AI Infrastructure Agent Skills
A set of eight skills for AI infrastructure: TileLang GPU kernels, Megatron memory, SLIME training and diagrams.
- 158 ★ on the repository
- Last commit
Minimal access
Why this level
- The skills give advice and code; the author warns that some material was not proofread by hand.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/ai-infrastructure-agent-skillsDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skills give advice and code; the author warns that some material was not proofread by hand.
This is third-party code. Review the repository files before installing.
What it does
The repository holds eight skills, each in its own folder with a `SKILL.md`. Technical ones: `tilelang-developer` writes high-performance GPU kernels in TileLang for NVIDIA, AMD and Ascend (GEMM, FlashAttention, MLA), `megatron-memory-estimator` estimates memory for dense and MoE models from a HuggingFace config across TP, PP, EP and CP options, and `slime-user` guides RL training setup in SLIME. For presentation work there are `tikz-flowchart`, `hf-architecture-tikz`, `anthropic-theme-flowchart`, `material-you-slides` and `openai-dotcom-viz`: flowcharts and architecture diagrams in LaTeX TikZ, Material You slides and visualisations.
Who it is for. Model training and inference engineers who need an agent that knows narrow frameworks.
Good fit when
- A matrix multiplication kernel for GPU is written in TileLang.
- Before a training run, you check whether an MoE model fits in memory.
- RL training in SLIME is configured with GRPO or PPO.
- A model architecture diagram in TikZ is prepared for a talk.
Not a fit when
- Production-grade code is needed with no verification.
- The work uses a framework other than TileLang, Megatron and SLIME.
- A memory estimate must match measurements on a live cluster.
Example request
Estimate whether this HuggingFace MoE config fits in memory under different TP and PP values and show compression options.Limitations
The author states in the README that the project is under active development and largely generated by language models without strict proofreading, so verify code before production use. Some skills were built from Gemini DeepResearch material, and the README marks individual skills as Complete. The README has no install instructions: copy a skill by hand into the agent's skills directory. The last commit in the repository dates from July 2026.
How to disable. Delete the chosen skill folders from the agent's skills directory. The skills create no other data.
FAQ
Which skills are included?
Eight: `tilelang-developer`, `megatron-memory-estimator`, `slime-user`, `tikz-flowchart`, `hf-architecture-tikz`, `anthropic-theme-flowchart`, `material-you-slides`, `openai-dotcom-viz`.
Which GPUs does the TileLang skill target?
NVIDIA, AMD and Ascend: matrix multiplication, FlashAttention and DeepSeek MLA operators.
What does the Megatron memory estimate rely on?
HuggingFace configs and a comparison of TP, PP, EP and CP strategies; the basis comes from `megatron_memory_estimator` on HuggingFace Spaces.
How should generated code be verified?
Run the skill's examples on test data and compare them with official documentation: the author explicitly asks for manual checking.
How do I install a skill?
The README gives no install steps. Copy the folder of the needed skill with its `SKILL.md` into the agent's skills directory.
Discussion
No questions or reviews yet.
Related
SkillsAI Research Skills
AI Research Skills Library
98 skills for ML research, from ideation and model training to inference, evaluation and paper writing
Ten skills for LLM serving: SGLang and vLLM benchmarks, profiler triage, memory planning and PR history.
A local MCP server that stores skills as versions and lets improved variants go live only after a human review
Skills keep project memory in a docs/ai folder: short size-capped files, a one-page handoff and a verifier for the install
A skill for the "Early Experience" method (ICML 2026) that helps an agent prepare reward-free training data for language agents.
A CLI installer and package manager for coding agent skills: one command installs skills into many runtimes and helps build your own
README in short
For each skill the README lists capabilities and readiness status. Several skills are built on the documentation of third-party projects: TileLang, Megatron and SLIME.