cuda-attention-kernel-patterns Skill
Patterns and pitfalls for the ONNX domain Attention operator (opset 23/24) CUDA implementation. Use when modifying the dispatch cascade in core/providers/cuda/llm/attention.cc, writing mask/bias CUDA kernels, debugging attention test routing, or adding features to the ONNX Attention op. NOT for contrib domain MultiHead Published by microsoft in onnxruntime.
What is cuda-attention-kernel-patterns Skill?
Patterns and pitfalls for the ONNX domain Attention operator (opset 23/24) CUDA implementation. Use when modifying the dispatch cascade in core/providers/cuda/llm/attention.cc, writing mask/bias CUDA kernels, debugging attention test routing, or adding features to the ONNX Attention op. NOT for contrib domain MultiHead Published by microsoft in onnxruntime. This profile combines repository metadata with install, compatibility, and usage signals so developers can quickly decide whether it fits their agent workflow before opening the source repository.
Automated repository signals based on public metadata such as recency, license, installation evidence, and adoption. These are not a security audit or endorsement.
Key capabilities
- Includes SKILL.md support
- Reusable instructions support
- Testing
- Writing
- Testing use cases
- Writing use cases
Technical details
- Install or run with Copy skill directory
When to use cuda-attention-kernel-patterns Skill
- Use it for testing.
- Use it for writing.
Built with
Editorial notes
Source
- Creator: microsoft
- Repository: microsoft/onnxruntime
- Skill file: .agents/skills/cuda-attention-kernel-patterns/SKILL.md
What it does
Patterns and pitfalls for the ONNX domain Attention operator (opset 23/24) CUDA implementation. Use when modifying the dispatch cascade in core/providers/cuda/llm/attention.cc, writing mask/bias CUDA kernels, debugging attention test routing, or adding features to the ONNX Attention op. NOT for contrib domain MultiHead
Skill instructions
ONNX Domain Attention (Opset 23/24) CUDA Patterns Reusable knowledge from ONNX Attention CUDA development in ORT. Scope: This skill covers the ONNX domain Attention operator (opset 23/24) implemented at core/providers/cuda/llm/attention.cc. This is separate from the contrib domain MultiHeadAttention / GroupQueryAttention at contribops/cuda/bert/. They share some underlying kernels (CUTLASS FMHA, Flash Attention) and infrastructure (attentionsoftmax.h) but have different dispatch logic, parameter structs, and eligibility checks. - Shared infrastructure: CUTLASS FMHA kernel, Flash kernel, unified unfused kernel (unfusedattention.cu), attentionsoftmax.h, attentionimpl.cu (contrib only) - ONNX-specific: Dispatch cascade in attention.cc, ConvertAttnMaskToBias, maskfiltervalue cap, parameter bridge to contrib structs, attentionmaskimpl.cu - Contrib-specific: Own dispatch in contrib MHA/GQA ops, uses contrib::AttentionParameters directly, has XQA kernel, past-present buffer sharing 1. Runner
Explore related resources
Frequently asked questions
What is cuda-attention-kernel-patterns?
cuda-attention-kernel-patterns is a open-source AI agent skill with Copy skill directory. Patterns and pitfalls for the ONNX domain Attention operator (opset 23/24) CUDA implementation.
Who is cuda-attention-kernel-patterns best for?
cuda-attention-kernel-patterns is best for reusing agent instructions, scripts, and references, testing workflows, writing workflows.
How do I install cuda-attention-kernel-patterns?
Install or run cuda-attention-kernel-patterns using Copy skill directory. Check cuda-attention-kernel-patterns for the latest setup command.
Is cuda-attention-kernel-patterns actively maintained?
cuda-attention-kernel-patterns may need a closer maintenance check before production use.
