Swiss AI
Official@swiss-ai · Switzerland
Swiss AI maintains engineering skills for the FLA (flash-linear-attention) repository, covering GPU kernel optimization, correctness testing, backend dispatch, and merge-request readiness.
Agent Skills by Swiss AI
Showing 7 vetted skills indexed across 1 GitHub repositories.
swiss-aifla-dispatch-backends
Implements and tests runtime backend dispatch for FLA operations using BaseBackend subclasses and verifiers.
swiss-aifla-optimization-loop
Guides reproducible multi-round optimization of FLA kernels under a frozen pytest correctness gate.
swiss-aifla-mr-readiness
Validates pull requests against CONTRIBUTING.md, test plans, and benchmark evidence requirements.
swiss-aifla-kda
Guides modification and review of KDA kernels in the Flash Linear Attention library.
swiss-aifla-nvidia-performance
Guides NVIDIA GPU kernel profiling and performance evidence collection for FLA repository changes.
swiss-aifla-correctness-coverage
Guides correctness testing and coverage analysis for Triton kernels in fla/ops.
swiss-aifla-triton-to-gluon
Ports Triton kernels in fla/ops to Gluon for explicit control over layouts, shared memory, and MMA scheduling.
Frequently Asked Questions About Swiss AI
FAQPage SchemaWhat tasks can I perform using Swiss AI's FLA skills?▼
You can optimize FLA GPU kernels across Triton, Gluon, TileLang, and CuTe, port Triton kernels to Gluon, implement backend dispatch decorators and BaseBackend subclasses, modify KDA gate and chunk kernels, add correctness tests for grid/addressing pitfalls, and prepare benchmark-backed merge requests.
Who are these FLA skills designed for?▼
They target GPU kernel engineers and ML systems contributors working in the flash-linear-attention repository—specifically developers touching fla/ops/**, KDA kernels, backend dispatch layers, or NVIDIA performance work who need disciplined correctness gates and reproducible optimization loops.
How does the FLA optimization loop ensure kernels stay correct?▼
It anchors every iteration on FLA's frozen pytest suite covering both forward and backward passes under NaN poisoning as the immutable correctness gate, using a task-contract, three-phase iteration protocol, and a silent-bug catalog drawn from kernel-optimization frameworks like KDA and FlashInfer contests.
When should I port a Triton kernel to Gluon using these skills?▼
Port when a kernel is register-bound, when num_stages pipelining underperforms, or when you need Hopper/Blackwell features like TMA, TMEM, tcgen05, cp.async, or warp specialization. The skill provides an incremental sequence preserving numerical parity plus a Triton-to-Gluon API mapping.
What evidence is required before submitting an FLA merge request?▼
MRs require CONTRIBUTING.md compliance, a defined test plan, benchmark evidence, and a structured PR body. NVIDIA performance changes additionally need profiling workflow documentation, hardware baselines, and Nsight Compute collection via the installed ncu-report-skill.