Feature matrix

One-pager

At a glance, this table shows which Infera features are available on each engine. Legend: ✅ supported · 🚧 work in progress · blank = not supported.

Overview

Feature	vLLM	SGLang	ATOM	Source
Disaggregated Serving (PD)	✅	✅	✅	PD Disaggregation
KV-Aware Routing	✅	✅	✅	KV-Aware Routing
KV-Aware Routing + DP-Attention	✅	✅	✅	KV-Aware Routing
Tiered KV Cache Offload (kvd)	✅	🚧	🚧	KV Cache Offload
SLA Planner (PD replica sizing)	✅	✅	✅	SLA Planner
Multimodal (image / audio / video)
Request Migration	✅	✅	✅	Request Migration

KV-cache offload (kvd), including AIC GPU-Direct, is vLLM-only today. Request migration does not depend on the engine, but does require the NATS request transport and mixed (non-PD) workers; it is off unless enabled.

The SLA planner is engine-agnostic — it reads the server’s metrics and decides how many replicas each pool needs — but it only covers disaggregated deployments, needs a profiling sweep of your own model to work from, and does not carry out the resize itself.