Changelog

For the complete and up-to-date version history of the Lutflow SDK, please refer to the main repository's changelog:

TPOT for vLLM v0.30.0 uses vllm:inter_token_latency_seconds when the older family is absent: a per-token-gap mean labeled derived.
The p95 interpolation warning identifies its bucket; the value and schema are unchanged.
TPOT continuity across families avoids false resets; a family switch keeps the window with TPOT unavailable.
Licensor and package author are Lutflow EAS; the other BSL 1.1 terms are unchanged.
Shorter, self-contained published README; the sdist no longer includes tests or fixtures.
Wrapper prices remain dated: separate input/output rates and gpt-4o-mini fallback for unknown models. This is not a price-table fix.
Source: 0.3.2 CHANGELOG.

Adds lutflow check to inspect one exposition against the vLLM v0.30.0 measurable profile. It generates no inference traffic and is not a measurement.
Exits 0/1/3/2 and a tier0.check.v1 document; see the SDK.
Does not yet include TPOT under vLLM's new names, the p95 bucket-bound warning or the wrappers' price update. Earlier limitations remain. The versioned CHANGELOG is the reference.

Behavior change: a window in which the engine restarts, a token counter disappears from some scrapes, or the engine's process identity changes now ends with exit code 1 and no report. 0.2.1 reported some of those windows with exit 0. Detecting a restart whose counters keep rising requires the engine to publish its process identity. tier0.measure.v1 is unchanged.
Python 3.14 and 3.15 are supported; the package requires Python 3.9 or newer.
Install instructions offer uvx, uv tool install and pipx first, all pinned to >=0.3.0,<0.4.
Known limitations: restart detection depends on the engine publishing its identity, TPOT is unavailable on recent vLLM, the p95 resolves only to a histogram bucket, and the wrappers' cost estimates use a dated price table. See SDK: Known limitations.
Releases 0.2.0 and 0.2.1 were withdrawn (yanked) from PyPI.