Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
62 lines (45 loc) · 2.9 KB

CONTRIBUTING.md

File metadata and controls

62 lines (45 loc) · 2.9 KB
Raw

Contributing to AIStudio Workloads

Thanks for your interest in contributing. This document covers how to get started, what areas are open for extension, and what to verify before opening a pull request.

Getting Started

  1. Fork the repo and create a branch from main.
  2. Make your changes inside the relevant workload directory (llm-inference/ or jupyter-notebook/).
  3. Build and test your image locally before opening a PR.
  4. Open a pull request against main with a clear description of what and why.

For new workload types or significant changes, open an issue first to discuss the approach.

Areas Open for Extension

Area Examples
New workload types Training jobs, embedding benchmarks, diffusion model inference, ASR
AMD ROCm support ROCm-compatible Dockerfiles alongside the existing NVIDIA images
Benchmark metrics Extend benchmark.py with additional vLLM metrics or cost models
Dataset formats Support for dataset formats beyond ShareGPT (e.g. OASST, Dolly)
Jupyter notebooks New GPU profiling or MLPerf benchmark notebooks
Hardware pricing Extend hardware_pricing.json with additional GPU types and cloud regions

Best Practices

  • One workload per directory. Each workload is self-contained with its own Dockerfile, requirements.txt, and entrypoint.
  • No bundled datasets. Workloads must not download or embed datasets automatically. Use --dataset-path and require the operator to supply the file.
  • Permissive licences only. New dependencies must be Apache-2.0, MIT, or BSD. No GPL, AGPL, or packages with contested provenance without prior discussion. Update the NOTICE file if you add a package with notice obligations.
  • Keep images lean. Do not add packages to requirements.txt that are already in the base image (vllm/vllm-openai for llm-inference).

Before Raising a PR

# 1. Build the image locally
cd llm-inference   # or jupyter-notebook
docker build -t llminference:dev .

# 2. Run a smoke test (requires a GPU node with Docker)
docker run --gpus all --rm \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -v /tmp/results:/results \
  -v /path/to/dataset.json:/path/to/dataset.json \
  --entrypoint python3 llminference:dev \
  /llm-inference/benchmark.py \
    --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 \
    --dataset-path /path/to/dataset.json \
    --concurrencies 1 \
    --output_data_path /results/smoke

# 3. Confirm BENCH_RESULT lines are printed and results written to /tmp/results

CI runs a lint and image build check on every PR. PRs that fail to build will not be merged.

Reporting Bugs

Open a GitHub issue with: steps to reproduce, expected vs actual behaviour, relevant log output, GPU type, and image tag used.

Security Issues

Do not open a public issue for security vulnerabilities. Email muraharirao@corespan.ai directly.