Docs

Quick start

The main branch contains runnable Rust code for G3.5 shared KV caching over CPU/TCP. Start with the software contract tests; no GPU or DPU is required.

In development

Source code is available. There is no installable release or production deployment qualification.

Build and test from source

Use stable Rust with Cargo, Git, and Python 3. Fetch dependencies on the first run. Test payloads are synthetic KV blocks, not output from an inference model.

git clone https://github.com/rustfs/GPUCache.git
cd GPUCache
cargo test --workspace --all-targets

Run the CPU demos

Run these demos without GPUCACHE_RUSTFS_* endpoint or credential settings. The cold demo uses the default mock; it is not a live RustFS validation.

cargo run -p gpucache-transport --bin share-demo
cargo run -p gpucache-transport --bin shard-demo
cargo run -p gpucache-cold --bin cold-demo
cargo run -p gpucache-storage --bin engine-spike
cargo run -p gpucache-dpu --bin dpu-demo

The expected labels are REUSE_OK, SHARD_OK, COLD_CONTRACT_OK, FAKE_BLOCK_CONTRACT_OK and DPU_CONTRACT_OK. The engine spike checks fake whole blocks. The DPU demo is a host simulation. Neither label means a real engine or device ran.

Reproduce the full CPU regression

The runner checks formatting, dependencies, Clippy, tests, optional features and demos. It disables live RustFS access and saves summary.md, results.json, JUnit and logs. Use a new output directory for each run; --offline requires dependencies to have been fetched.

cargo fetch --locked
python3 scripts/ci/vm_regression.py --profile cpu --offline --output artifacts/cpu-docs

What this does not qualify

Real vLLM export/import, RDMA, GDS and BlueField-3 remain NOT_RUN_HARDWARE; the GPU/model campaign is NOT_RUN_BUDGET. No active zero-copy path, hardware performance result or hot-tier durability replica is implied.

Quick start

Build and test from source

Run the CPU demos

Reproduce the full CPU regression

What this does not qualify

Related documentation