plasma-umass/scalene

Scalene: a high-performance, high-precision CPU, GPU, and memory profiler for Python with AI-powered optimization proposals

What it solves

Scalene is a high-performance profiler for Python that identifies bottlenecks in CPU, GPU, and memory usage. It solves the problem of imprecise or slow profiling by providing line-level detail and separating time spent in Python code from time spent in native libraries or system I/O, allowing developers to focus on the code they can actually optimize.

How it works

Scalene uses sampling instead of instrumentation to keep overhead low (typically 10-20%). It tracks CPU usage, NVIDIA GPU time, and memory allocations via a specialized memory allocator. It also attributes await time for async coroutines and captures stitched Python and native (C/C++) call stacks. To provide optimization suggestions, it integrates with AI providers like OpenAI, Azure, Amazon Bedrock, and Ollama.

Who it’s for

It is designed for Python developers who need to optimize the performance of their applications, particularly those using heavy native libraries, asynchronous code, or GPU acceleration.

Highlights

  • AI-Powered Optimizations: Generates proposed code improvements using LLMs based on the profiled data.
  • Detailed Resource Tracking: Separates Python, native, and system time, and identifies memory leaks and copying volume.
  • Comprehensive Visualizations: Includes a web-based GUI with flame charts for call stacks, memory-weighted stacks, and a wall-clock timeline.
  • Async Profiling: Measures per-line wall-clock attribution for suspended coroutines and tracks concurrency levels.
  • Low Overhead: Runs significantly faster than many other profilers while maintaining high accuracy.

Related

  • Project
    krrishnarraj/clpeak

    clpeak is an open‑source cross‑platform benchmark that measures peak compute throughput, memory bandwidth and latency on CPUs, GPUs and NPUs. It supports many backends (CUDA, ROCm, Metal, oneAPI, Vulkan, OpenCL, ONNX Runtime, Core ML, LiteRT), offers both CLI and GUI, and can export results as JSON for comparison.

  • Project
    stumpy-dev/stumpy

    A scalable Python library for computing matrix profiles to identify patterns, anomalies, and motifs within time series data.

  • Dispatch
    Profiling in PyTorch: A Beginner's Guide to torch.profiler

    Hugging Face provides a comprehensive guide to using torch.profiler to identify bottlenecks, understand the CPU-GPU dispatch chain, and analyze the impact of torch.compile on kernel execution.

  • Project
    sknetwork-team/scikit-network

    A Python library for machine learning on graphs that uses sparse matrices for memory-efficient network analysis.

  • Project
    earthgecko/skyline

    A real-time Python-based anomaly detection and performance monitoring system that automatically identifies anomalies in high-resolution time series data without requiring manual threshold configuration.