# ParallelHue

A terminal visualization tool for local LLM inference that labels a streaming response as exact token provenance or plain SSE chunk mode, and colors verified speculative-decoding steps rather than arbitrary transport chunks.

ParallelHue renders a local, OpenAI-compatible streaming response with a truthful label describing how much is actually known about it: exact token provenance when scheduler and detokenizer telemetry prove each emitted token's role, or plain SSE chunk mode for ordinary streaming where that proof is not available. It is built for the reality that local LLM serving is increasingly a multi-agent, parallel, batched workload, where a single aggregate tokens-per-second number hides concurrent contribution, stalls and interleaving between requests, so it places each stream's observed accepted-draft, target and bonus tokens side by side instead.

Color is only used when the selected backend profile actually runs speculative decoding; ordinary non-speculative backends stay monochrome, since without verified speculative-step telemetry there is no real step progression for color to encode, and an environment variable can force monochrome for any backend. In exact mode, each newly observed verified step takes the next color in a four-color cycle that wraps around, while a chunk color only ever describes a transport chunk and never claims to identify a token's role. It installs with pip into a Python virtual environment and has no other runtime dependencies.

Website: https://github.com/hikarioyama/ParallelHue