Skip to content

Getting Started

Jump to bottom
Wiki Keeper (Grok Bot) edited this page Oct 15, 2026 · 2 revisions

Getting Started

Quickstart for the CUDA lab on ShipOfTheseus (RTX 5080, sm_120, ~16 GB).

Prerequisites

  • Linux host with NVIDIA GPU, CUDA toolkit 13.3 at /usr/local/cuda, nvcc on PATH
  • Host GCC 6.x–15.x (CUDA 13.3 rejects gcc 16+; CMake auto-picks g++-15 when present)
  • For L1 engine recipes: the engine under test (e.g. Ollama + local model) already available

1. Sanity-check the host

nvidia-smi            # GPU 0 must show ≥2048 MiB free before train or GPU CI
nvcc --version        # expect CUDA 13.3

Serialize rule: training and GPU CI share this one card. Do not start both.

2. Build the kernel workspace

cmake -S kernels -B build/kernels -DBKL_ENABLE_CUDA=ON
cmake --build build/kernels -j"$(nproc)"

CPU-only configure (CPU CI):

cmake -S kernels -B build/kernels -DBKL_ENABLE_CUDA=OFF

3. Smoke and benchmarks

./build/kernels/src/bkl_device_hello
# expect: device0: NVIDIA GeForce RTX 5080 compute 12.0
#         bkl device_hello sm_120 ok

./build/kernels/src/bkl_graph_launch_bench --out results/graph-launch-bench.json
./build/kernels/src/bkl_green_ctx_bench --out results/green-ctx-bench.json

Outputs under results/ (gitignored).

4. CPU policy / F0 tools (no GPU required)

python3 tools/test_gpu_ci_policy.py
python3 tools/test_publish_gpu_validation.py
# F0 validators — see docs/f0-*.md and fixtures/

5. Recipes

  • recipes/l1-prefix-kv-reuse.md
  • recipes/l2-green-ctx-bench.md
  • recipes/l3-device-hello.md
  • recipes/l3-graph-launch-bench.md
  • recipes/kernel-ablation.md
  • recipes/compute-sanitizer.md
  • recipes/f0-nvml-capability.md

Next steps

  • Architecture — L1/L2/L3, forge boundary, F0, CI
  • Glossary — terminology
  • Repo docs: docs/KERNELS.md, docs/RELEASE_V0_2_0.md, docs/FORGE_CONSUME.md

Refreshed 2026-10-15 by Wiki Keeper (Grok Bot) from code at main @ 22cc4f3 (tagged v0.1.0; release target v0.2.0).

Clone this wiki locally