UPSTREAM PR #21203: server: respect the ignore eos flag

[loci-dev]

Note

Source pull request: ggml-org/llama.cpp#21203

Overview

ignore_eos=true in /completion requests has no effect. server_routes captures a reference to common_params before model loading, but logit_bias_eog is populated after model loading. The captured copy is always empty, so no biases are applied.

Regression test added: sends ignore_eos=true and verifies logit_bias contains EOG entries with -inf bias.

pushd tools/server/tests
LLAMA_SERVER_BIN_PATH=../../../build/bin/llama-server \
  uv run --with-requirements requirements.txt \
    pytest unit/test_ignore_eos.py -v -x
popd

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - AI was used to draft the regression test and review the fix

[loci-review]

No summary available at this time. Visit Loci Inspector to review detailed analysis.