Deep-Learning-Profiling-Tools is a collection of open-source tools for profiling, analyzing, debugging, and optimizing GPU and accelerator programs in AI and high-performance computing workloads. Our projects include:
- Proton: A lightweight, multi-level profiler for Triton that connects Python context and user annotations with GPU kernel metrics to help identify performance bottlenecks.
- TileLens (formerly Triton-Viz): A kernel visualization and analysis toolkit that helps developers understand tensor operations, inspect memory accesses, and detect memory-safety issues. Many debugging workflows run without requiring a GPU.
- TileBench: An accelerator benchmarking framework for comparing kernel implementations across programming models using standardized correctness checks, timing protocols, and hardware-aware performance metrics.
- HPC-Performance-AI: A reproducible benchmark suite supporting AI-driven performance prediction and optimization for HPC applications, spanning standalone GPU kernels, mini-applications, and production applications with multi-GPU execution.
- PerfSkills: A collection of AI-agent skills for profiling-guided performance optimization, connecting bottleneck diagnosis with code changes, correctness validation, and measured performance improvements.
- Triton Samples: Examples covering Triton operators, compilation, and Proton profiling to support learning and experimentation with GPU kernel development.
- Fasten: A library of optimized GPU segment operators that accelerates heterogeneous graph neural networks through efficient segmented matrix multiplication and integration with PyG.
