Triton Tutor Setup — Build an Obsidian Triton StudyVault
CWD Boundary Rule (ALL MODES)
NEVER access files outside the current working directory (CWD).
All source scanning, reading, and vault output MUST stay within CWD and its subdirectories.
If the user provides an external repo path, ask them to either clone or symlink it into CWD first.
Mode Detection
On invocation, auto-detect the mode:
- Scan CWD for code markers:
setup.py, pyproject.toml, CMakeLists.txt, *.py containing import triton / @triton.jit / triton.language as tl, triton/ source dir, python/triton/, lib/Dialect/Triton, third_party/nvidia, third_party/amd.
- Scan CWD for document markers:
*.pdf, *.txt, *.md (excluding README.md), *.html, *.epub, *.ipynb (tutorials).
- Decide:
- Triton codebase markers found (
@triton.jit/triton.language/lib/Dialect/Triton) → Codebase Mode.
- No code markers AND document markers found → Document Mode.
- Neither found (empty dir or only
.git/) → Curriculum Mode (the default for this skill).
- Announce detected mode and ask the user to confirm or override (they may force any mode).
Default bias: when in doubt between Document mode and Curriculum mode, choose Curriculum Mode — it is this skill's headline use case.
Curriculum Mode (default)
Generates a complete Triton learning StudyVault organized around the 6-topic curriculum.
Full workflow: see curriculum-workflow.md.
Per-topic content lives in references/topic-*.md (loaded only for the topic being authored).
Templates: templates.md.
Topic List & Order (fixed)
The 6 topics, ordered by prerequisite chain. Build the StudyVault in this order. Do not reorder.
| # |
Topic |
Reference File |
Vault Folder |
| 1 |
Triton Basics (foundation) |
topic-triton-basics.md |
01-Triton-Basics/ |
| 2 |
Tiling & Autotuning |
topic-tiling-autotuning.md |
02-Tiling-Autotuning/ |
| 3 |
Matmul Patterns |
topic-matmul-patterns.md |
03-Matmul-Patterns/ |
| 4 |
Attention & Reductions |
topic-attention-reductions.md |
04-Attention-Reductions/ |
| 5 |
Compiler Internals |
topic-compiler-internals.md |
05-Compiler-Internals/ |
| 6 |
Ecosystem & Production |
topic-ecosystem-production.md |
06-Ecosystem-Production/ |
Prerequisite chain (encode in the Dashboard's prereq DAG):
┌────────────────────────────┐
│ 1. Triton Basics │
│ (@triton.jit, masks, │
│ tl.load/store) │
└──────┬──────────────────────┘
│
┌──────┴──────┐
▼ ▼
┌──────────┐ ┌──────────────┐
│ 2. Tiling│ │ 3. Matmul │
│ & Auto- │◀▶│ Patterns │
│ tuning │ │ (tl.dot) │
└─────┬────┘ └───────┬──────┘
│ │
└───────┬───────┘
▼
┌─────────────────────┐
│ 4. Attention & │
│ Reductions │
│ (FlashAttention) │
└────────┬────────────┘
│
┌──────┴──────┐
▼ ▼
┌────────────┐ ┌──────────────┐
│ 5. Compiler│ │ 6. Ecosystem │
│ Internals│ │ & Production│
│ (TTIR/TTGIR│ │ (Inductor, │
│ /PTX) │ │ proton) │
└────────────┘ └──────────────┘
Phase Summary (TU1-TU9)
| Phase |
Name |
Key Action |
| TU1 |
Pre-flight |
Ask user about target hardware (NVIDIA A100/H100/B200, AMD MI250/MI300, Intel PVC), prior knowledge (PyTorch user / CUDA writer / Triton kernel author), language preference. Skip topics already mastered if user opts in. |
| TU2 |
Topic Plan |
Build the topic dependency DAG. Materialize the prereq chain in the dashboard. |
| TU3 |
Tag Standard |
Define tag registry: #topic-triton-basics, #topic-tiling-autotuning, #topic-matmul-patterns, #topic-attention-reductions, #topic-compiler-internals, #topic-ecosystem-production, plus #concept-*, #milestone-*, #pitfall-*. English kebab-case. |
| TU4 |
Vault Structure |
Create StudyVault/ with 00-Dashboard/ + the 6 numbered topic folders + 99-Exercises/. |
| TU5 |
Dashboard |
MOC + Prereq DAG + Glossary + Quick Reference + Pitfall Index. |
| TU6 |
Per-Topic Notes |
For each of the 6 topics: load references/topic-{slug}.md (lazy, one at a time to manage context), then generate concept notes per templates.md. For topics with a CUDA/CUTLASS/cuTile peer (Triton Basics, Tiling/Autotuning, Matmul Patterns, Attention/Reductions, Compiler Internals), the topic reference file ends with a "Cross-Stack Equivalent" appendix table — surface this in the generated vault notes so vault readers see the Rosetta mapping while studying the Triton-side concept. The full Rosetta is at ../tutor-core/references/cross-stack-rosetta.md (consumed by triton-tutor at quiz time via symlink). |
| TU7 |
Hands-On Milestones |
Per topic, materialize 3-7 "build/run/measure X" milestones as standalone notes with success criteria. |
| TU8 |
Pitfall Notes |
Per topic, materialize the pitfalls section as Pitfalls.md with fold callouts. |
| TU9 |
Interlinking + Self-Review |
Cross-link all notes per the prereq DAG, then verify against quality-checklist.md Curriculum Mode section. |
See curriculum-workflow.md for the full procedural detail of each phase.
Lazy-Loading Discipline for Topic Content (IMPORTANT)
The 6 topic reference files (topic-*.md) are large. Load them one at a time in Phase TU6 — read topic-triton-basics.md while authoring 01-Triton-Basics/, drop it from working memory before reading topic-tiling-autotuning.md, etc. Do not pre-load all 6.
Codebase Mode
Generate an onboarding StudyVault from a real Triton / Inductor / kernel-library codebase.
Full workflow: codebase-workflow.md. Templates: triton-codebase-templates.md.
Phase Summary (C1-C9)
| Phase |
Name |
Key Action |
| C1 |
Project Exploration |
Detect Triton version (parse setup.py/pyproject.toml/requirements.txt), target backends (NVIDIA / AMD / Intel via third_party/), identify role of the repo (kernel library / compiler fork / inference engine / training framework). |
| C2 |
Architecture Analysis |
Trace a representative kernel from Python entry (kernel[grid](...)) through @triton.jit compile cache, autotune flow, generated IR (dump with TRITON_KERNEL_DUMP=1), into PTX/AMDGCN. Identify host-side wrappers and PyTorch integration points. |
| C3 |
Tag Standard |
Same kebab-case rules; add #backend-*, #arch-*, #kernel-role-*, #api-*, #ir-stage-* categories. |
| C4 |
Vault Structure |
Dashboard, Architecture, per-module, DevOps (build flags, CI, benchmarks), Exercises. |
| C5 |
Dashboard |
MOC + Module Map + Kernel Surface table + Backend Matrix + Getting Started (build & run a representative benchmark). |
| C6 |
Module Notes |
Per-module note: Purpose, Key Kernels, Python Wrappers, Autotune Configs, Block Sizes, IR Lowering Notes, Dependencies, How to Test. |
| C7 |
Exercises |
Code reading (trace a kernel + dump IR), config (change BLOCK_SIZE, swap backend), debugging (TRITON_INTERPRET, tl.device_print, IR dump), extension (add a new kernel or autotune config). |
| C8 |
Interlinking |
Cross-link Python wrapper ↔ @triton.jit kernel ↔ autotune config ↔ test. |
| C9 |
Self-Review |
quality-checklist.md Codebase Mode section. |
Common Triton-adjacent codebases this mode targets: triton-lang/triton, pytorch/pytorch (Inductor codegen), Dao-AILab/flash-attention (Triton variants), facebookresearch/xformers, linkedin/Liger-Kernel, unslothai/unsloth, pytorch-labs/applied-ai, meta-pytorch/tritonbench, openai/triton examples, plus user repos containing @triton.jit kernels.
Document Mode
Transform PDFs (Triton papers, OpenAI/Meta blog posts in PDF, conference slides on Triton), text, and web pages into study notes. Same workflow as upstream tutor-skills document mode — see document-workflow.md.
Phase Summary (D1-D9)
| Phase |
Name |
| D1 |
Source Discovery & Extraction (use pdftotext CLI, NEVER Read directly on PDF; for .ipynb use jupyter nbconvert --to markdown) |
| D2 |
Content Analysis |
| D3 |
Tag Standard |
| D4 |
Vault Structure |
| D5 |
Dashboard |
| D6 |
Concept Notes |
| D7 |
Practice Questions (active recall fold callouts) |
| D8 |
Interlinking |
| D9 |
Self-Review |
See document-workflow.md for detail.
Special handling for Triton documents: when processing the official Triton docs, the original Triton paper (Tillet et al., MAPL 2019), FlashAttention papers (Dao et al.), or PyTorch 2.x Inductor design docs, prefer reorganizing extracted content into the 6-topic structure (same folders as Curriculum mode) rather than mirroring the source's TOC. The 6-topic structure is the canonical taxonomy.
Language
- Match the user's language (Korean → Korean prose, etc.).
- Tags, keywords, and Triton identifiers stay verbatim in English regardless:
@triton.jit, triton.autotune, tl.load, tl.dot, tl.constexpr, BLOCK_SIZE_M, num_warps, num_stages, TTIR, TTGIR, #blocked, #mma, #shared, #dot_op, torch.compile, Inductor, proton, TRITON_INTERPRET, etc.
- Fold-callout labels can be localized (e.g., Korean:
정답 보기, English: View Answer).
Pairing with triton-tutor and tutor-handouts
The generated StudyVault is the input to two paired skills:
triton-tutor — interactive quiz tutor over the vault (4-question rounds, concept-level proficiency tracking).tutor-handouts — produces a Coursepack/ of CMU 15-418 / Stanford CS149-style PDF lecture handouts, programming assignment writeups, cheatsheets, problem sets, capstone project, syllabus, plus matching LeetGPU-style graded exercise scaffolds (CUDA / Triton / PyTorch starters + autograder harness).
Once Phase TU9 / C9 / D9 self-review passes, tell the user:
"Vault built at ./StudyVault/. Run the triton-tutor skill in the same directory to start quiz sessions, or tutor-handouts to generate a printable course pack with programming exercises."
1---2name: triton-tutor-setup3description: Generate an Obsidian-format Triton StudyVault for learning OpenAI Triton's GPU programming stack. Three modes: (1) Curriculum mode (default, no source files needed) builds a prereq-ordered vault around 6 topics: Triton basics, tiling & autotuning, matmul patterns (tiled GEMM, split-K, FP8 `tl.dot`), attention & reductions (FlashAttention), compiler internals (TTIR/TTGIR/LLIR, NVIDIA/AMD backends, WGMMA + TMA lowering), and ecosystem/production (`torch.compile`/Inductor, AOT, `proton`); (2) Codebase mode generates an onboarding vault from a real Triton/Inductor source tree; (3) Document mode turns Triton docs, blog posts, and tutorial PDFs into study notes. Mode is auto-detected from CWD markers. Use when the user wants to learn Triton from scratch, onboard onto a Triton codebase, build a Triton exercise list, or preprocess Triton PDFs/docs into study material. Pair with the `triton-tutor` skill to quiz against the resulting vault.4---56# Triton Tutor Setup — Build an Obsidian Triton StudyVault78## CWD Boundary Rule (ALL MODES)910> **NEVER access files outside the current working directory (CWD).**11> All source scanning, reading, and vault output MUST stay within CWD and its subdirectories.12> If the user provides an external repo path, ask them to either clone or symlink it into CWD first.1314## Mode Detection1516On invocation, auto-detect the mode:17181. **Scan CWD** for code markers: `setup.py`, `pyproject.toml`, `CMakeLists.txt`, `*.py` containing `import triton` / `@triton.jit` / `triton.language as tl`, `triton/` source dir, `python/triton/`, `lib/Dialect/Triton`, `third_party/nvidia`, `third_party/amd`.192. **Scan CWD** for document markers: `*.pdf`, `*.txt`, `*.md` (excluding `README.md`), `*.html`, `*.epub`, `*.ipynb` (tutorials).203. **Decide:**21 - **Triton codebase markers found** (`@triton.jit`/`triton.language`/`lib/Dialect/Triton`) → **Codebase Mode**.22 - **No code markers AND document markers found** → **Document Mode**.23 - **Neither found** (empty dir or only `.git/`) → **Curriculum Mode** (the default for this skill).244. **Announce** detected mode and ask the user to confirm or override (they may force any mode).2526> **Default bias:** when in doubt between Document mode and Curriculum mode, choose **Curriculum Mode** — it is this skill's headline use case.2728---2930## Curriculum Mode (default)3132> Generates a complete Triton learning StudyVault organized around the **6-topic curriculum**.33> Full workflow: see [curriculum-workflow.md](references/curriculum-workflow.md).34> Per-topic content lives in `references/topic-*.md` (loaded only for the topic being authored).35> Templates: [templates.md](references/templates.md).3637### Topic List & Order (fixed)3839The 6 topics, ordered by prerequisite chain. Build the StudyVault in this order. Do not reorder.4041| # | Topic | Reference File | Vault Folder |42|---|-------|----------------|--------------|43| 1 | Triton Basics (foundation) | [topic-triton-basics.md](references/topic-triton-basics.md) | `01-Triton-Basics/` |44| 2 | Tiling & Autotuning | [topic-tiling-autotuning.md](references/topic-tiling-autotuning.md) | `02-Tiling-Autotuning/` |45| 3 | Matmul Patterns | [topic-matmul-patterns.md](references/topic-matmul-patterns.md) | `03-Matmul-Patterns/` |46| 4 | Attention & Reductions | [topic-attention-reductions.md](references/topic-attention-reductions.md) | `04-Attention-Reductions/` |47| 5 | Compiler Internals | [topic-compiler-internals.md](references/topic-compiler-internals.md) | `05-Compiler-Internals/` |48| 6 | Ecosystem & Production | [topic-ecosystem-production.md](references/topic-ecosystem-production.md) | `06-Ecosystem-Production/` |4950Prerequisite chain (encode in the Dashboard's prereq DAG):5152```53 ┌────────────────────────────┐54 │ 1. Triton Basics │55 │ (@triton.jit, masks, │56 │ tl.load/store) │57 └──────┬──────────────────────┘58 │59 ┌──────┴──────┐60 ▼ ▼61 ┌──────────┐ ┌──────────────┐62 │ 2. Tiling│ │ 3. Matmul │63 │ & Auto- │◀▶│ Patterns │64 │ tuning │ │ (tl.dot) │65 └─────┬────┘ └───────┬──────┘66 │ │67 └───────┬───────┘68 ▼69 ┌─────────────────────┐70 │ 4. Attention & │71 │ Reductions │72 │ (FlashAttention) │73 └────────┬────────────┘74 │75 ┌──────┴──────┐76 ▼ ▼77 ┌────────────┐ ┌──────────────┐78 │ 5. Compiler│ │ 6. Ecosystem │79 │ Internals│ │ & Production│80 │ (TTIR/TTGIR│ │ (Inductor, │81 │ /PTX) │ │ proton) │82 └────────────┘ └──────────────┘83```8485### Phase Summary (TU1-TU9)8687| Phase | Name | Key Action |88|-------|------|------------|89| TU1 | Pre-flight | Ask user about target hardware (NVIDIA A100/H100/B200, AMD MI250/MI300, Intel PVC), prior knowledge (PyTorch user / CUDA writer / Triton kernel author), language preference. Skip topics already mastered if user opts in. |90| TU2 | Topic Plan | Build the topic dependency DAG. Materialize the prereq chain in the dashboard. |91| TU3 | Tag Standard | Define tag registry: `#topic-triton-basics`, `#topic-tiling-autotuning`, `#topic-matmul-patterns`, `#topic-attention-reductions`, `#topic-compiler-internals`, `#topic-ecosystem-production`, plus `#concept-*`, `#milestone-*`, `#pitfall-*`. English kebab-case. |92| TU4 | Vault Structure | Create `StudyVault/` with `00-Dashboard/` + the 6 numbered topic folders + `99-Exercises/`. |93| TU5 | Dashboard | MOC + Prereq DAG + Glossary + Quick Reference + Pitfall Index. |94| TU6 | Per-Topic Notes | For each of the 6 topics: load `references/topic-{slug}.md` (lazy, one at a time to manage context), then generate concept notes per [templates.md](references/templates.md). For topics with a CUDA/CUTLASS/cuTile peer (Triton Basics, Tiling/Autotuning, Matmul Patterns, Attention/Reductions, Compiler Internals), the topic reference file ends with a "Cross-Stack Equivalent" appendix table — surface this in the generated vault notes so vault readers see the Rosetta mapping while studying the Triton-side concept. The full Rosetta is at `../tutor-core/references/cross-stack-rosetta.md` (consumed by `triton-tutor` at quiz time via symlink). |95| TU7 | Hands-On Milestones | Per topic, materialize 3-7 "build/run/measure X" milestones as standalone notes with success criteria. |96| TU8 | Pitfall Notes | Per topic, materialize the pitfalls section as `Pitfalls.md` with fold callouts. |97| TU9 | Interlinking + Self-Review | Cross-link all notes per the prereq DAG, then verify against [quality-checklist.md](references/quality-checklist.md) **Curriculum Mode** section. |9899See [curriculum-workflow.md](references/curriculum-workflow.md) for the full procedural detail of each phase.100101### Lazy-Loading Discipline for Topic Content (IMPORTANT)102103The 6 topic reference files (`topic-*.md`) are large. Load them **one at a time** in Phase TU6 — read `topic-triton-basics.md` while authoring `01-Triton-Basics/`, drop it from working memory before reading `topic-tiling-autotuning.md`, etc. Do not pre-load all 6.104105---106107## Codebase Mode108109> Generate an onboarding StudyVault from a real Triton / Inductor / kernel-library codebase.110> Full workflow: [codebase-workflow.md](references/codebase-workflow.md). Templates: [triton-codebase-templates.md](references/triton-codebase-templates.md).111112### Phase Summary (C1-C9)113114| Phase | Name | Key Action |115|-------|------|------------|116| C1 | Project Exploration | Detect Triton version (parse `setup.py`/`pyproject.toml`/`requirements.txt`), target backends (NVIDIA / AMD / Intel via `third_party/`), identify role of the repo (kernel library / compiler fork / inference engine / training framework). |117| C2 | Architecture Analysis | Trace a representative kernel from Python entry (`kernel[grid](...)`) through `@triton.jit` compile cache, autotune flow, generated IR (dump with `TRITON_KERNEL_DUMP=1`), into PTX/AMDGCN. Identify host-side wrappers and PyTorch integration points. |118| C3 | Tag Standard | Same kebab-case rules; add `#backend-*`, `#arch-*`, `#kernel-role-*`, `#api-*`, `#ir-stage-*` categories. |119| C4 | Vault Structure | Dashboard, Architecture, per-module, DevOps (build flags, CI, benchmarks), Exercises. |120| C5 | Dashboard | MOC + Module Map + Kernel Surface table + Backend Matrix + Getting Started (build & run a representative benchmark). |121| C6 | Module Notes | Per-module note: Purpose, Key Kernels, Python Wrappers, Autotune Configs, Block Sizes, IR Lowering Notes, Dependencies, How to Test. |122| C7 | Exercises | Code reading (trace a kernel + dump IR), config (change BLOCK_SIZE, swap backend), debugging (`TRITON_INTERPRET`, `tl.device_print`, IR dump), extension (add a new kernel or autotune config). |123| C8 | Interlinking | Cross-link Python wrapper ↔ `@triton.jit` kernel ↔ autotune config ↔ test. |124| C9 | Self-Review | [quality-checklist.md](references/quality-checklist.md) **Codebase Mode** section. |125126Common Triton-adjacent codebases this mode targets: `triton-lang/triton`, `pytorch/pytorch` (Inductor codegen), `Dao-AILab/flash-attention` (Triton variants), `facebookresearch/xformers`, `linkedin/Liger-Kernel`, `unslothai/unsloth`, `pytorch-labs/applied-ai`, `meta-pytorch/tritonbench`, `openai/triton` examples, plus user repos containing `@triton.jit` kernels.127128---129130## Document Mode131132> Transform PDFs (Triton papers, OpenAI/Meta blog posts in PDF, conference slides on Triton), text, and web pages into study notes. Same workflow as upstream tutor-skills document mode — see [document-workflow.md](references/document-workflow.md).133134### Phase Summary (D1-D9)135136| Phase | Name |137|-------|------|138| D1 | Source Discovery & Extraction (use `pdftotext` CLI, NEVER `Read` directly on PDF; for `.ipynb` use `jupyter nbconvert --to markdown`) |139| D2 | Content Analysis |140| D3 | Tag Standard |141| D4 | Vault Structure |142| D5 | Dashboard |143| D6 | Concept Notes |144| D7 | Practice Questions (active recall fold callouts) |145| D8 | Interlinking |146| D9 | Self-Review |147148See [document-workflow.md](references/document-workflow.md) for detail.149150**Special handling for Triton documents**: when processing the official Triton docs, the original Triton paper (Tillet et al., MAPL 2019), FlashAttention papers (Dao et al.), or PyTorch 2.x Inductor design docs, prefer reorganizing extracted content **into the 6-topic structure** (same folders as Curriculum mode) rather than mirroring the source's TOC. The 6-topic structure is the canonical taxonomy.151152---153154## Language155156- Match the user's language (Korean → Korean prose, etc.).157- **Tags, keywords, and Triton identifiers** stay verbatim in English regardless: `@triton.jit`, `triton.autotune`, `tl.load`, `tl.dot`, `tl.constexpr`, `BLOCK_SIZE_M`, `num_warps`, `num_stages`, `TTIR`, `TTGIR`, `#blocked`, `#mma`, `#shared`, `#dot_op`, `torch.compile`, `Inductor`, `proton`, `TRITON_INTERPRET`, etc.158- Fold-callout labels can be localized (e.g., Korean: `정답 보기`, English: `View Answer`).159160## Pairing with `triton-tutor` and `tutor-handouts`161162The generated StudyVault is the input to two paired skills:163164- **`triton-tutor`** — interactive quiz tutor over the vault (4-question rounds, concept-level proficiency tracking).165- **`tutor-handouts`** — produces a `Coursepack/` of CMU 15-418 / Stanford CS149-style PDF lecture handouts, programming assignment writeups, cheatsheets, problem sets, capstone project, syllabus, plus matching LeetGPU-style graded exercise scaffolds (CUDA / Triton / PyTorch starters + autograder harness).166167Once Phase TU9 / C9 / D9 self-review passes, tell the user:168169> "Vault built at `./StudyVault/`. Run the `triton-tutor` skill in the same directory to start quiz sessions, or `tutor-handouts` to generate a printable course pack with programming exercises."
drunkcoding