[QST] Cutlass Python not showing custom mofifications

[IzanCatalan]

What is your question?
Hi, I would like to know how to import Cutlass from source for Python. I have read the guidelines in readme and installed with pip install -e . from /cutlass main folder. But if I modify the source code from any of the other folders, for instance, /inlcude/cutlass/conv/kiernel and compile it ( via make make -j$(nproc) in /cutlass/build folder) and after that, I uninstalled cutlass (pip uninstall), and I installed it again, the modifications didn't appear. If I use the examples (like example 16 ) with C++ and execute them from inside/cutlass/build, it works.

How can I also make the modifications in the source code appear in Python?

[hwu36]

@jackkosaian

[jackkosaian]

What do you mean when you say that "the modifications didn't appear"? How are you checking for the presence of your modifications?

[IzanCatalan]

@jackkosaian I use printf functions throughout all the code in both files https://github.com/NVIDIA/cutlass/blob/24f991e87930e1159f1f5a47e329d43bcfbd76b9/include/cutlass/conv/kernel/implicit_gemm_convolution.h (mainly in operator() function) and in https://github.com/NVIDIA/cutlass/blob/24f991e87930e1159f1f5a47e329d43bcfbd76b9/include/cutlass/conv/device/implicit_gemm_convolution.h (mainly in run() function).

When I modify, add, or eliminate these prints, as I commented, executing example 16 the output is correct. Still, when I compile and install Cutlass for Python, the output is the one when I first added or modified the prints; any further later modification no longer appears.

[jackkosaian]

The CUTLASS Python interface caches compiled code in a local file called compiled_cache.db. On subsequent runs, you're likely getting a cache hit, which results in not recompiling the kernel.

Can you please try deleting compiled_cache.db before trying the Python script again?

[IzanCatalan]

Yes, @jackkosaian, it was exactly what you said. I removed compiled_cache.db, uninstalled the package and installed it again, and it worked. I was not aware of this. Thank you for your help.