The infrastructure between model and machine.
I created AOTInductor, PyTorch's ahead-of-time compiler for deploying models without Python, and I'm the tech lead behind it. For five years I've worked on the layers that turn PyTorch programs into fast, deployable code. That runs from getting TorchInductor ready for PyTorch 2.0 to bringing these technologies into production at Meta.
Three things I've built, and why they matter.
AOTInductor: deploy PyTorch without Python.
I created AOTInductor and lead its development. It compiles an exported PyTorch model into a self-contained artifact behind a stable C ABI. That artifact can be served from C++ and stays portable across GPU architectures and PyTorch releases. It has been widely adopted for production deployment, replacing earlier generations of native deployment solutions.
See the PRs →TorchInductor: from prototype to PyTorch 2.0.
I joined TorchDynamo and TorchInductor in 2022, while they were still a prototype in a separate repo. I worked across the compiler: scheduling, op lowerings, and a long tail of codegen correctness. I also built the quality bar it shipped against: model-suite CI, accuracy checks, and PyTorch's core test suites running under Inductor. Later I built the nightly PT2 performance dashboard that tracks speedup, compile time, and memory.
See the PRs →Taking the host out of the hot path.
I turned Inductor's C++ wrapper into a first-class mode for both JIT and AOT, removing Python from the kernel-launch path, and added lazy kernel compilation for faster cold starts. Now I'm driving CUDA Graphs and regional compilation, which give large models per-region control over how they're compiled and captured.
See the PRs →Leading the work, not just doing it.
As tech lead for AOTInductor and Inductor's C++ wrapper, I led the PyTorch engineers building them in open source. Today I lead a larger team driving production adoption of these technologies at Meta.
Along the way: 1,300+ code reviews, the AOTInductor tutorial and docs, and integrations with vLLM, ExecuTorch, and torchchat.
ML compilers
TorchInductorTorchDynamoTriton codegenAOT compilationRuntime & deployment
C++ runtimeStable C ABIModel packagingMulti-arch CUDAPerformance
CUDA GraphsAutotuningCompile-time optimizationCUTLASSInfrastructure
BenchmarkingPerformance CICorrectness testing at scaleFive years across PyTorch's compiler stack.
- 2025 → nowRuntime overhead, large-model compilation, and production enablement
- 2023 → 2025Created and led AOTInductor
- 2022 → 2023TorchInductor, from prototype to PyTorch 2.0
- 2021 → 2022TorchScript and Lazy Tensor Core