• Add some minor fixes for OpenCL in tests and functionality
• Clarify some error messages
• Update the list of cuda versions for windows to include 9.1

• Fix for initialization on OpenCL
• Add cuda 9.0 to the list of versions.

• Add support for tensor cores on Volta
• Check the compute capability against the loaded version of cuda.
• Fix an error in the generated code for elemwise scalars

• Fix some race conditions in the CUDA blas code.
• Add more extension checking in the OpenCL core code.
• Add better python setup.py clean

• New automated conda package building for releases (uploaded to channel 'mila-udem')
• Fix a bunch of memory leaks
• Fix some possible synchronisation problems in BLAS dot.

• Fix for collectives tests
• Fix for header inclusion in ext_cuda.h

• Support for cuda9 float16
• Support for NCCL 2.0 (support for 1.0 dropped)
• Extensible API to configure contexts
• OpenCL minimum version requirement bumped to 1.2

This release is API and ABI incompatible with the 0.6.X series and will require some code adjustments from the clients.

• Improve cuda detection, especially on windows.
• Fix bug in kernel cache related to multiple kernels in the same source block.
• Fix a couple of possible segfaults on errors.
• Fix a number of new compiler warnings, mostly in the tests.

• Fix build on windows
• Add a test with appveyor so that the windows build will keep working.

• Fix broken pygpu.test() in 0.6.6.