Skip to main content
affpapa.org
RU EN
skill CX Codex

flashkda-delta-attention — Claude Code skill: install

This skill guides FlashKDA setup for Kimi Delta Attention CUDA kernels, with direct flash_kda.fwd use and chunk_kda integration.

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓ 
npx skills add reason-machines/trending-skills --skill flashkda-delta-attention -a claude-code
Open on GitHub

What flashkda-delta-attention does

The skill covers FlashKDA, a CUTLASS-based CUDA kernel package for Kimi Delta Attention and recurrent-state inference. It explains the low-level flash_kda.fwd call and the recommended path through flash-linear-attention, where chunk_kda can dispatch to the FlashKDA backend.

The setup needs an SM90-or-newer GPU such as H100 or H20, CUDA 12.9+, PyTorch 2.4+, and Python 3.8+. The flash-linear-attention integration at version 0.5.0 or newer is recommended. Hardware below SM90 does not meet the backend requirement.

The package build uses https://github.com/MoonshotAI/FlashKDA.git, initializes its submodules, and runs pip install -v .. Kernel inputs require K == V == 128; packed variable-length input with cu_seqlens requires batch size 1.

What it helps with

  • wire FlashKDA into chunk_kda
  • call the low-level flash_kda.fwd kernel
  • carry an initial state and collect the final state
  • run packed sequences with cu_seqlens

Example prompts for the agent

  • Use flashkda-delta-attention to connect FlashKDA to chunk_kda.
  • Show a flash_kda.fwd call that passes initial_state and returns final_state.
  • Check the GPU, CUDA, and tensor-shape requirements for a Kimi Delta Attention implementation.

How to install flashkda-delta-attention

1Skills CLI (Vercel)

npx skills add reason-machines/trending-skills --skill flashkda-delta-attention -a claude-code

2Manually: put the files in place

~/.claude/skills/flashkda-delta-attention/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/flashkda-delta-attention/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about flashkda-delta-attention

How do I install flashkda-delta-attention in Claude Code?

Run npx skills add reason-machines/trending-skills --skill flashkda-delta-attention -a claude-code, then check ~/.claude/skills/flashkda-delta-attention/SKILL.md.

What does flashkda-delta-attention do?

It provides guidance for using FlashKDA CUDA kernels with Kimi Delta Attention, either through flash_kda.fwd or the chunk_kda backend.

What constraints does FlashKDA have?

It requires an SM90+ GPU and K == V == 128. With cu_seqlens, batch size must be 1, and initial_state and final_state must use matching dtypes.

Similar skills

All skills for AI and ML →
nvidia-cuda Fandhe-AI This skill gives an agent CUDA, PTX, Blackwell, and CUTLASS references. Use it when working through GPU code or planning CUDA kernels. CX Codex +1 326installs spark-training-gotchas Seth Hobson Diagnoses ten known ML training failure modes on NVIDIA DGX Spark GB10. It catches memory leaks, power throttling, and CUDA mismatches early. CX Codex 2,500installs nemo-mbridge-perf-memory-tuning NVIDIA Helps diagnose GPU OOM failures in Megatron Bridge training and choose settings that reduce peak memory use. CX Codex +1 1,900installs nemo-mbridge-perf-moe-comm-overlap NVIDIA A skill for MoE expert-parallel communication overlap in Megatron Bridge: dispatch/combine, flex dispatcher backends, expert wgrad scheduling. CX Codex +1 1,800installs mcp-server-colab-exec pdwi2020 Runs Python on Google Colab GPU runtimes and can return generated files. Useful for CUDA, PyTorch, or TensorFlow work when no local GPU is available. CX Codex +1 ★ 9repo stars optimizing-attention-flash Orchestra Research A skill that moves transformer attention onto Flash Attention for 2-4x speedup and 10-20x memory reduction. CX Codex +1 967installs

From the same set186

permiso-accessibility-permissions A skill for coding agents that shows how to add a guided macOS accessibility permission panel to a Swift app. CX Codex +1 386installs text-to-cad-harness This skill has your coding agent write build123d Python parts from a description, export them as STEP, STL or URDF, and open the geometry in a local viewer. CX Codex +1 380installs gpt2api-openai-gateway This skill walks your agent through deploying a self-hosted Anthropic-compatible gateway over chatgpt.com and wiring up an account pool. CX Codex +1 378installs openai-privacy-filter A skill that wires a token-classification model into your project to find and blank out personal data. Useful before shipping logs or documents to an LLM. CX Codex +1 369installs masterhttprelayvpn-proxy A skill that walks your agent through running a local HTTP/SOCKS5 proxy, with traffic tunnelled through Google Apps Script and dressed up as Google requests. CX Codex +1 365installs tilekernels-gpu-kernels This skill guides agents through TileKernels GPU kernels for MoE, quantization and other LLM operations. Use it when writing or checking TileLang code. CX Codex +1 363installs aris-autonomous-ml-research A Markdown skill that links ML idea discovery, experiments, paper writing, review, and rebuttal. It fits workflows where one model checks another. CX Codex +1 360installs freellmapi-proxy This skill sets up a self-hosted Anthropic-compatible proxy that puts free-tier keys from many providers behind one endpoint. CX Codex +1 351installs

Project sponsors

Companies that keep this analytics open
Gold sponsor
Nutra.Club

Impressions, clicks and code copies — open statistics. We keep the numbers public for the same reason NeBlask itself is public. Become a sponsor