Federico Bruzzone

CS PhD Candidate. Programming Languages and Compilers enthusiast. Diving into AI compilers. MLIR • LLVM • IREE.

[Federico Bruzzone @fedebruzzone7]

https://t.co/NpcVrEQGFz

[Federico Bruzzone @fedebruzzone7]

Anatomy of High-Performance Matrix Multiplication

🔗 https://t.co/FwUC3sN3TU

[Patrick C Toulme @PatrickToulme]

SemiAnalysis’ OpenAI Jalapeño blog shows an obvious industry trend. OpenAI wrote the entire inference stack in pure assembly. No framework required. All from hillclimbing.

Traditional ML frameworks like PyTorch are vanishing from the frontier stack. We already saw this with @elonmusk announcement that the entire SpaceXAI stack is in C.

I think you will see all frontier labs and inference providers converge to AI generated training and inference stacks in C / assembly.

It is a bit nostalgic, as the AI industry would never have started without ML frameworks but the future of serving hill climbed assembly is obviously coming.

[Federico Bruzzone @fedebruzzone7]

Multidirectional Propagation of Sparsity Information
across Tensor Slices

🔗https://t.co/ODQBdKF0Uy

[[email protected] @HaskRay]

When code is this cheap to produce, saying no is the scarce resource. I don't think maintainers can absorb the volume that's coming....

[Federico Bruzzone @fedebruzzone7]

Tensor Seeks Layout: Formalizing Layout Selection for
ML Compilers

🔗https://t.co/tkkDjdy0N1

[[email protected] @HaskRay]

https://t.co/JjytNvZy49
Alexis Engelke总结了LLVM 23在编译速度上的优化工作：整体编译时间降低了6.75%，sqlite3甚至提升了10.53%。能参与到这样一次系统性的性能提升里，挺有成就感的😺。这种成熟稳定软件的进步，往往就是靠一点一点抠出来的效率积累起来的🥹

[Federico Bruzzone @fedebruzzone7]

Generative AI at the Edge: A Comprehensive Survey of
Architectures, Hardware and Applications

🔗https://t.co/A7wQvgFLmX

[Federico Bruzzone @fedebruzzone7]

For anyone who missed the live stream, or just wants to watch the keynote by @clattner_llvm, @iamtimdavis, and other folks again 🚀

A huge thanks to @Modular and @Qualcomm for making this ecosystem available as open source after all their hard work 🔥

https://t.co/UDdLkeCoCD

[Federico Bruzzone @fedebruzzone7]

@secona0 That's definitely the only way!
But can you compare the beauty of "—" to "--"? :'D

[Federico Bruzzone @fedebruzzone7]

The only sad thing about the AI ​​era is that I can no longer use, except in papers, the em dash (i.e., "—"), because otherwise people think I'm an AI myself 🤣

[Federico Bruzzone @fedebruzzone7]

@clattner_llvm Absolutely incredible milestone, thank you! Really looking forward to diving into it, seeing how MLIR/LLVM is used behind the scenes, and hopefully contributing along the way 🚀

[Federico Bruzzone @fedebruzzone7]

The Modular ecosystem is now Open Source! 🚀

From Max to the Mojo compiler, built on MLIR/LLVM.

Breaking news, huge thanks to @clattner_llvm and everyone who has worked on this over the past 4.5 years!

[Federico Bruzzone @fedebruzzone7]

@adityazero_ @clattner_llvm I haven't dug into it thoroughly yet, but from what little I've seen, Max is implemented in Mojo, and dialects seem to be upstream. I'm really curious to see how it differs from other AI compilers, too!

If you discover anything interesting, don't hesitate to reach out to me 🙏

[Federico Bruzzone @fedebruzzone7]

🔗 https://t.co/m9bf48QBfB
🎥 https://t.co/kuyDpJcBdG

[Federico Bruzzone @fedebruzzone7]

@ReubenConducts Hi Ruben, thank you and Colfax for this fantastic article and your research! 🫶

[Federico Bruzzone @fedebruzzone7]

From a category theory perspective: Categorical Foundations for CuTe Layouts

🔗https://t.co/BuL7RnPHkH

[Federico Bruzzone @fedebruzzone7]

CuTe Layout Representation and Algebra

🔗https://t.co/4yF1T5LQ3N

[Modular @Modular]

Keynote starting in a few minutes! Let us know where you're joining from in the livestream chat:

https://t.co/3niTn9rBXU

#ModCon2026

[Federico Bruzzone @fedebruzzone7]

CuTe Layout Representation and Algebra

🔗https://t.co/4yF1T5LQ3N

[Federico Bruzzone @fedebruzzone7]

@AartBik This is true! Maranello has produced cars that are masterpieces.

I'm not saying this because I'm Italian ;)

[Federico Bruzzone @fedebruzzone7]

Mind the Abstraction Gap: Bringing Equality Saturation to Real-World ML Compilers

🔗https://t.co/9dJ3tQ9aTw