confidential-gemma4-31b

Tinfoil confidential enclave configuration for google/gemma-4-31B-it.

This branch ports the confidential-computing and MTP optimization stack to vLLM v0.25.1. The candidate has completed full-CC correctness, API, stress, and performance validation on an NVIDIA B300; HANDOFF.md records the immutable identities, results, and release procedure.

Repo layout

confidential-gemma4-31b/
├── Dockerfile                 builds on the digest-pinned vLLM v0.25.1 base
├── HANDOFF.md                 candidate status and required validation
├── patches/                   ordered deployment patch series
│   ├── 0101-0104-...          minimal CC, correctness, and upstream MTP fixes
│   ├── 0105-0119-...          V1 CC/MTP performance and safety fixes
│   └── README.md              provenance and regeneration instructions
├── validation/                API and source-level verification tools
├── tinfoil-config.yml         enclave configuration
└── .github/workflows/
    ├── tinfoil-release.yml    review image and release build
    └── tinfoil-release-publish.yml
                               release measurement and publication

The authoritative implementation is tinfoilsh/vllm-cc-opt:milestone/gemma4-cc-v0251-v1-perf-b300-20260722. Experiment evidence is in the private tinfoilsh/vllm-cc-gemma4-lab repo.