Collective Knowledge (CK) in a community project to develop open-source tools, platforms and automation recipes that can help researchers and engineers automate their repetitive, tedious and time-consuming tasks to build, run, benchmark and optimize AI, ML and other applications and systems across diverse and continuously changing models, data, software and hardware.

CK consists of several ongoing sub-projects:

• Collective Mind framework (CM) (~1MB) - a very light-weight Python-based framework with minimal dependencies to help users implement, share and reuse cross-platform automation recipes to build, benchmark and optimize applications on any platform with any software and hardware. CM attempts to extends the cmake concept with reusable automation recipes and workflows written in plain Python or native OS scripts, accessible via a human readable interface with simple tags, and shareable in public and private repositories in a decentralized way. Furthermore, in comparison with cmake, these automation recipes can not only detect missing code but also download artifacts (models, data sets), preprocess them, build missing dependencies, install them and run the final code on diverse platforms in a unified and automated way. You can read more about the CM concept in this presentation.

  • CM automation recipes for MLOps and DevOps (~6MB) - a small collection of portable, extensible and technology-agnostic automation recipes with a human-friendly interface (aka CM scripts) to unify and automate all the manual steps required to compose, run, benchmark and optimize complex ML/AI applications on diverse platforms with any software and hardware: see online catalog and source code.

  • CM automation recipes to reproduce research projects (~1MB) - a unified CM interface to help researchers and engineers access, prepare and run diverse research projects and make it easier to validate them in the real world across rapidly evolving models, data, software and hardware (see our reproducibility initatives and motivation behind this project).

  • CM automation recipes for ABTF (~1MB) - a unified CM interface and automation recipes to run automotive benchmark across different models, data sets, software and hardware from different vendors.

  • Modular C++ harness for MLPerf loadgen

  • Modular Python harness for MLPerf loadgen

• Collective Knowledge Playground - an external platform being developed by cKnowledge to list CM scripts similar to PYPI, aggregate AI/ML Systems benchmarking results in a reproducible format with CM workflows, and organize public optimization challenges and reproducibility initiatives to find the most performance and cost-effective AI/ML Systems.

  • CK GUI to run modular benchmarks - such benchmarks are composed from CM scripts and can run via a unified CM interface.

Apache 2.0

MLCommons is updating the CM documentation based on user feedback - please stay tuned for more details.

• CM Getting Started Guide and FAQ
  • Common CM interface to run MLPerf inference benchmarks
  • Common CM interface to re-run experiments from ML and Systems papers including MICRO'23 and the Student Cluster Competition @ SuperComputing'23
  • CM automation recipes for MLOps and DevOps
  • Other CM tutorials
• Full documentation
• CM development tasks
• CM and CK history

Please use this BibTex file.

Collective Knowledge automation framework (deprecated CK v1 and v2), Collective Mind automation framework (CM), CM4MLOPS and CM4ABTF were originally developed by Grigori Fursin and donated to MLCommons to benefit everyone. You can learn more about the motivation behind these projects from the following presentations:

• ACM REP'23 keynote about the MLCommons CM automation framework: [ slides ]
• ACM TechTalk'21 about automating research projects: [ YouTube ] [ slides ]

We would like to thank all our great volunteers, collaborators and contributors for their support, fruitful discussions, and useful feedback! We thank the cTuning foundation, cKnowledge.org and MLCommons for sponsoring this project!