메뉴
How to author a rigorous report/LAB.md for a curriculum lab — the writeup that is graded on the Writeup axis of the rubric and shipped publicly with the repo. Use when running /lab-report or finishing a lab.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Working with NVIDIA Cosmos World Foundation Models (Cosmos-Predict, Cosmos-Reason, Cosmos-Transfer) on DGX Spark — pulling containers, running baseline inference, fine-tuning on custom data, and exporting for production. Use when the user is on Lab 11+ or otherwise interacting with Cosmos.
Step-by-step workflow for writing a new CUDA kernel from spec to profile-validated implementation on DGX Spark (sm_121). Use when starting a new kernel, refactoring an existing one, or porting Python/PyTorch logic to CUDA.
The master 4-month cuda-spatial-intelligence-lab curriculum. Use when the user asks about the overall plan, what week they are on, what to study next, or for the rubric used to grade weekly checkpoints. Read the relevant month file for week-level detail.
Bootstrap a fresh NVIDIA DGX Spark for this curriculum — verify hardware, log into NGC, pull baseline ARM64 containers, install CUDA-13/CMake/Ninja/Nsight host tooling, and prove the toolchain works with the standard CMake template. Use when starting work on a new Spark.
Build the Month-4 NextJS application backed by a LangChain DeepAgent that consumes home cameras/sensors and routes inference to Spark (Triton) or AWS (SageMaker/Bedrock). Use when scaffolding the app, defining tools, wiring SSE streaming, or persisting agent memory.
Canonical pattern for wrapping our CUDA kernels as PyTorch custom ops via torch.utils.cpp_extension. Use proactively when a lab from Lab 02 onward needs a python/ wrapper, or when a production kernel needs to survive torch.compile / TensorRT export.
SOC 직업 분류 기준
| name | lab-notebook |
| description | How to author a rigorous report/LAB.md for a curriculum lab — the writeup that is graded on the Writeup axis of the rubric and shipped publicly with the repo. Use when running /lab-report or finishing a lab. |
report/LAB.md) authoringEvery labs/lab-NN-*/ ships a report/LAB.md. It is the write-up
recruiters will read. Treat it as a mini-paper.
# Lab N — <Title>
## 1. Spec
- Inputs / outputs (shapes, dtypes, layouts)
- Numerical tolerance
- Performance target (cite the curriculum file)
- Constraints
## 2. Hypothesis
*Written before any optimization.* What do you expect the bottleneck to
be? Why? Which Nsight metric will move first?
## 3. Method
For each version (v0 naive → vN optimized), one paragraph:
- What changed.
- Why (cite PMPP §, Iglberger guideline, paper).
- One sentence on what you expected.
Include any non-obvious build flags.
## 4. Results
A table. Always. With units.
| Version | Time (ms) | GFLOP/s | % of <baseline> | Achieved BW (GB/s) |
|---|---|---|---|---|
| v0 naive | ... | ... | ... | ... |
| v1 tiled | ... | ... | ... | ... |
| ... | ... | ... | ... | ... |
Reference Nsight reports inline:
> Roofline analysis: see `report/ncu_<name>_v3.ncu-rep` — kernel sits at
> 73% of FP32 ceiling, memory-bound on L2.
## 5. Discussion
- Where did the predicted bottleneck NOT show up? Why?
- What surprised you?
- What did you learn that you'll carry forward?
## 6. What I would do next
At least one concrete next experiment (with the metric you'd watch).
"Try TMA + thread-block clusters and see if SOL Memory drops below 60%."
## 7. References
- PMPP 4e Ch X §Y.Z
- Paper title (Author et al, year, arXiv:XXXX.XXXXX)
- Curriculum file: `.cursor/skills/curriculum-plan/month-N-*.md`
## Setup only.# Lab 02 — Tiled SGEMM
## Setup
- Hardware: DGX Spark, sm_121, driver <X>, CUDA <Y>
- Toolchain: GCC 13, CMake 3.28, Ninja
- Reference: cuBLAS sgemm
- Workload: M=N=K=4096 single precision, 100 trials
## 1. Spec
...
## 2. Hypothesis
...
## 3. Method
### v0 — naive
### v1 — 32×32 tiled
### v2 — 64×64 tiled, padded
### v3 — async-copy double-buffered
## 4. Results
| Version | Time (ms) | GFLOP/s | % cuBLAS |
|---|---|---|---|
| v0 | ... | ... | ... |
| v1 | ... | ... | ... |
| v2 | ... | ... | ... |
| v3 | ... | ... | ... |
## 5. Discussion
...
## 6. What I would do next
...
## 7. References
- PMPP 4e §5.4 (tiling)
- PMPP 4e §6 (performance)
- CUTLASS docs, "Efficient GEMM in CUDA"
- `.cursor/skills/curriculum-plan/month-1-foundations.md`