Roles format with JSON wrapper for batching multiple texts.

Call /v1/completions API.

Send raw JSON payload.

Clean up harmony tokens from response.

Get thread-local session for keep-alive.

Initialize DuckDB table.

Initialize DuckDB table.

Load from FLORES-200 (any pair via English pivot).

Load FLORES-200 dataset.

Load from WMT24++ (en->xx only).

Load WMT24++ dataset.

Normalize language to 'FullName (code)' format, e.g. 'ru' -> 'Russian (ru)'.

Parse JSON response from roles format. Returns list of translation strings.

Run summarization using harmony format for gpt-oss models.

Translate multiple texts using JSON batching (roles only). Returns list of results.

* * SC MONITOR * */

Calculate timing statistics from results with timing info.

CLI entry point for benchmark command.

Benchmark audio transcription speed and latency.

Evaluate audio language identification with per-language metrics.

Evaluate audio transcription with WER metrics.

Run audio task (audio_langid, audio_transcribe, etc.).

Send raw JSON payload directly to API.

Compute scores with per-sample caching. 1. Check cache for each sample 2. Batch compute uncached samples 3. Store new

Get FDs and sockets for all processes in a cgroup + docker for vllm

Create local configuration file for SOCKS5 proxy.

Create policies from configuration

Create remote configuration file for reverse proxy.

Download War and Peace from Gutenberg.