[Bug]: The Python GEMM example crashes due to master branch changes.

[lcy-seso]

The Python GEMM example crash after the recent master branch updates.

[Parxd]

Hello! I was looking to contribute to any issues, but noticed this Python GEMM example was removed; are there any other good first issues that I could work on, or is this repo not accepting any open-source contributions at the moment?

[lcy-seso]

@Parxd Thank you for your interest in contributing to this repository. We greatly welcome and appreciate open-source contributions.

The Python example was intended to demonstrate the use of the header-only files of TileFusion to implement a tuner that adjusts the tile size across three levels of the memory hierarchy in GPU memory.

However, I temporarily removed the example after discovering some bugs that caused it to fail numeric checks with cuBLAS (I think it is caused by some simple bugs due to a lack of maintenance, as the core C++ implementations passed the correctness unittests). To prevent any potential confusion, I decided to remove it for now. But it is definitely in my plan to add it back.

I believe that adding a Python GEMM example would be a good first issue. I am more than happy to provide the necessary design ideas and documentation to support this effort.

I am also developing documentation to assist with contributions, which you can find here: TileFusion Documentation. This task is currently in progress.

Recently, I have been making the repository more robust and doing some cleanup for the next steps, which include Python binding, tuner, and support for Hopper. I will clarify the goals of the project and update the roadmaps accordingly in the coming days.

Thank you once again for your interest. I am more than happy to provide the necessary support.

[Parxd]

@lcy-seso I see, thanks! I would love to contribute in that case. I am happy to help with any other (more urgent) good first issues depending on your roadmap, as I've been able to build on my machine. Otherwise, I can help with this Python example as well. I see there's already bindings for the FlashAttention and ScatterND kernels; I assume the GEMM kernel was written but removed? I would really appreciate any more documentation or even a meeting to understand the project details further.

Thank you again!

[lcy-seso]

@Parxd Yes, I have initialized the bindings for Flash Attention and ScatterND, allowing us to keep the core components in C++ while moving tasks such as generating unit test data, running unit tests, and benchmarking to Python gradually.

And, yes, the GEMM kernel was written (you can find it here), but removed. The deleted Python example was intended to demonstrate a possible usage of the TileFusion macro kernel, specifically implementing a simple JIT that tunes the tile size across the GPU's three memory hierarchies. I do plan to add it back.

I believe I can prepare a quick write-up by updating the documentation to clarify the project's design goals, current status, and roadmaps. I will create a series of issues and provide a detailed description for each of them by the end of tomorrow (it is evening on my side now).

I will let you know as soon as this is completed. I think this could be a good first step to provide you with more details about the project. I am more than happy to help.

Thank you again. 😊