More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    GPU Stack Inspector: why CUDA works in one env, not another

    1

    Explains why torch.cuda.is_available() is False in one Python environment and True in another, with the evidence, before you reinstall anything.

    Free

    Security scanned
    GPU Stack Inspector: why CUDA works in one env, not another

    GPU Stack Inspector: why CUDA works in one env, not another

    Example session with this skill installed

    torch.cuda.is_available() is True in my conda env but False in the new venv I made. nvidia-smi works. Why?

    • Read your context and instructions
    • Compiled the gpu stack inspector

    Summary

    Ran uvx gpu-stack-inspector compare --python ~/envs/sapiens2/bin/python --python ~/envs/cpu-venv/bin/python. The machine is fine: NVIDIA driver 595.99.02 answers, supports CUDA up to 13.2, and sees the RTX 3050 Laptop GPU. The venv has a

    CPU-only build of torch, so CUDA can never be available there.

    FieldA: conda env (works)B: venv (fails)
    Python3.13.133.12.13
    torch2.11.0+cu1282.15.0+cpu
    torch buildCUDA 12.8CPU only
    nvidia-* CUDA wheels15 installednone

    Findings

    FAIL GSI-TORCH-002 (B): torch 2.15.0+cpu is a CPU-only build. This build has no CUDA support, so torch.cuda.is_available() is False there whatever the driver can do.

    WARN GSI-DIFF-001: the environments have different PyTorch builds (A CUDA 12.8, B CPU only), the most likely reason CUDA works in one and not the other.

    • PASS: the driver supports CUDA 12.8, which A's torch was built for. NOT TESTED: CUDA itself, since only metadata was read.

    Next step

    Install a CUDA build of torch in the venv using the selector on pytorch.org (Get Started), with a CUDA version up to 13.2, for example CUDA 12.8 to match the conda env. Then check with gpu-stack-inspector inspect --python ~/envs/cpu-venv/bin/python.

    I haven't installed anything. Want me to show the exact pip command first?

    Connects securely to your tools. The creator never sees your data.

    About this skill

    nvidia-smi shows your GPU, but PyTorch runs on the CPU. Or CUDA works in your conda environment and not in the new venv. This skill has your agent compare your Python environments with the open-source gpu-stack-inspector CLI, find the difference that matters, and suggest the exact fix, before anyone reinstalls drivers or wheels at random.

    What it does

    • Inspects several interpreters from one command and shows every field that differs: Python, venv or conda, the torch version and build (CUDA and its version, CPU-only, ROCm), and the nvidia-* CUDA wheels
    • Decides the torch build from package metadata, without importing torch
    • Checks the NVIDIA driver against the CUDA version torch was built for, using dated, source-tagged compatibility data
    • Every finding has a stable code, the evidence, what to do next and the command that checks it afterwards
    • A redacted report you can paste into a bug tracker

    Safety

    Read-only: it never installs, upgrades or changes anything, and never needs sudo. The skill tells your agent to show any install command and wait for your yes, and to run the runtime probe (which imports torch and uses GPU memory) only when you ask.

    Network

    The first run downloads gpu-stack-inspector from PyPI through uvx. The tool itself is offline.

    MIT licensed. Source and docs: https://github.com/BhuwanJoshi-01/gpu-stack-inspector

    Requires Python 3.11+ for the tool. Inspects Python 3.8+. Runtime probing requires explicit user consent as it may hang on broken systems. Redacts paths but requires local CLI access.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 11 days ago

    • Free with an Agensi account

    Requires

    Python
    uv (Python tool runner)

    Permissions

    Terminal / Shell
    Write Files
    Network Access
    Read Files
    Environment Variables

    Runs the read-only gpu-stack-inspector CLI in the terminal. It starts each Python interpreter you name with a fixed probe that reads package metadata (torch is imported only with probe --runtime), runs nvidia-smi and nvcc to read versions, and reads CUDA_VISIBLE_DEVICES. Network: one package download from PyPI through uvx.

    Listed11 days ago

    What's inside

    Frequently Asked Questions

    Instead of guessing, it compares the interpreters side by side and shows the evidence: the torch build from its version tag, torch/version.py, the CUDA libraries in torch/lib, the nvidia-* requirements and conda's records, plus the CUDA version the driver supports. Each finding says what to do next and how to check it afterwards.

    Any Python 3.8+ interpreter: venvs, conda environments and the system Python. The tool itself needs Python 3.11+ and runs on Linux, macOS and Windows.

    Yes. gpu-stack-inspector report --output gpu-report.md writes a Markdown (or JSON) report with home paths, user and host names, secrets and tokens redacted.

    Not by default: it reads metadata and never imports torch. Only probe --runtime imports torch and runs a tiny CUDA operation, which creates a CUDA context and uses some GPU memory; the skill runs it only when you ask.

    It is free: a SKILL.md that teaches your agent to find your interpreters, compare them, read the findings and propose the fix, without installing anything on its own. The CLI is MIT licensed on GitHub and installs from PyPI.

    New in LLM Ops