adlee-was-taken 519fa4aee7 fix(config): meter an in-process encoder regardless of classifier.base_url ...

local_energy_call_sites["classify"] gated on whether
classifier.base_url resolves to loopback. That is the right question in
local_llm mode, where base_url IS the machine doing the work.

It is meaningless in local_encoder mode. The encoder places no HTTP call
at all -- the model runs in-process on this host's own GPU -- so
base_url describes nothing about where the work happens.

The consequence was silent and backwards: docs/local-models.md
recommends pointing classifier.base_url at a VPN address rather than
0.0.0.0, and doing that while in encoder mode switched off metering for
a GPU whose electricity is on this machine's own bill, then warned about
a URL nothing calls.

Narrow on purpose: local_llm mode still gates on base_url, because a
remote Ollama genuinely is unmeterable from here.
test_a_remote_ollama_classifier_is_still_not_metered pins that the
narrowing stays narrow.

Found while enabling local_energy on this deployment, which runs
classifier.mode: local_encoder. Metering is confirmed working live --
facebook/bart-large-mnli, 63.35 W over 0.206 s, 3.63e-06 kWh -- since
this host's base_url happens to be localhost. A VPN classifier would
have lost it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U

2026-09-09 21:00:24 -04:00

20 KiB

Raw Permalink Blame History

View Raw