NVIDIA Nsight Python 1.1: Jupyter Profiling Needs Specific Import Order and Nsight Compute 2026.2.1
The upcoming release of NVIDIA’s open-source profiling library introduces experimental backend support and explicit tool management for developers working in Jupyter environments.
Key points
- Version 1.1 adds an experimental CUPTI backend and explicit tool management features.
- The library requires Nsight Compute version 2026.2.1 or later to function correctly.
- Import order is critical; nsight must load before torch to capture CUDA work accurately.
NVIDIA is preparing to release version 1.1 of Nsight Python, an open-source interface that allows developers to profile GPU kernels directly from Python code, according to MarkTechPost. This update introduces an experimental CUPTI backend and explicit tool management capabilities, enabling more granular control over hardware metric analysis within interactive environments such as Jupyter and Google Colab. The tool attaches Nsight Compute to the running Python process, allowing users to measure kernel fusion, cache controls and clock settings without leaving their development environment.
Root cause
The previous version, 1.0, already allowed the library to attach Nsight Compute to a running Python process, making it usable in notebook environments. The upcoming 1.1 release builds on this foundation by adding a new backend option. According to MarkTechPost, this experimental CUPTI backend provides an alternative method for collecting hardware metrics. The update also includes explicit tool management and custom information collectors, which help developers structure how data is gathered from the GPU during execution.
Attack path
This release does not involve a security vulnerability or attack vector. Instead, it focuses on improving the developer experience for performance profiling. Users mark specific code regions with nsight.annotate and decorate benchmark functions with nsight.analyze.kernel. This process returns a pandas DataFrame containing hardware metrics per annotated region, configuration and run. The tool helps engineers distinguish between compute-bound and memory-bound work by analysing hardware counters and measuring bytes moved during kernel fusion.
Affected versions
- Nsight Python 1.1 (upcoming pre-release)
- Nsight Compute 2026.2.1 or later (required dependency)
- Nsight Compute 2026.3.1 (tested in the tutorial)
- Colab’s default Nsight Compute (older version, requires manual update)
Mitigation
- Ensure Nsight Compute version 2026.2.1 or later is installed and accessible via PATH.
- Install the cupti-python package in a separate directory to avoid library conflicts with PyTorch.
- Import the nsight library before importing torch to ensure proper attachment to CUDA work.
- Pin Nsight Python to the specific 1.1.0 pre-release commit for consistent behaviour.
Why it matters
Security and operations teams often overlook developer tools when assessing risk, but profiling software interacts deeply with hardware drivers and runtime libraries. Understanding how these tools attach to processes is crucial for maintaining system integrity. Teams usually verify that profiling utilities are sourced from official repositories and check for dependency conflicts that could introduce unstable libraries.
It is important to monitor how these tools access hardware counters, as misconfigured profilers can expose sensitive runtime metrics or cause performance degradation. Regular audits of development environments help ensure that profiling tools do not inadvertently bypass security controls or expose internal system states to unauthorised users.
What to do and how to stay safe: NVIDIA
- Verify that all profiling dependencies, such as Nsight Compute, are installed from official vendor repositories.
- Check import orders in Python scripts to ensure profiling libraries load before other GPU-intensive libraries.
- Monitor for library version conflicts, especially when using pre-releases in shared environments like Colab.
- Review documentation for any new experimental features, such as the CUPTI backend, before deploying in production workflows.
Step-by-step guide: Password Policies Examples That Reduce Password Fatigue
General security guidance from the Vector Bulletin newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the minimum version of Nsight Compute required for Nsight Python?
Nsight Python requires Nsight Compute version 2026.2.1 or later to function correctly.
Can Nsight Python be used in Google Colab?
Yes, but users may need to install a newer version of Nsight Compute because Colab’s default image ships an older version.
Why must nsight be imported before torch?
The nsight library loads Nsight Compute’s injection library into the Python process, which can only attach to CUDA work started after that import.
