Fixes for pip wheels (#1042)

* Fixes for wheels
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix paddle wheel test
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

build_tools/wheel_utils/build_wheels.sh

@@ -11,9 +11,8 @@ BUILD_PYTORCH=${4:-true}
 BUILD_PADDLE=${5:-true}
 
 export NVTE_RELEASE_BUILD=1
-export TARGET_BRANCH=${TARGET_BRANCH:-wheels}
-mkdir /wheelhouse
-mkdir /wheelhouse/logs
+export TARGET_BRANCH=${TARGET_BRANCH:-}
+mkdir -p /wheelhouse/logs
 
 # Generate wheels for common library.
 git config --global --add safe.directory /TransformerEngine

qa/L0_paddle_wheel/test.sh

@@ -15,7 +15,6 @@ cd transformer_engine/paddle
 python setup.py bdist_wheel
 
 export NVTE_RELEASE_BUILD=0
-cd $TE_PATH
 pip install dist/*
 
 python $TE_PATH/tests/paddle/test_sanity_import.py

transformer_engine/common/CMakeLists.txt

@@ -37,6 +37,7 @@ include_directories(${PROJECT_SOURCE_DIR}/..)
 # Configure Transformer Engine library
 set(transformer_engine_SOURCES)
 list(APPEND transformer_engine_SOURCES
+     pycudnn.cpp
      transformer_engine.cpp
      transpose/cast_transpose.cu
      transpose/transpose.cu
@@ -72,6 +73,8 @@ add_library(transformer_engine SHARED ${transformer_engine_SOURCES})
 target_include_directories(transformer_engine PUBLIC
                            "${CMAKE_CURRENT_SOURCE_DIR}/include")
 
+target_compile_definitions(transformer_engine PUBLIC NV_CUDNN_FRONTEND_USE_DYNAMIC_LOADING)
+
 # Configure dependencies
 target_link_libraries(transformer_engine PUBLIC
                       CUDA::cublas

transformer_engine/common/pycudnn.cpp

+/*************************************************************************
+ * Copyright (c) 2022-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ *
+ * See LICENSE for license information.
+ ************************************************************************/
+
+namespace cudnn_frontend {
+
+// This is needed to define the symbol `cudnn_dlhandle`
+// When using the flag NV_CUDNN_FRONTEND_USE_DYNAMIC_LOADING
+// to enable dynamic loading.
+void *cudnn_dlhandle = nullptr;
+
+}  // namespace cudnn_frontend