can't install te v1.13 in databricks

[Amalktb]

While trying to install te in databricks (16.2 ml run time and l4 as the node type), the machine fails to build the wheels for te. initially it shown some error which indicates cuda.h file were missing. After figuring out this, it shows cudnn.h file was not found. I figured some of these error by manually copying and pasting the .h files. But still could not complete the installation. Can any one help me, in completing the installation of te in databricks.

[ptrendx]

Do you need specifically TE 1.13 or may it be newer version? @ksivaman worked on overhauling the installation process of TE and starting from TE 2.4 you should be able to build TE with CUDA/cuDNN from pip wheels.

[ksivaman]

It is true that starting v2.4 the framework extensions' build/installation will not require the cuda toolkit, and the python packages provided by cuDNN can be used instead. However it is noteworthy that if you're installing from source (as opposed to from PyPI), the cuda toolkit and cuDNN system installation will still be required for compilation of the core lib. Support for full installation without cuda toolkit or system libs will be added in later versions.

[Amalktb]

I was working with Evo2, which requires TE version 1.13. Thats why TE 1.13 was necessary.