Skip to content
Skills Library
DevOps & cloud · Analysis

select-nvidia-accelerated-stack

/select-nvidia-accelerated-stack

Resolve and lock a CUDA, PyTorch, CuPy, or CUDA Python toolchain for one observed Linux or Windows NVIDIA worker. Use when a Literate AI Component selects an NVIDIA accelerator Flavor, when matching driver/toolkit/package/SM compatibility, or when diagnosing why a CUDA variant is ineligible.

Repository
literate-ai ↗
Licence
Apache-2.0
Last repository change
27.10.2026

Assessment and security

86/ 100Class: good
Format25 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content21 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
1
Scripts
0
HTML comments
0
URLs in SKILL.md
4
Deprecation

No signals detected. Repository changed on 27.10.2026, without references to retired models or APIs.

Licence

Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.

What it does and when to use it

Resolve and lock a CUDA, PyTorch, CuPy, or CUDA Python toolchain for one observed Linux or Windows NVIDIA worker. Use when a Literate AI Component selects an NVIDIA accelerator Flavor, when matching driver/toolkit/package/SM compatibility, or when diagnosing why a CUDA variant is ineligible.

Original author description

Resolve and lock a CUDA, PyTorch, CuPy, or CUDA Python toolchain for one observed Linux or Windows NVIDIA worker. Use when a Literate AI Component selects an NVIDIA accelerator Flavor, when matching driver/toolkit/package/SM compatibility, or when diagnosing why a CUDA variant is ineligible.

How it works

SKILL.md has no level-two headings · 52 lines, 446 words

SKILL.md preview
---
name: select-nvidia-accelerated-stack
description: Resolve and lock a CUDA, PyTorch, CuPy, or CUDA Python toolchain for one observed Linux or Windows NVIDIA worker. Use when a Literate AI Component selects an NVIDIA accelerator Flavor, when matching driver/toolkit/package/SM compatibility, or when diagnosing why a CUDA variant is ineligible.
metadata:
  author: Literate AI maintainers <literate-ai-maintainers@users.noreply.github.com>
---

# Select an NVIDIA accelerated stack

Resolve one exact stack from current official evidence. Never infer a GPU architecture
from its marketing name or treat the `CUDA Version` banner in `nvidia-smi` as an
installed toolkit.

1. Run `litai config paths` and load the selected worker's current typed observation
   from its reported `worker_observations` path. Require Linux or Windows,
   `nvidia_status: ok`, at least one device, exact driver version, per-device memory,
   and compute capability. A missing or degraded observation makes the CUDA variant
   ineligible before generation.
2. Load the Component's locked OS, architecture, language/Python ABI, accelerator,
   package, and build Flavors. Explicit specification and Flavor pins outrank this
   skill; reject contradictions rather than substituting another backend.
3. Consult the current official CUDA release notes and compatibility guide. Confirm the
   selected toolkit supports every targeted compute capability and the observed driver
   supports the toolkit. Minor-version compatibility has feature and PTX restrictions;
   when relying on it, compile explicit `sm_<major><minor>` targets and record that
   choice. Do not install a forward-compatibility package implicitly.
4. Resolve framework packages only from their official compatibility selectors:
   PyTorch's install matrix for the exact OS/package/Python combination, and CuPy's
   CUDA-specific distributions for NumPy/SciPy-compatible GPU arrays. NumPy and SciPy
   themselves remain CPU implementations; name a CuPy backend honestly and constrain
   code to the tested compatibility surface.
5. Detect `nvcc` and required native compilers before use. An absent toolkit may be
   satisfied by a locked wheel/runtime only when the selected library officially ships
   it and no native compilation is required. Never derive toolkit presence from the
   driver alone or mutate global/system packages without authorization.
6. Retain the consulted primary-source matrices as a typed compatibility document with
   exact package hashes, then run `litai worker resolve-nvidia --compatibility FILE
   --worker-id ID --python-abi ABI` plus explicit pins. Python validates the observation,
   driver floors, compute capabilities, ABI, packages, hashes, and selects the oldest
   eligible exact candidate. Feed that content-identified record into the Component lock
   and CycloneDX source/resolved SBOMs; do not reproduce selection mechanics in prose.
7. Generate a preflight plus a minimal device computation. Require known output and an
   independent CPU oracle; verify the selected GPU device actually executed the kernel.
   A runtime fallback to CPU is a failure for a CUDA-qualified variant.

Prefer the oldest currently supported toolkit/package combination that satisfies the
locked Component behavior and every observed device; this limits artifact churn without
overriding an explicit user pin. Re-resolve mutable upstream matrices at major rebuild
or release time and retain their fetched content identity with the lock.

Use primary sources: NVIDIA's CUDA
[compatibility guide](https://docs.nvidia.com/deploy/cuda-compatibility/),
[toolkit release notes](https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/), and
[architecture matrix](https://docs.nvidia.com/datacenter/tesla/drivers/cuda-toolkit-driver-and-architecture-matrix.html);
PyTorch's [local install selector](https://docs.pytorch.org/get-started/locally/);
CuPy's official installation guide; and NumPy/SciPy's Array API and interoperability
documentation. Do not use blogs or remembered tables as resolution authority.

Versions

The database has no other skills named select-nvidia-accelerated-stack.

Identical content appears in 1 other GitHub folders. We show one copy with the highest score.

Database snapshot: 13.11.2026 · package version 5884cd0d5b94 (SHA-256 digest of the entire package)

Downloads in the last 30 days: 0 · ranking

Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.

Similar skills

  • 96generate-nvidia-cuda-application
    Generate and test a C++ CUDA or Python CUDA application from a Literate AI specification after an exact NVIDIA stack has been selected. Use when the selected accelerator Flavor is nvidia-cuda and generated source must prove real device execution, deterministic output, portable…
  • 100pixi-install-nvidia
    Use when the user says "use pixi to install <some nvidia tool>" (or similar) and wants NVIDIA/CUDA/GPU packages installed via Pixi (no sudo/apt), e.g., CUDA toolkit pieces, cuDNN/NCCL, PyTorch CUDA builds, RAPIDS.
  • 100gpu-containers
    Use when the user needs to configure Docker/container environments for GPU workloads, mentions CUDA, ROCm, GPU passthrough, or asks about running PyTorch/TensorFlow in containers. Configures NVIDIA/AMD GPU runtime access, selects compatible base images, and validates…
  • 95hyperpod-version-checker
    Check and compare software component versions on SageMaker HyperPod cluster nodes - NVIDIA drivers, CUDA toolkit, cuDNN, NCCL, EFA, AWS OFI NCCL, GDRCopy, MPI, Neuron SDK (Trainium/Inferentia), Python, and PyTorch. Use when checking component versions, verifying CUDA/driver…
  • 100deepstream-inspect-stack
    Use when diagnosing NVIDIA DeepStream SDK installation, version compatibility, container tag, JetPack, CUDA, TensorRT, driver, GPU/Jetson, WSL2, pyservicemaker, or pyds issues; also use to reroute DeepStream pipeline-building or model-import requests to sibling skills without producing code.
  • 100tao-setup-nvidia-gpu-host
    Host setup for TAO GPU backends. Checks and, after user approval, installs minimum-compatible NVIDIA driver, CUDA Toolkit, and NVIDIA Container Toolkit versions for Docker/local-Docker and Kubernetes GPU worker hosts. TAO-wide defaults can be overridden by the selected model's…

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.