Agent skill

Tilelang Skill

by slowlyC in slowlyC/agent-gpu-skills

Write, debug, and optimize TileLang kernels from local upstream language, JIT, autotuning, profiling, compiler, test, and example source.

MITAuto-check passedDevelopment

Install Tilelang Skill

skills CLI
$ npx skills add slowlyC/agent-gpu-skills --skill tilelang-skill -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install slowlyC/agent-gpu-skills tilelang-skill --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/slowlyC/agent-gpu-skills.git skills-src && mkdir -p .claude/skills && cp -r skills-src/skills/tilelang-skill .claude/skills/tilelang-skill && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
tilelang-skill
GitHub stars
169
Token cost
~1.8k tokens
SKILL.md length
590 words
Files
2
Skills in repo
4
Repo updated
First seen
Licence
MIT

At a glance

Write, debug, and optimize TileLang kernels from local upstream language, JIT, autotuning, profiling, compiler, test, and example source.

  • Works in 5 steps: Classify the task as kernel DSL,… → Find the closest current example for the… → Verify each DSL operation against its… → …
  • The task explicitly involves tilelang
  • SKILL.md covers Locate the checkout, Choose the source surface, Query workflow and Implementation discipline, plus 2 more sections
  • Calls rg, bash and python3

What it does

Tilelang Skill is an agent skill from slowlyC/agent-gpu-skills. Write, debug, and optimize TileLang kernels from local upstream language, JIT, autotuning, profiling, compiler, test, and example source. Use when the task explicitly involves tilelang, tilelang.language, @tilelang.jit, @T.primfunc, T.Kernel, T.copy, T.gemm, TileLang Profiler, Carver, TileLang passes, or TileLang CUDA, ROCm, Metal, and CPU backends. Use triton-skill for Triton or Gluon, cutlass-skill for direct CUTLASS, CuTe, or CuTeDSL work, and cuda-skill for raw CUDA, PTX, NVIDIA architecture, or…

Its SKILL.md is about 1.8k tokens, which your agent loads only when the skill is triggered. The skill folder holds 1 other file (for example `quick-reference.md`).

It sits in Development, covering Performance optimization. It works with CUDA and NVIDIA AI Platform. The licence is MIT.

When your agent uses it

  • The task explicitly involves tilelang
  • Tilelang.language
  • TileLang Profiler
  • TileLang passes

Example prompts

  • “/tilelang-skill”

Requirements

  • Python 3

Workflow steps

5 steps, taken from the first numbered list in SKILL.md.

  1. Classify the task as kernel DSL, JIT/runtime, autotuning/profiling, or compiler/backend work.
  2. Find the closest current example for the operation, target backend, architecture, dtype, and shape regime.
  3. Verify each DSL operation against its Python definition, then trace lowering or backend code only as far as the question requires.
  4. Preserve the workload's shapes, dtypes, layouts, memory scopes, target, launch configuration, and numerical contract.
  5. Establish correctness and confirm the actual generated path before tuning.

What it can do on your machine

Read from SKILL.md and the folder at commit ae02d07. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    Shell commands in SKILL.md call:

    • rg
    • bash
    • python3

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    No URLs in SKILL.md.

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Tilelang Skill loads about 1.8k tokens when it runs. Until then it costs about 136 tokens; SKILL.md has 590 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~136
When it runs · the whole SKILL.md, loaded when a task matches
~1.8k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from slowlyC/agent-gpu-skills at commit ae02d07, republished under its MIT licence (© slowlyC). 590 words, ~1,815 tokens.

Download SKILL.mdSave it as .claude/skills/tilelang-skill/SKILL.md (or your agent's skills folder). This skill also uses 1 other file; get the full folder from GitHub.
name
tilelang-skill
description
Write, debug, and optimize TileLang kernels from local upstream language, JIT, autotuning, profiling, compiler, test, and example source. Use when the task explicitly involves tilelang, tilelang.language, @tilelang.jit, @T.prim_func, T.Kernel, T.copy, T.gemm, TileLang Profiler, Carver, TileLang passes, or TileLang CUDA, ROCm, Metal, and CPU backends. Use triton-skill for Triton or Gluon, cutlass-skill for direct CUTLASS, CuTe, or CuTeDSL work, and cuda-skill for raw CUDA, PTX, NVIDIA architecture, or profiling-tool facts.

TileLang development

Use the local TileLang checkout as the primary source for APIs, implementation patterns, and repository-specific build instructions. Prefer current source and examples over remembered signatures because the DSL, compiler, and backend integrations evolve quickly.

Locate the checkout

Resolve the directory containing this SKILL.md, then use its repos/tilelang/ child. The installer links that path to agent-gpu-skills/third_party/tilelang/ or to the checkout supplied through TILELANG_REPO.

In commands below, replace TILELANG_REPO with the resolved absolute path:

bash
TILELANG_REPO=/absolute/path/to/tilelang-skill/repos/tilelang

If the checkout is missing, run this from the agent-gpu-skills repository and reinstall the Skill:

bash
bash update-repos.sh tilelang
bash install.sh --skill tilelang-skill

The default checkout is a source-reference snapshot. It does not initialize TileLang's nested third-party submodules. Use a complete development checkout through TILELANG_REPO when building TileLang or changing its compiler.

Choose the source surface

TaskStart here
Language syntax and operationstilelang/language/
JIT wrappers, adapters, and kernel objectstilelang/jit/
Lowering and semantic checkstilelang/engine/
Autotuningtilelang/autotuner/
Benchmarking and profiler helperstilelang/profiler/
Layout inference and representationstilelang/layout/, tilelang/analysis/
Schedule recommendation and Carvertilelang/carver/
CUDA, ROCm, Metal, and CPU Python backendstilelang/cuda/, tilelang/rocm/, tilelang/metal/, tilelang/cpu/
Compiler transforms and target code generationsrc/
Current kernels and end-to-end workloadsexamples/
Language, compiler, backend, and regression teststesting/python/
Repository-specific build and test commands.agents/skills/tilelang-build/SKILL.md

Read quick-reference.md when mapping an operation, backend, or compiler question to a concrete path in the validated checkout.

Query workflow

  1. Classify the task as kernel DSL, JIT/runtime, autotuning/profiling, or compiler/backend work.
  2. Find the closest current example for the operation, target backend, architecture, dtype, and shape regime.
  3. Verify each DSL operation against its Python definition, then trace lowering or backend code only as far as the question requires.
  4. Preserve the workload's shapes, dtypes, layouts, memory scopes, target, launch configuration, and numerical contract.
  5. Establish correctness and confirm the actual generated path before tuning.

Discover current documentation and examples before relying on a remembered filename:

bash
find "$TILELANG_REPO/docs" -type f -name '*.md' | sort
find "$TILELANG_REPO/examples" -type f -name '*.py' | sort

Search kernel patterns and their definitions together:

bash
rg -n '@tilelang\.jit|@T\.prim_func|T\.Kernel|T\.Pipelined' \
  "$TILELANG_REPO/examples"

rg -n 'T\.copy|T\.gemm|T\.alloc_shared|T\.alloc_fragment' \
  "$TILELANG_REPO/examples"

rg -n '^def (copy|gemm|alloc_shared|alloc_fragment)|class Kernel' \
  "$TILELANG_REPO/tilelang/language"

Trace JIT, lowering, and target selection:

bash
rg -n 'def (compile|lower)|class .*Kernel|class .*Adapter' \
  "$TILELANG_REPO/tilelang/jit" \
  "$TILELANG_REPO/tilelang/engine"

rg -n 'target|backend|codegen' \
  "$TILELANG_REPO/tilelang/backend" \
  "$TILELANG_REPO/tilelang/cuda" \
  "$TILELANG_REPO/tilelang/rocm" \
  "$TILELANG_REPO/tilelang/metal" \
  "$TILELANG_REPO/tilelang/cpu"

Trace compiler transformations and generated source:

bash
rg -n 'Pass|Lower|Legalize|Layout' \
  "$TILELANG_REPO/src/transform" \
  "$TILELANG_REPO/tilelang/engine"

rg -n 'CodeGen|codegen|Build' \
  "$TILELANG_REPO/src/cuda" \
  "$TILELANG_REPO/src/rocm" \
  "$TILELANG_REPO/src/metal" \
  "$TILELANG_REPO/src/cpu"

Search tuning, profiling, and nearby tests:

bash
rg -n 'autotune|Profiler|do_bench|get_profiler' \
  "$TILELANG_REPO/examples" \
  "$TILELANG_REPO/tilelang/autotuner" \
  "$TILELANG_REPO/tilelang/profiler"

rg -n 'copy|gemm|pipelined|layout' \
  "$TILELANG_REPO/testing/python/language" \
  "$TILELANG_REPO/testing/python"
Show full SKILL.md (261 more words)Show less

Implementation discipline

Keep these layers separate during diagnosis:

text
TileLang Python kernel and compile-time parameters
  → TileLang IR, semantic checks, and compiler transforms
  → backend source generation and native compilation
  → runtime dispatch on the selected target

A TileLang source construct does not prove which hardware instruction or memory path the generated kernel uses. Inspect get_kernel_source(), compiler traces, or profile data when that distinction matters. Add cuda-skill for PTX semantics, compute capability, Nsight, or Compute Sanitizer details.

TileLang contains a CuTeDSL integration under tilelang/contrib/cutedsl/. Use this Skill when tracing how TileLang selects or calls that integration. Use cutlass-skill when the task is about CuTeDSL APIs or implementations themselves.

For correctness work:

  • compare against an independent reference over representative and boundary shapes;
  • test tails, dynamic dimensions, layouts, dtype conversions, and backend-specific paths;
  • separate DSL parsing, lowering, native compilation, runtime, and numerical failures;
  • reproduce the original target and compile-time parameters before minimizing the case.

For performance work:

  • freeze the benchmark shape set, warmup, repetition count, and synchronization method;
  • confirm the measured call dispatches to the intended compiled kernel;
  • inspect generated source before attributing a result to TMA, tensor cores, tcgen05, or a CuTeDSL path;
  • change one tile, thread count, stage count, layout, or pass configuration at a time.

Build and test routing

Before building, installing, or testing TileLang itself, read the upstream repository instructions:

text
.agents/skills/tilelang-build/SKILL.md

Those instructions are versioned with the checkout and are authoritative for the current build commands. Most compiler and kernel tests require an appropriate device. Verify the active import path, build directory, target backend, and GPU before interpreting a result.

Updating the source

From the agent-gpu-skills repository:

bash
bash update-repos.sh tilelang
python3 scripts/validate_repo.py --require-sources

The checkout follows TileLang main, while third_party/UPSTREAMS.toml records the commit last accepted by this Skill. Review source-map drift before updating that record.

© slowlyC, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 1 other file in skills/tilelang-skill of slowlyC/agent-gpu-skills.

  • SKILL.md
  • quick-reference.md

Open the folder on GitHubat commit ae02d07

Compare with similar skills

Tilelang Skill next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Tilelang Skill compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Tilelang Skill this skillslowlyC/agent-gpu-skills169—~1.8kAutomated safety check: PassMIT
Nemo Mbridge Perf Moe Optimization WorkflowNVIDIA/skills3.5k—~3.3kAutomated safety check: PassApache-2.0
LLM Torch Profiler Trace AnalysisBBuf/AI-Infra-Auto-Driven-SKILLS925—~2.8kAutomated safety check: PassNone
Cuda Profilingmohitmishra786/low-level-dev-skills253—~1.6kAutomated safety check: NotesMIT
Graphsignalgraphsignal/graphsignal257—~6.2kAutomated safety check: PassApache-2.0
LLM Torch Profiler Analysissgl-project/sglang37k2 repos~6.4kAutomated safety check: PassApache-2.0

Similar skills

  • LLM Torch Profiler Trace Analysis

    BBuf/AI-Infra-Auto-Driven-SKILLS

    Analyzes Torch Profiler traces from SGLang, vLLM and TensorRT-LLM servers into kernel attribution, overlap and fusion tables.

    925 GitHub stars~2.8k tokensUpdated 3 days ago
    AI & LLM EngineeringAuto-check passed
  • Cuda Profiling

    mohitmishra786/low-level-dev-skills

    CUDA profiling skill for NVIDIA GPU performance analysis. An agent skill from mohitmishra786/low-level-dev-skills.

    253 GitHub stars~1.6k tokensUpdated 3 mo ago
    AI & LLM EngineeringAuto-check: notes
  • Graphsignal

    graphsignal/graphsignal

    Profile AI inference workloads (vLLM, SGLang, TensorRT-LLM, PyTorch, any GPU application) with the Graphsignal profiler and read the results from its local /signals JSON endpoint.

    257 GitHub stars~6.2k tokensUpdated 11 days ago
    AI & LLM EngineeringAuto-check passed
  • LLM Torch Profiler Analysis

    sgl-project/sglang

    Unified LLM torch-profiler triage skill for sglang, vllm, TensorRT-LLM, and TokenSpeed.

    37k GitHub starsUsed in 2 repos~6.4k tokens
    DevelopmentAuto-check passed
  • The Art of Debugging

    stas00/the-art-of-debugging

    Condensed debugging method and tool recipes for Unix, Python and PyTorch programs: crashes, hangs, segfaults, wrong output, CUDA OOM, NaN values and slowness.

    1.7k GitHub stars~6.1k tokensUpdated 2 days ago
    DevelopmentAuto-check: notes

More from slowlyC/agent-gpu-skills

  • Cutlass Skill

    slowlyC/agent-gpu-skills

    Write, debug, and optimize CUTLASS, CuTe, and CuTeDSL GPU kernels from local upstream source, examples, and headers.

    169 GitHub stars~1.3k tokensUpdated 2 mo ago
    Auto-check passed
  • Triton Skill

    slowlyC/agent-gpu-skills

    Write, debug, and optimize Triton and Gluon GPU kernels from local upstream tutorials, production kernels, language definitions, and compiler source.

    169 GitHub stars~1.3k tokensUpdated 2 mo ago
    Auto-check passed
  • Cuda Skill

    slowlyC/agent-gpu-skills

    Query current NVIDIA CUDA, PTX ISA, Runtime API, Driver API, Programming Guide, Best Practices, Nsight Compute, and Nsight Systems references.

    169 GitHub stars~1.8k tokensUpdated 2 mo ago
    Auto-check passed

Categories

Questions about Tilelang Skill

What does Tilelang Skill do?

Write, debug, and optimize TileLang kernels from local upstream language, JIT, autotuning, profiling, compiler, test, and example source. Tilelang Skill is an agent skill from slowlyC/agent-gpu-skills. Write, debug, and optimize TileLang kernels from local upstream language, JIT, autotuning, profiling, compiler, test, and example source.

When should I use Tilelang Skill?

Tilelang Skill fits situations like: the task explicitly involves tilelang; tilelang.language; tileLang Profiler; tileLang passes.

How do I install Tilelang Skill in Claude Code?

Run `npx skills add slowlyC/agent-gpu-skills --skill tilelang-skill -a claude-code`. Or copy the skill folder (skills/tilelang-skill in slowlyC/agent-gpu-skills) into .claude/skills/tilelang-skill in your project. Claude Code loads it when a task matches its description.

How do I install Tilelang Skill in Codex?

Run `npx skills add slowlyC/agent-gpu-skills --skill tilelang-skill -a codex`. Or copy the skill folder (skills/tilelang-skill in slowlyC/agent-gpu-skills) into .agents/skills/tilelang-skill in your project. Codex loads it when a task matches its description.

Can I use Tilelang Skill in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add slowlyC/agent-gpu-skills --skill tilelang-skill -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/tilelang-skill, .gemini/skills/tilelang-skill, .github/skills/tilelang-skill and .opencode/skills/tilelang-skill in your project.

What does Tilelang Skill need to run?

Going by SKILL.md and its folder, Tilelang Skill needs the command-line tools its instructions call (rg, bash and python3). Our summary lists: Python 3.

Does Tilelang Skill access the network?

SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.

Is Tilelang Skill safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.

What licence does Tilelang Skill use?

Tilelang Skill is published under the MIT licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Tilelang Skill use?

About 1.8k tokens (SKILL.md is roughly 7.3k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full.

What are the alternatives to Tilelang Skill?

Skills that share tags, products or a category with Tilelang Skill: Nemo Mbridge Perf Moe Optimization Workflow (NVIDIA/skills, 3.5k stars), LLM Torch Profiler Trace Analysis (BBuf/AI-Infra-Auto-Driven-SKILLS, 925 stars), Cuda Profiling (mohitmishra786/low-level-dev-skills, 253 stars) and Graphsignal (graphsignal/graphsignal, 257 stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Tilelang Skill?

slowlyC (a GitHub user) maintains it in slowlyC/agent-gpu-skills, which has 169 GitHub stars. The repository holds 4 skills in this directory. The repository was last updated on August 8, 2026.

Source: slowlyC/agent-gpu-skills on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.