fla-orgOfficial·9 Agent Skills Included

flash-linear-attention

Optimize and test GPU kernels for linear attention models safely

Guides AI agents through optimizing, testing, and shipping GPU kernels for the Flash Linear Attention library. Covers backend dispatch, NVIDIA and Ascend profiling, correctness coverage, and pull request preparation. Eliminates silent numerical bugs, broken benchmarks, and rejected contributions through frozen test gates and checklists.

5,677|685

↗

npx skills add fla-org/flash-linear-attention --all -g -y

Available:

Tells the AI agent how to work safely in this repo, including git safety rules, when to ask before breaking changes, and which task-specific skill to load for kernel optimization, testing, or PR preparation.

All Skills in This Repository (9)

Pure Emerald Level Indicators

📦 In Repo

fla-org

fla-dispatch-backends

Manage and optimize backend dispatch for Flash Linear Attention operations.

fla-optimization-loop

Optimize Flash Linear Attention kernels across Triton, Gluon, TileLang, and CuTe backends.

fla-mr-readiness

Check pull requests against CONTRIBUTING guidelines and PR body format.

fla-kda

Manage KDA development tasks within the Flash Linear Attention framework.

fla-nvidia-performance

Profile and benchmark NVIDIA GPU kernels in the FLA repository.

fla-correctness-coverage

Automate kernel correctness testing and coverage analysis in FLA/ops/.

fla-triton-to-gluon

Port Triton kernels to Gluon with explicit tensor layout and memory control.

fla-design-coverage

Defines contract-first design and coverage requirements for FLA kernel and numerical changes.

fla-ascend-performance

Profile, diagnose, and optimize Triton-Ascend NPU kernels in the flash-linear-attention repository.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install flash-linear-attention skills?▼

Run `npx skills add fla-org/flash-linear-attention --all -g -y` in your terminal to install all skills in this suite globally.

How to optimize Triton kernels without breaking correctness?▼

The fla-optimization-loop skill enforces a frozen pytest gate so your agent benchmarks and logs each change while never loosening tolerances or editing tests.

How to profile GPU kernels on NVIDIA or Ascend?▼

Dedicated skills walk your agent through NCU profiling on NVIDIA GPUs and torch_npu PipeUtilization analysis on Ascend NPUs, with bottleneck diagnosis tables.

Do these skills work with Claude Code and Codex?▼

Yes. The AGENTS.md orchestrator and all SKILL.md files follow the universal standard and run in Claude Code, Codex, and compatible agents.

How to prepare a pull request for the FLA repo?▼

Load the fla-mr-readiness skill, which gives your agent the full checklist for tests, benchmark evidence, and the required PR body structure.