[T0mSIlver]

Proposed body for review on the fork. Rewrite it in your own words before opening upstream (llama.cpp does not accept AI-written PR text).

Overview

Adds tools/server/tests/unit/test_realtime.py, the first tests for the /v1/realtime WebSocket endpoint.

• test_realtime_unsupported_model runs in CI on tinygemma3: a model without voxtral_rt audio streaming gets a single error event.
• test_realtime_control_events (slow): session.update, input_audio_buffer.clear, an unknown event and a final commit produce session.updated, input_audio_buffer.cleared, error and completed, in that order.
• test_realtime_commit (slow): the clip, a non-final commit, the clip again, then a final commit. It checks one committed, two completed with matching item ids, and that each item holds the whole sentence. This is the regression test for Misc. bug: server: realtime commit completes the item before the flushed words are decoded ggml-org/llama.cpp#30968: before server: realtime API: send complete after committed audio drains ggml-org/llama.cpp#30994, the words held back by the model delay were attributed to the next item.

The slow tests use ggml-org/Voxtral-Mini-4B-Realtime-2602-GGUF:Q4_K_M and whisper.cpp's jfk.wav, and run with --temp 0. The model is not a ServerPreset on purpose, because load_all() would download it for every test session.

Additional information

Run on an RTX 3090, CUDA Release build of master 9145ddb:

./tests.sh unit/test_realtime.py                  1 passed, 2 skipped
SLOW_TESTS=1 ./tests.sh unit/test_realtime.py     3 passed

At 1242be6 (before ggml-org#30994), test_realtime_commit fails as expected:

item 1: "And so, my fellow Americans, ask not what your country can do for you."
item 2: "Ask what you can do for your country. ..."
AssertionError (AUDIO_RE)

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES.

---

---

Generated by Claude Code