Hello, I am Jasjeet Singh, currently pursuing Btech in computer science from Punjab University. I’ve spent the past 6-8 months learning CUDA for machine learning AI purposes and have developed working knowledge of GEMM and BLAS concepts, which is what drew me to PyTensor’s rewriting infrastructure specifically. Ive already made some contribution with Argmax/Argmin and Max/Min monotonic rewrite in PR #1869 and also a Dot and Solve rewrite for Diagonal matrices in PR #1932. I am interested in Linear Algebra rewrites projects for GSOC 2026 and am already investigating eigs diagonal rewrite from COLA. I also found PyTensor has the Eig Op in nlinalg.py but no rewrite rules for it anywhere. In my PR #1932 I used AllocDiag parent node detection and x * eye pattern matching for diagonal rewrites — would that same approach work for eig, or is there a reason eig rewrites haven’t been done yet? I also noticed PR #1126 was opened 2 years ago was working on a similar thing but was abandoned? Is that considered fair game for GSoC scope?
While working on my GSOC proposal for linear algebra rewrites project, I made a prototype for eig rewrite for diagonal and identity matrices as listed in issue #573 since i was quite familiar with diagonal and identity matrices rewrites from my previous work and a significant amount of work had already been completed but since my prototype was so near being PR ready, I went ahead and created a PR as well.
Any feedback on the approach or direction would be appreciated as I work on my proposal
Hi @jessegrabowski and @zaxtax — I’ve been working on my proposal draft and wanted to share the latest version before the March 31 deadline. Any feedback would be greatly appreciated: draft. Thanks for your time.