Profine is a tool designed to profile and optimize PyTorch training code running on GPUs. By executing code on real hardware, it identifies genuine performance bottlenecks rather than relying on synthetic benchmarks. compile, scaled_dot_product_attention (SDPA), fused AdamW, bf16 autocast, and TF32 matmul precision, all while ensuring that the original training semantics remain unchanged.
A key feature of Profine is its transparent rewriting process, which produces a reviewable diff for every change made. Each modification comes with a stated justification, allowing users to confidently assess and validate the optimizations before deploying them. The tool reports the actual end-to-end speedup achieved in the training loop, providing measured results instead of theoretical projections. This approach enables users to understand the impact of each optimization prior to committing to lengthy training runs.
Profine operates through a deterministic pipeline that includes profiling, planning, rewriting, validating, and reporting in a single pass. The tool is delivered as a Python package available for installation via PyPI using pip. It is also referenced as an open-source project on GitHub, indicating accessibility for those interested in examining or contributing to the source code.
The service is aimed at individuals or teams working with PyTorch training workloads who seek to improve their GPU training efficiency through automated, reviewable optimizations and measurable performance gains.
Profine is a LLM eval & observability product. It helps machine learning engineers optimize PyTorch training code for better GPU performance and efficiency. It is built as an open-source project for machine learning engineers. Profine is open source under the Apache-2.0 license. The product ships for the command line.
Behind Profine is ProfineAI, and the product first shipped in 2026. Development happens publicly on GitHub with 17 stars and 46 commits in the last 90 days. PulseGate's similarity index finds few close equivalents — Profine occupies a relatively distinct niche. Key capabilities include pyTorch profiling, GPU optimization, and code rewriting.
Latest indexed changes and source events
Other apps tracked under the same category.
