التحسين الآلي للشيفرة

Compilers that optimise deep-learning code automatically need the model in a form they can work on, such as MLIR. During my remote research internship at NYU Abu Dhabi, my task was to find a reliable way to translate PyTorch models into MLIR. The long-term aim was large language models; in the time I had, I got it working for NLP models and other simpler models.

What I built

I worked on it alone. The pipeline imports a PyTorch model through torch-mlir's FX importer, the only torch-mlir importer still enabled at the time, and emits an MLIR module in one of three forms: the Torch dialect, Linalg-on-tensors or TOSA. The conversions themselves are torch-mlir's; my part was finding the path that worked and turning it into a pipeline with examples for BERT, RoBERTa, ResNet-18 and a linear regression. Documentation for that path was scarce, and people in the LLVM community on Discord pointed me in the right direction.

What I took from it

Sometimes the hardest part of the work is not the code but information that is scattered and hard to understand, and knowing whom to ask.