diff --git a/CMakeLists.txt b/CMakeLists.txt
index 40142f7..b048aaf 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -4,7 +4,7 @@
 
 set(OWL_VERSION_MAJOR 1)
 set(OWL_VERSION_MINOR 2)
-set(OWL_VERSION_PATCH 0)
+set(OWL_VERSION_PATCH 2)
 
 cmake_minimum_required(VERSION 3.24)
 
@@ -79,6 +79,10 @@ include(embed_optixir)
 option(OWL_BUILD_SHARED "Build OWL as a shared library? (otherwise static)" OFF)
 set(BUILD_SHARED_LIBS ${OWL_BUILD_SHARED}) # use 'OWL_' naming convention
 
+
+include(owl/cmake/configure_tbb.cmake)
+
+
 # ------------------------------------------------------------------
 # owl library itself, including owl-common
 # ------------------------------------------------------------------
diff --git a/include/owl/owl_host.h b/include/owl/owl_host.h
index 584d0fe..36568c0 100644
--- a/include/owl/owl_host.h
+++ b/include/owl/owl_host.h
@@ -5,7 +5,8 @@
 
 #pragma once
 
-#include <cuda.h>
+// #include <cuda.h>
+#include <cuda_runtime.h>
 #include <driver_types.h>
 #include <optix.h>
 
@@ -445,7 +446,7 @@ owlSetMaxInstancingDepth(OWLContext context,
                          int32_t maxInstanceDepth);
 
 /* return the cuda stream associated with the given device. */
-OWL_API CUstream
+OWL_API cudaStream_t
 owlContextGetStream(OWLContext context, int deviceID);
 
 /* return the optix context associated with the given device. */
@@ -723,7 +724,8 @@ owlTexture2DDestroy(OWLTexture texture);
 
 /*! returns the device handle of the given texture for the given
     device ID. Useful for custom texture object arrays. */
-OWL_API CUtexObject
+OWL_API cudaTextureObject_t
+// OWL_API CUtexObject
 owlTextureGetObject(OWLTexture texture, int deviceID);
 
 /*! returns the dimensions (number of pixels in) a given texture. */
@@ -893,7 +895,7 @@ owlAsyncLaunch2DOnDevice(OWLRayGen rayGen, int dims_x, int dims_y,
                         int deviceID, OWLParams params);
 
 
-OWL_API CUstream
+OWL_API cudaTextureObject_t//CUstream
 owlParamsGetCudaStream(OWLParams params, int deviceID);
 
 /*! wait for the async launch to finish */
diff --git a/owl/APIHandle.h b/owl/APIHandle.h
index f3dcbac..b888e2e 100644
--- a/owl/APIHandle.h
+++ b/owl/APIHandle.h
@@ -48,8 +48,9 @@ namespace owl {
     assert(object);
     std::shared_ptr<T> asT = std::dynamic_pointer_cast<T>(object);
     if (object && !asT) {
-      const std::string objectTypeID = typeid(*object.get()).name();
-	
+      auto *obj = object.get();
+      const std::string objectTypeID = typeid(*obj).name();
+       
       const std::string tTypeID = typeid(T).name();
       OWL_RAISE("could not convert APIHandle of type "
                 + objectTypeID
diff --git a/owl/CMakeLists.txt b/owl/CMakeLists.txt
index b954a68..a39e391 100644
--- a/owl/CMakeLists.txt
+++ b/owl/CMakeLists.txt
@@ -3,8 +3,6 @@
 
 find_package(OptiX REQUIRED)
 
-include(cmake/configure_tbb.cmake)
-
 set(OWL_SOURCES
   # -------------------------------------------------------
   # infrastructure
@@ -191,11 +189,17 @@ endif()
 
 target_include_directories(owl-config
   INTERFACE
-    ${PROJECT_SOURCE_DIR}
-    $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
-    #CMAKE_CURRENT_LIST_DIR}/include
+  $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}>
+  $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
+  $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
+  #CMAKE_CURRENT_LIST_DIR}/include
+)
+target_include_directories(owl  PUBLIC
+  $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
+)
+target_include_directories(owl_static PUBLIC
+  $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
 )
-
 if (WIN32)
   target_compile_definitions(owl-config INTERFACE -DNOMINMAX)
 endif()
diff --git a/owl/Group.h b/owl/Group.h
index 1f40790..2f2a80a 100644
--- a/owl/Group.h
+++ b/owl/Group.h
@@ -107,7 +107,7 @@ namespace owl {
     int getSBTOffset() const override { return sbtOffset; }
     
     /*! pretty-printer, for printf-debugging */
-    std::string toString() const;
+    std::string toString() const override;
     
     /*! list of child geometries to use in this BVH */
     std::vector<Geom::SP> geometries;
diff --git a/owl/LaunchParams.cpp b/owl/LaunchParams.cpp
index bc348ed..a8db7ba 100644
--- a/owl/LaunchParams.cpp
+++ b/owl/LaunchParams.cpp
@@ -24,6 +24,8 @@ namespace owl {
   {
     return std::make_shared<Object::DeviceData>(device);
   }
+
+  std::string LaunchParamsType::toString() const { return "LaunchParamsType"; }
   
   
   // ------------------------------------------------------------------
diff --git a/owl/LaunchParams.h b/owl/LaunchParams.h
index d1c7b09..ce795e0 100644
--- a/owl/LaunchParams.h
+++ b/owl/LaunchParams.h
@@ -27,7 +27,7 @@ namespace owl {
 
     Object::DeviceData::SP createOn(const DeviceContext::SP &device) override;
     
-    virtual std::string toString() const { return "LaunchParamsType"; }
+    virtual std::string toString() const override;
   };
 
   /*! an object that stores the variables used for building the launch
diff --git a/owl/UserGeomGroup.cpp b/owl/UserGeomGroup.cpp
index 21fcd3f..df65aab 100644
--- a/owl/UserGeomGroup.cpp
+++ b/owl/UserGeomGroup.cpp
@@ -27,6 +27,9 @@ namespace owl {
     buildFlags( (_buildFlags > 0) ? _buildFlags : defaultBuildFlags)
   {}
 
+  std::string UserGeomGroup::toString() const
+  { return "UserGeomGroup"; }
+  
   void UserGeomGroup::buildOrRefit(bool FULL_REBUILD)
   {
     for (auto child : geometries) {
diff --git a/owl/UserGeomGroup.h b/owl/UserGeomGroup.h
index 4adceed..e4cf964 100644
--- a/owl/UserGeomGroup.h
+++ b/owl/UserGeomGroup.h
@@ -17,7 +17,7 @@ namespace owl {
     UserGeomGroup(Context *const context,
                    size_t numChildren,
                    unsigned int buildFlags);
-    virtual std::string toString() const { return "UserGeomGroup"; }
+    virtual std::string toString() const override;
 
     /*! build() and refit() share most of their code; this functoin
         does all that code, with only minor specialization based on
diff --git a/owl/cmake/embed_ptx.cmake b/owl/cmake/embed_ptx.cmake
index fb29a8a..4049f9d 100644
--- a/owl/cmake/embed_ptx.cmake
+++ b/owl/cmake/embed_ptx.cmake
@@ -74,6 +74,7 @@ function(embed_ptx)
   
   target_sources(${PTX_TARGET} PUBLIC ${EMBED_PTX_SOURCES})
 
+  set(CUDA_TOOLKIT_VERSION "13.0.0")
   target_link_libraries(${PTX_TARGET} PRIVATE ${EMBED_PTX_PTX_LINK_LIBRARIES})
   set_property(TARGET ${PTX_TARGET} PROPERTY CUDA_PTX_COMPILATION ON)
   set_property(TARGET ${PTX_TARGET} PROPERTY CUDA_ARCHITECTURES OFF)
diff --git a/owl/cmake/owlConfig.cmake.in b/owl/cmake/owlConfig.cmake.in
index f030e85..e6d5ec3 100644
--- a/owl/cmake/owlConfig.cmake.in
+++ b/owl/cmake/owlConfig.cmake.in
@@ -23,9 +23,9 @@ set(OWL_DATAROOTDIR
   ${CMAKE_CURRENT_LIST_DIR}/../../../@CMAKE_INSTALL_DATAROOTDIR@/owl
   )
 
-if (@OWL_MPI@)
-  find_dependency(MPI)
-  set(OWL_FEATURE_MPI ON)
+if (@OWL_HAVE_TBB@ AND @TBB_FOUND@)
+  find_dependency(TBB)
+  set(OWL_HAVE_TBB ON)
 endif()
    
 
diff --git a/owl/impl.cpp b/owl/impl.cpp
index 21dbe4c..81f5fae 100644
--- a/owl/impl.cpp
+++ b/owl/impl.cpp
@@ -68,10 +68,12 @@ inline Buffer::SP checkGet(OWLBuffer _buffer)
 }
 
 /* return the cuda stream associated with the given device. */
-OWL_API CUstream owlContextGetStream(OWLContext _context, int deviceID)
+OWL_API cudaStream_t owlContextGetStream(OWLContext _context, int deviceID)
 {
   LOG_API_CALL();
-  return checkGet(_context)->getDevice(deviceID)->getStream();
+  CUstream s = checkGet(_context)->getDevice(deviceID)->getStream();
+  
+  return (cudaStream_t&)s;
 }
 
 /* return the optix context associated with the given device. */
@@ -748,14 +750,15 @@ owlTexture2DCreate(OWLContext _context,
   return (OWLTexture)context->createHandle(texture);
 }
 
-OWL_API CUtexObject
+// OWL_API CUtexObject
+OWL_API cudaTextureObject_t
 owlTextureGetObject(OWLTexture _texture, int deviceID)
 {
   LOG_API_CALL();
   assert(_texture);
   Texture::SP texture = ((APIHandle *)_texture)->get<Texture>();
   assert(texture);
-  return texture->getObject(deviceID);
+  return (cudaTextureObject_t)texture->getObject(deviceID);
 }
 
 OWL_API owl2ui
@@ -876,14 +879,15 @@ owlGroupGetTraversable(OWLGroup _group, int deviceID)
   return group->getTraversable(group->context->getDevice(deviceID));
 }
 
-OWL_API CUstream
+OWL_API cudaTextureObject_t
+// OWL_API CUstream
 owlParamsGetCudaStream(OWLLaunchParams _lp, int deviceID)
 {
   LOG_API_CALL();
   assert(_lp);
   LaunchParams::SP lp = ((APIHandle *)_lp)->get<LaunchParams>();
   assert(lp);
-  return lp->getCudaStream(lp->context->getDevice(deviceID));
+  return (cudaTextureObject_t)lp->getCudaStream(lp->context->getDevice(deviceID));
 }
 
 OWL_API void