Get thread-local connection.

CHARGE AND TOP UP */

Format value with diff from best, using rich markup for colors.

Collect all metrics

Helper: find cgroup path for a service (tries system.slice then user.slice)

Helper: get docker container cgroup paths for cocoon-vllm

Get the headers for API requests.

Simple hex dump utility for debugging

Calculate overheads at each stage. Returns: (client_overhead, proxy_overhead, worker_overhead) - worker_overhead: time spent

Parse systemctl show output for service metrics

Print statistics.

UTILS */

Unified translation function. Args: text: Single text (str) or list of texts (List[str]) target_lang: Target language

Validate port number

Aggregate per-sample scores into final metrics. Default: mean score. Override for custom aggregation.

Compute per-sample scores. Returns list of floats (one per sample). This is the raw computation - no caching. Override in subclass.

Download model using hf CLI.

Create TimingInfo from debug JSON response. Args: debug_data: Dict with 'client', 'proxy', 'worker' keys, each containing