pytorch/kineto

A CPU+GPU Profiling library that provides access to timeline traces and hardware performance counters.

What it solves

Kineto provides performance observability and diagnostics for machine learning workloads, helping developers identify and resolve bottlenecks in common ML components.

How it works

It functions as a library integrated into the PyTorch Profiler. Its core component, Libkineto, is an in-process profiling library specifically designed for low-overhead GPU timeline tracing.

Who it’s for

Developers and researchers using PyTorch who need to analyze the execution timeline of their models to optimize performance.

Highlights

  • Low-overhead GPU timeline tracing via Libkineto.
  • Integration with external system-level profiling tools.
  • Actionable recommendations for common performance issues.
  • Compatibility with popular visualization platforms and analysis pipelines.

Related

  • Project
    meta-pytorch/tritonbench

    A collection of PyTorch operators used to evaluate the performance of Triton and its integration with PyTorch.

  • Project
    Denys88/rl_games

    A high-performance reinforcement learning library implemented in PyTorch that optimizes training throughput for robotics, multi-agent systems, and complex simulations.

  • Project
    pytorch/ao

    A PyTorch-native library for model optimization that provides quantization and sparsity tools to accelerate training and inference while reducing memory usage.

  • Project
    meta-pytorch/tnt

    A library of training tools and utilities for PyTorch users to streamline the model training process.

  • Project
    facebookresearch/HolisticTraceAnalysis

    Holistic Trace Analysis (HTA) is a Python library that reads PyTorch Profiler (Kineto) traces from distributed training jobs and produces data‑frames and visualisations showing where GPUs spend time—compute, communication, memory, or idle. It offers kernel‑level breakdowns, idle‑time causes, comm‑compute overlap, frequent kernel patterns, launch‑time stats, augmented memory‑bandwidth/queue‑length counters, and a trace‑comparison tool. Install via `pip install HolisticTraceAnalysis` (Linux/macOS, Python ≥ 3.10) and use the `TraceAnalysis` class in a notebook to get pandas DataFrames for each analysis. Documentation, examples, and an experimental CUPTI counter API are provided.