TECH RADAR / SERVING ENGINES
SGLang
Model serving with RadixAttention and prefix reuse.
What it covers
Prefix reuse, speculative decoding, and model-serving paths.
What we would check
Compare warm and cold cache conditions and record accepted draft tokens under the same quality requirements.
Go to the source
A curated ecosystem entry. The evaluation guidance is our interpretation, not a report of a completed test or a performance ranking.