
     forceIntercept = getenv_tobool("UR_ENABLE_LOADER_INTERCEPT");