Hello,
I’m a PhD candidate employed at one of the largest HPC centers in Europe and I have been getting deeper into ML compilers. We don’t have a lot of expertise in this particular niche at our institute so we are looking for outside collaborators.
My work so far had me doing a deep dive into TorchInductor, so now I have a really good understanding of how the computational graph is lowered, fusions are scheduled and so forth. Besides, that I have a decent understanding of GPU architectures and programming paradigms. And also a good understanding of automatic differentiation. I also looked briefly at OpenXLA but I have not had time to dig deep in that direction.
I’m interested in exploring many topics within the domain of ML compilers, for example, more intelligent fusions, briding the gap between template compilers like TorchInductor that use templated kernels (for matmul, sdpa, convs) and actual compilers like Triton/Gluon/cuTile/…etc and the dichotomy between polyhedral optimization based compilers and scheduler-type compilers.
Whether you are in academia or industry, if you have expertise and interest in ML compilers and you’d like to act as collaborator/2nd supervisor, let me know.
Contact me at fpmatwork@gmail.com
I’ll then reply with my academic email and we can arrange an online meeting.