An open, agent-native control plane above LLM inference engines.

RIDE Lab studies the policy, coordination, and evidence layer between agent workflows and inference engines. Engines execute prefill, decode, cache, and communication paths. RIDE observes workload intent and live system state, then decides what should run, where, when, with which state, and under which SLO, quality, cost, and correctness contract.

Website · Members · GitHub organization

RIDE turns agent intent and live serving state into accountable execution decisions above one or more inference engines.

Each decision episode is recorded as state -> action and reason -> actuation receipt -> measured outcome. This makes policy behavior reviewable across engines, workloads, and hardware rather than hiding it in ad hoc serving scripts.

RIDE builds cross-engine policy and coordination. It does not replace model kernels, inference-engine scheduling primitives, transport implementations, compiler/runtime internals, hardware backends, or application-level tool planning.

Agent workloads are not interchangeable chat requests. They expose DAG readiness, tool waits, branches, resumptions, deadlines, quality bounds, context reuse, persistent state, and recovery intent. An agent-native inference control plane makes those signals explicit inputs to serving decisions instead of hiding them behind a request-only API.

Agents, workflows, applications
              |
              v
RIDE: observe -> decide -> actuate -> verify
              |
              v
LLM inference engines and serving runtimes
              |
              v
Accelerators, memory, storage, and networks

RIDE owns above-engine decisions: admission, scheduling, routing, placement, budgeting, state/KV policy, recovery, and evidence. It is not another inference engine. Engine patches are admitted only as minimal reusable telemetry or actuation seams; kernels, transport implementations, compiler internals, and hardware backends remain execution-plane work.

Control-plane projects may live in different GitHub organizations while sharing the same RIDE-Lab research architecture. Repository location does not change research-family membership.

Project names and technical roles remain visible throughout development; repository links appear only after public release.

The current project names, links, research status, and evidence boundaries are maintained in the website's versioned portfolio manifest and rendered on the RIDE Lab homepage. This keeps one source of truth instead of duplicating project records across public pages.

• Workflow and SLO: DAG scheduling, admission, deadlines, fairness, and capacity allocation.
• Routing and placement: backend, replica, cache-holder, model, and continuation placement.
• State and KV: reuse prediction, materialization, recomputation, tiering, recovery, and lifecycle control.
• Quality and economics: quality bounds, token budgets, service risk, cost, and safe fallback.

RIDE keeps decision inputs, selected actions, engine receipts, and measured outcomes traceable across workloads and backends. This makes correctness, latency, quality, cost, and fallback behavior visible to both operators and researchers.

Projects retain independent owners and repositories while sharing the RIDE control-plane architecture. Public releases include a clear abstraction, named maintainers, reproducible evaluation, documentation, tests, licensing, and a sustainable release path.

Mao Yancan leads RIDE-Lab operations and the Inference Control Plane program. Project-level academic ownership remains with each named project owner.

• SAGE provides programmable agent and workflow surfaces.
• vLLM-HUST develops inference-engine and hardware-execution capabilities.
• DataSys develops framework-neutral data systems.
• IntelliStream incubates cross-layer research before graduation.

See our contribution guide, security policy, and support guide.