v0.1.0a2 — Cleaner configuration and safer delivery
August 30, 2026
This alpha focuses on making experiments easier to describe, reports easier to recover, and contributions safer to review.
Highlights
- Fixed vLLM arguments now live directly under
server. - Tunable flags, fixed environment variables, and tunable environment variables have clear top-level sections.
- Reports can be regenerated offline from validated run artifacts.
- Pull requests validate Python 3.11–3.12 packages and documentation before merge.
Configuration
- Made
server.modelthe required local model path. - Removed the extra
server.argsnesting. - Moved tunable vLLM flags to
tune, fixed environment variables toenv, and tunable environment variables totune_env. - Documented categorical, boolean, numeric-range, list, and environment forms.
Reporting
- Added offline HTML report regeneration without starting vLLM or GuideLLM.
- Preserved immutable source artifacts and refused accidental destination overwrites.
Delivery and documentation
- Added package checks for Python 3.11 and 3.12 on pull requests and
main. - Added strict documentation builds on pull requests and Pages deployment from
main. - Pinned GitHub Actions to immutable Node 24 revisions.
- Added code ownership, CUDA guidance, and a shorter quick start.
Validation
- All 113 private tests and 17 subtests passed.
- Strict documentation and Python compilation checks passed.
- The universal wheel and source archive passed metadata checks.
- Python 3.11 and 3.12 GitHub package checks passed before release preparation.
- The latest real compatibility run completed with vLLM
0.28.0and GuideLLM0.7.3on WSL2 with an RTX 3080. This was a compatibility workload, not a general performance claim.
Install
The universal py3-none-any wheel installs on Linux, Windows, and macOS.
Running vLLM experiments remains Linux-only.