Roadmap¶
This page describes the direction of the LoxiLB Inference Gateway. It is a guide to intent, not a commitment to dates — priorities are shaped by the community and by real deployment feedback. The authoritative, up-to-the-minute view is the issue tracker.
“Available” below describes the current documented implementation. A capability is part of a production release only when it is present in a published, immutable image and has passed the deployment's release qualification.
Available today¶
- Model-name routing — per-model backend pools with wildcard fallback.
- Engine-aware routing for vLLM, SGLang, TensorRT-LLM, and llama.cpp, with validation that rejects unsupported engine/feature combinations.
- KV-cache-aware routing for vLLM, SGLang, and TensorRT-LLM through engine-specific event contracts; llama.cpp uses CHWBL or session affinity without KV events.
- Engine-specific prefill/decode disaggregation for vLLM, SGLang, and TensorRT-LLM.
- Consistent-hash cache routing (CHWBL) and GPU-aware selection.
- OpenAI-compatible SSE streaming with token accounting.
- MCP gateway with session-sticky routing.
- API-key authentication, model authorization, request-rate limits, and tenant/model token quotas enforced on the inference path.
- Frontend client-certificate verification for qualified FullProxy paths; backend identity verification remains an implementation gap.
- Versioned Gateway configuration snapshots with dry-run/commit restore, automatic rollback, atomic persistence, and boot replay.
- HTTP/L4 tracing and OTLP export in the trace-enabled Ubuntu 24.04 build profile.
- A Prometheus + Grafana monitoring stack and the
loxilb-mcpmanagement surface. - AI-aware
loxicmdcommands for load-balancer rules, keys, limits, metrics, GPU state, OPA, SNI, and KV inventory.
Current focus¶
- Broader scenario coverage across supported engine versions and deployment topologies.
- Safer multi-node promotion, state reconciliation, and upgrade evidence. See HA and Upgrade Limitations for the current boundaries.
- Additional dashboards and alerts for authorization, token quotas, engine event health, and QoS.
- Release packaging and representative validation for build-tag-dependent AI safety and DPU capabilities.
Maturing / experimental¶
- Adaptive routing controller (
loxilb-ai-controller) — TTFT-aware, capacity-weighted prefill selection for heterogeneous fleets. Experimental and off by default. - Tiered caching with LMCache — a CPU (and optional remote) KV tier beneath the GPU cache. Advanced; gate before production rollout.
- OPA L4 policy watcher — desired-state firewall synchronization is implemented, but authenticated internal rule apply, meaningful fail-mode behavior, atomic document validation, cache permissions, and complete SSRF controls remain release gates.
- Presidio PII scanning — optional source build, currently request-direction only; standard release images use the stub implementation.
- Llama Firewall — source integration and API contract exist, but standard release images use a stub and no supported release packaging path is defined.
- NVIDIA DOCA DPU offload — optional hardware/source build that requires target-specific SDK, driver, firmware, correctness, fallback, and performance qualification.
Longer term¶
- Broader serving-engine coverage and more turnkey deployment recipes.
- Expanded dashboards and alerting for inference-specific SLOs.
Have a use case or a feature request? Open an issue in the code repository or say hello on Slack.