Skip to content

github.md

Latest commit

 

History

History
107 lines (79 loc) · 2.48 KB

github.md

File metadata and controls

107 lines (79 loc) · 2.48 KB
Raw

Publish the repository on GitHub

1. Extract and enter the repository

unzip llm-serving-kernels.zip
cd llm-serving-kernels

2. Create the main environment

uv venv --python 3.12
source .venv/bin/activate
uv pip install --torch-backend=auto -e '.[dev]'

3. Run local release gates

make check
make test

On the NVIDIA L4, run the complete hardware gate:

python tools/environment.py | tee results/environment.json
make gpu-test
make benchmark
make cutlass

Project 08 uses a separate environment:

cd projects/08_vllm_plugin
uv venv --python 3.12 --seed .venv
source .venv/bin/activate
uv pip install --torch-backend=auto -e . pytest
VLLM_PLUGINS=llm_serving_swiglu pytest -q test_operator.py
VLLM_PLUGINS=llm_serving_swiglu \
  python benchmark_operator.py > ../../results/08_vllm_operator.json
cd ../..

Do not commit benchmark output until the corresponding correctness suite passes in the same environment.

4. Inspect the files Git will receive

git status --short
git diff --check --no-index /dev/null README.md || true
find . -type d \( -name .venv -o -name __pycache__ -o -name build \) -prune

The final command should print only directories intentionally excluded by .gitignore.

5. Create the Git history

git init -b main
git add .
git status --short
git commit -m "Initial release: LLM serving kernels"

6. Create and push the GitHub repository

With GitHub CLI:

gh auth login
gh repo create llm-serving-kernels \
  --public \
  --description "Eight GPU-kernel projects for modern LLM serving" \
  --source . \
  --remote origin \
  --push

Without GitHub CLI, create an empty GitHub repository named llm-serving-kernels, then run:

git remote add origin https://github.com/YOUR_USERNAME/llm-serving-kernels.git
git push -u origin main

7. Configure the repository

Use these topics:

cuda triton cutlass cute-dsl vllm llm-inference gpu-kernels fp8 paged-attention moe

Keep GitHub Actions enabled. Add a repository description and pin the repository to your profile only after verified L4 evidence is visible.

8. Commit verified L4 evidence

git add results
git commit -m "Add verified NVIDIA L4 results"
git push

A strong public release includes correctness results, benchmark JSON, one Nsight Systems timeline, one Nsight Compute report, and one short PTX/SASS note explaining an observed instruction or memory behavior without claiming unsupported speedups.