Pull requests / #3244

#3244 [Docs] the three things 0.1.42 changed that no page told

open · @Buddychek · 0 コメント · GitHub で見る

Setup & installServer & APIAMD / HIPNVIDIA / CUDAModels & quantsSecurityDocumentationWindows

本文

### Problem and result

Three things 0.1.42 changed that no page told:

1. `/metrics/prometheus` and the llama.cpp-compatible discovery endpoints are not listed as endpoints.
2. `STRATA_VCVARS=<path to vcvars64.bat>` (naming a Visual Studio install `vswhere` cannot use) is not documented.
3. The Monitor's prompt-speed row (prompt tokens minus `reused` over `prompt_ms`) is not described.

After: `docs/DETAILS.md`, `SECURITY.md` and `docs/TROUBLESHOOTING.md` state all three, matching the code. Docs-only.

### Related discussion

Closed in the 0.2.0-beta tracker reset; rebased on main and re-verified per the reset note. #985 (CUDA 13.0-13.2 = VS 2019/2022, VS 2026 only from CUDA 13.3) and #881 (CPU image encoder built by MSVC alone) are the compiler-acceptance reports the text cites.

### Approach

Docs text only, verified against the code on current main: `/metrics/prometheus` in `serve/server.py` (the route and the Prometheus text format), `STRATA_VCVARS` in `setup.py` (the override), `reused`/`prompt_ms` in `serve/server.py` (the prompt-speed row).

### Validation

Passed (static verification, rebased on 0.2.0-beta main `e88ee39d`): every statement re-checked against `serve/server.py` and `setup.py` on current main — endpoint, env var, and the reused/prompt_ms fields all exist as documented.
Not run: no code change, nothing to run.

### Limits

Docs only; describes shipped 0.1.42 behavior.

関連リンク

インストール・モデル・リリースへの站内リンク。