LMCache InProcess

lmcache_inprocess packages LMCache's legacy in-process engine separately from the main LMCache distribution. It keeps the existing Python import path, lmcache.inprocess, so current vLLM, TensorRT-LLM, configuration, and plugin integrations do not need an API rename.

The wheel also supplies lmcache.integration.vllm.vllm_v1_adapter as a compatibility re-export because vLLM's connector imports that canonical path. The implementation remains under lmcache.inprocess.integration.vllm.

The initial source snapshot is copied from LMCache dev commit 95ced2b137e3d854de662a45dd3c6ca3fee0753b. The exact source revision is also recorded in UPSTREAM_REVISION.

Install

Install a compatible LMCache build first, then this package:

python -m pip install 'lmcache>=0.5.7rc3'
python -m pip install lmcache_inprocess

For development against LMCache dev:

NO_GPU_EXT=1 python -m pip install \
  'lmcache @ git+https://github.com/LMCache/LMCache.git@dev'
python -m pip install -e . --no-deps

The distribution is pure Python. Native extensions remain owned and shipped by the lmcache dependency. A CLI-only NO_NATIVE_EXT=1 LMCache installation is not sufficient because the in-process engine requires LMCache's common native operations; CPU validation uses NO_GPU_EXT=1.

Compatibility contract

• Distribution name: lmcache_inprocess
• Import namespace: lmcache.inprocess
• Base dependency: lmcache>=0.5.7rc3
• Python: 3.10 through 3.14
• Supported source of shared native extensions: the installed lmcache package

Because both distributions contribute files below the top-level lmcache directory, install and uninstall them with the same environment manager. A future LMCache release should stop shipping lmcache/inprocess before both packages are published as stable releases, avoiding overlapping wheel ownership.

Development and validation

python -m pip install -r requirements/test.txt
python -m build
python -m twine check dist/*
python -m pytest -q tests/smoke
python -m pytest -q tests/inprocess/v1/test_mock_gpu_connector.py

CI builds and installs the wheel in a clean environment, runs CPU smoke tests, and tests compatibility against a pinned LMCache revision and current LMCache dev. GPU end-to-end validation is available through the manual workflow in .github/workflows/gpu-e2e.yml on a repository runner labeled self-hosted, linux, x64, gpu, and h200.

Repository layout

• lmcache/inprocess/: package source
• tests/inprocess/: copied unit and integration tests
• examples/inprocess/: in-process examples
• benchmarks/inprocess/: in-process benchmarks
• scripts/sync_from_lmcache.sh: reproducible upstream refresh helper

Release

Create a GitHub release with a vX.Y.Z tag. The release workflow builds the source distribution and universal wheel, validates both, uploads them as GitHub release artifacts, and publishes to PyPI through Trusted Publishing when the pypi environment is configured for this repository.