CUPTI Python
Type: Technology Tags: NVIDIA, CUPTI, Python, profiling, CUDA, observability Related: CUDA-Python, CUPTI, Nsight-Python, Nsight-Systems, Nsight-Compute, NVIDIA-CUDA, PyTorch, TensorRT Sources: https://docs.nvidia.com/cupti-python/, https://nvidia.github.io/cuda-python/latest/, https://docs.nvidia.com/nsight-python/index.html Last Updated: 2026-04-29
Summary
CUPTI Python provides Python-facing documentation and bindings around the CUDA Profiling Tools Interface. It is listed from the current CUDA-Python hub and brings CUPTI profiling and tracing capabilities closer to Python-based AI, HPC, and data-processing workflows.
Detail
Purpose
Much of the modern NVIDIA software ecosystem is driven from Python, including PyTorch, TensorRT tooling, RAPIDS, and custom orchestration. CUPTI Python helps expose CUDA profiling capabilities in that environment instead of requiring every workflow to be built directly in C/C++.
Key capabilities
- Python documentation for CUDA profiling interface usage.
- Access to CUPTI concepts used by profiling tools targeting CUDA applications.
- Useful for custom telemetry, tracing, and performance instrumentation around Python GPU workloads.
NVIDIA context
CUPTI Python complements user-facing profilers like Nsight-Systems and Nsight-Compute by enabling programmable profiling workflows. It is especially relevant when teams want to collect CUDA activity from Python applications and frameworks.
Connections
- CUPTI - underlying CUDA Profiling Tools Interface.
- CUDA-Python - umbrella for NVIDIA Python CUDA components.
- Nsight-Python - higher-level Python profiling automation in the CUDA Python ecosystem.
- Nsight-Systems - system profiler built around CUDA activity visibility.
- Nsight-Compute - kernel profiler that uses lower-level profiling concepts.
- PyTorch - Python GPU framework where custom profiling can be valuable.
- TensorRT - inference workflows often need profiling and trace integration.
Source Excerpts
- NVIDIA’s CUPTI Python docs introduce Python documentation for CUPTI, the profiling interface for CUDA applications.