bk99.de entertain the web since 1997

bk99.de
HF · GPUCurated

Developing production-ready CUDA kernels

Summary

The guide leads from a simple CUDA kernel to tests, benchmarks, variants and automated delivery. It treats kernel work as a software product rather than a one-off speed hack. The workflow covers implementation, tests, benchmarks and publication.

Ideas

  • A fast kernel needs correctness tests across shapes, data types and devices.
  • Autotuning shifts optimisation from a fixed assumption to measurable hardware variants.

Insights

  • Only specialised kernels turn mathematical possibilities into measurable hardware performance.

Facts

  • Several kernel variants can be selected depending on the input shape.

Critique

  • Benchmark gains on one GPU can disappear or reverse on another architecture.

Recommendations

  • Compare every kernel against a numerically stable reference and test edge sizes automatically.

References

Read the original article on Hugging Face

Search the Web Archive

HF · 2025

Reusing GPU kernels via the Hub

The Kernel Hub distributes optimised GPU operators as versionable artefacts and loads suitable variants at runtime. This allows specialised kernels to be used without building them into every package. HF Kernels provides pre-built and custom GPU kernels via the Hub.

HF · 2026

Security incident at Hugging Face in July 2026

Hugging Face describes a break-in in which an autonomous agent combined a prepared data source with insecure execution paths. Internal data sets and credentials were reached; according to what was known at the time, public Hub artefacts were not manipulated. The attack combined a malicious data set with remote code execution and template injection.

HF · 2026

Moving GitHub CI to Hugging Face Jobs

The guide moves compute-intensive CI steps from GitHub Actions to Hugging Face Jobs. GitHub remains the trigger and status interface, while specialised hardware runs externally. Hugging Face Jobs can run GPU and CPU tasks from GitHub workflows.