This directory holds TensorRT-LLM's third-party C++ dependencies (driven by cmake FetchContent from fetch_content.json) plus tooling that accelerates repeat clones of those dependencies.

The markdown files in this directory contain playbooks for how to add new third-party dependencies. Please see the document that matches the kind of dependency you want to add:

• For C++ dependencies compiled into the extension modules via the cmake build and re-distributed with the wheel, see cpp-thirdparty.md
• For python dependencies declared via wheel metadata and installed in the container via pip, see py-thirdparty.md

scripts/build_wheel.py --use-3rdparty-cache enables a local bare-repo cache that accelerates cmake FetchContent clones via git clone --reference. It is opt-in; without the flag, no -D is added and the cache code path in 3rdparty/CMakeLists.txt is bypassed entirely.

For flow, cache layout, threat model, and design rationale, see fetch-cache.md.