What torch.compile Sees, and What It's Blind To
The first time I put torch.compile in front of an LLM inference server, it spent the better part of a minute compiling and handed back a 6% speedup. I nearly wrote the feature off. The next morning...
Aditya Kumar
Triton: The Compiler That Pretends to Be a Library
Triton is a compiler with a Python frontend. The @triton.jit decorator does not decorate a function. It parses the function’s AST, runs it through an MLIR pipeline, and emits a GPU binary. The Pyth...
Anatomy of a CUDA Binary
When you compile a CUDA kernel, the final artifact is a cubin — a CUDA binary. It is a standard ELF64 file with NVIDIA-specific sections that encode everything the CUDA driver needs to load and lau...
Where Should Your Code Live?
A survey of code repository hosting in 2026 Most developers do not think about where their Git repository lives until something goes wrong. An account gets suspended. A DMCA notice takes down a pr...
Algorithms + Data Structures = Programs
A review and critical analysis of Niklaus Wirth’s foundational text Over the past decade, I have been thinking about how programming languages evolve and what survives the churn. Languages rise an...
The Anatomy of an Instruction Pipeline Hazard
A case study of the B200 pipeline model A note on methodology: Everything in this article is based on my analysis of microbenchmarks executed directly on B200 silicon. Nvidia does not publish i...
The Mathematics of Dataflow Analysis: Lattices, Fixpoints, and Semirings
When building a dataflow analysis framework in a compiler, you quickly encounter an array of mathematical terms: semilattices, semirings, Kleene algebras, Galois connections, and Knaster-Tarski fix...
Three Ways to Take a Gradient: Tape, Trace, and Source Transform
Training is gradient descent, and gradient descent needs gradients, so every ML framework has an automatic differentiation engine at its core. What is less obvious is that the representation a fram...
A tour of MLIR: The Dialect Stack Everyone Depends On
If you train or serve models, you depend on MLIR whether or not you have ever written a line of it. XLA lowers through it, Triton is built on it, Mojo is MLIR-native, and Torch-MLIR, IREE, and ONNX...
Crossing the Boundary: Custom Kernels and the C++/Python ABI in vLLM
Python is a productive orchestration language for inference serving, but it is the wrong tool for the critical path of token generation. Large-model inference is bound by memory bandwidth and by pe...