Show HN: Profine – optimize your PyTorch training script before the run

Category: devtools

Tags: pytorch, training-optimization, llm-agent

Score: 7.5/10 (Innovation: 7, Technical: 8, Documentation: 8, Utility: 7)

Profine is a CLI tool that profiles PyTorch training scripts on real GPUs, then uses an LLM agent to automatically suggest and apply optimizations like mixed precision and torch.compile, delivering measured speedups before a full training run. Its combination of deterministic profiling, LLM-driven bottleneck analysis, and safe multi-file editing makes it a novel approach to ML training optimization. The project is particularly interesting for its transparent, reproducible pipeline that bridges the gap between profiling and actual code changes.

Target audience: ML engineers, deep learning researchers

Repository: https://github.com/ProfineAI/profine-cli · Python · Apache-2.0 · 12 stars

View on Hacker News