Pull requests / #3396

#3396 [Engine] Add opt-in Galahad session storage

open · @j-luwierski · 0 Kommentare · Auf GitHub

Setup & installServer & APIAMD / HIPNVIDIA / CUDASecurityLinux

Beschreibung

## Problem and result

Strata's existing slot API can persist its current conversation to a local file. Users who need encrypted, durable named sessions on supported NVIDIA systems have no integrated storage backend. With this opt-in change, the same API saves and restores complete Strata session images through Galahad; the existing file backend and default inference path stay unchanged.

## Related discussion

Previous closed submission: https://github.com/Niko1221/Strata/pull/1761. The maintainer asked for a rebase on current main, a title prefix and the current PR template. After updating the old submission, the `/reopen` check passed but GitHub refused to reopen it and instructed us to open a new pull request. This PR contains that rebased update.

## Approach

Add a separate Galahad adapter for full session images, configuration and CLI options, size/error handling, and user documentation. Galahad remains an optional import and requires a separately installed licensed package. Session operations remain explicit, limited to slot 0, and unavailable with parallel requests or the ordinary file backend. The rebase preserves the current session-save retry behavior. Galahad shutdown is guarded against duplicate close calls from normal shutdown and `atexit`.

## Validation

**Passed:**

- `.venv/bin/python -m unittest serve.test_galahad serve.test_slots serve.test_security` — 67 tests passed on Linux x86-64, Python 3.13, at revision `7e3494582faecba5934c1d43b692b98ecfe080e1`, based on main `5bda973a`.
- `.venv/bin/python serve/server.py --help` — Galahad options are present.
- `git diff origin/main --check` — no whitespace errors.

**Failed:**

- None observed.

**Not run:**

- Licensed native Galahad save/restart/restore, GPU output parity, and performance checks: this environment has no licensed NVIDIA setup.
- CUDA, HIP, and SYCL builds: no engine or backend kernel files changed.

## Limits

Requires Linux x86-64, a licensed NVIDIA GPU, and `galahad-kv==1.31.5`. Names are immutable, only slot 0 is supported, and the feature does not add automatic prefix caching. Mock tests establish the adapter and HTTP contract, not native persistence, encryption, GPU correctness, or performance.

Mehr auf der Site

Links zu Install, Modellen, Releases.