Gateway API Inference Extension

Gateway API Inference Extension optimizes self-hosting Generative Models on Kubernetes. This is achieved by leveraging Envoy's External Processing (ext-proc) to extend any gateway that supports both ext-proc and Gateway API into an inference gateway.

New!

Inference Gateway has partnered with vLLM to accelerate LLM serving optimizations with llm-d!

[!IMPORTANT] The Endpoint Picker (EPP), InferenceObjective and InferenceModelRewrite APIs, and Body Based Router (BBR) packages have moved to new repositories:

• EPP and associated APIs: llm-d/llm-d-router
• BBR: llm-d/llm-d-inference-payload-processor

No new code will be accepted to these packages in this repository, and they will be archived soon. This move was proposed and discussed in issue #2430.

This repository will continue to host the lightweight EPP (LWEPP) and the InferencePool API, and will remain the primary location for the development and maintenance of conformance tests.

Concepts and Definitions

The following specific terms to this project:

• Inference Gateway (IGW): A proxy/load-balancer which has been coupled with an Endpoint Picker. It provides optimized routing and load balancing for serving Kubernetes self-hosted generative Artificial Intelligence (AI) workloads. It simplifies the deployment, management, and observability of AI inference workloads.
• Inference Scheduler: An extendable component that makes decisions about which endpoint is optimal (best cost / best performance) for an inference request based on Metrics and Capabilities from Model Serving.
• Metrics and Capabilities: Data provided by model serving platforms about performance, availability and capabilities to optimize routing. Includes things like Prefix Cache status or LoRA Adapters availability.
• Endpoint Picker (EPP): A data-plane component that communicates via the Envoy external processing protocol, and acts as the Router. It intercepts incoming inference requests and routes each request to the optimal model server replica. This repository provides a reference Lightweight Endpoint Picker (lwepp) for conformance test purposes.

The following are key industry terms that are important to understand for this project:

• Model: A generative AI model that has learned patterns from data and is used for inference. Models vary in size and architecture, from smaller domain-specific models to massive multi-billion parameter neural networks that are optimized for diverse language tasks.
• Inference: The process of running a generative AI model, such as a large language model, diffusion model etc, to generate text, embeddings, or other outputs from input data.
• Model server: A service (in our case, containerized) responsible for receiving inference requests and returning predictions from a model.
• Accelerator: specialized hardware, such as Graphics Processing Units (GPUs) that can be attached to Kubernetes nodes to speed up computations, particularly for training and inference tasks.

For deeper insights and more advanced concepts, refer to our proposals.

Technical Overview

This extension upgrades an ext-proc capable proxy or gateway - such as Envoy Gateway, kgateway, or the GKE Gateway - to become an inference gateway - supporting inference platform teams self-hosting Generative Models (with a current focus on large language models) on Kubernetes. This integration makes it easy to expose and control access to your local OpenAI-compatible chat completion endpoints to other workloads on or off cluster, or to integrate your self-hosted models alongside model-as-a-service providers in a higher level AI Gateway like LiteLLM, Solo AI Gateway, or Apigee.

The Inference Gateway:

• Improves the tail latency and throughput of LLM completion requests against Kubernetes-hosted model servers using an extensible request scheduling algorithm that is kv-cache and request cost aware, avoiding evictions or queueing as load increases
• Provides Kubernetes-native declarative APIs to route client model names to use-case specific LoRA adapters and control incremental rollout of new adapter versions, A/B traffic splitting, and safe blue-green base model and model server upgrades
• Adds end to end observability around service objective attainment
• Ensures operational guardrails between different client model names, allowing a platform team to safely serve many different GenAI workloads on the same pool of shared foundation model servers for higher utilization and fewer required accelerators

Model Server Integration

IGW’s pluggable architecture was leveraged to enable the llm-d Router.

Llm-d customizes vLLM & IGW to create a disaggregated serving solution. We've worked closely with this team to enable this integration. IGW will continue to work closely with llm-d to generalize the disaggregated serving plugin(s), & set a standard for disaggregated serving to be used across any protocol-adherent model server.

IGW has enhanced support for vLLM via llm-d, and broad support for any model servers implementing the protocol. More details can be found in model server integration.

Status

This project is GA'd! The latest release can be found here.
Please file any bugs or feature requests you have. We are always happy to welcome new collaborators and users.

Getting Started

Follow our Getting Started Guide to get the inference-extension up and running on your cluster!

See our website for detailed API documentation on leveraging our Kubernetes-native declarative APIs

Roadmap

As Inference Gateway builds towards a GA release. We will continue to expand our capabilities, namely:

1. Prefix-cache aware load balancing with interfaces for remote caches
2. Recommended LoRA adapter pipeline for automated rollout
3. Fairness and priority between workloads within the same criticality band
4. HPA support for autoscaling on aggregate metrics derived from the load balancer
5. Support for large multi-modal inputs and outputs
6. Support for other GenAI model types (diffusion and other non-completion protocols)
7. Heterogeneous accelerators - serve workloads on multiple types of accelerator using latency and request cost-aware load balancing
8. Disaggregated serving support with independently scaling pools

Contributing

Community meetings have moved to the llm-d Router community meeting. See the llm-d Router contributing section for current meeting details.

We currently utilize the #gateway-api-inference-extension channel in Kubernetes Slack workspace for communications.

Contributions are readily welcomed!

Code of conduct

Participation in the Kubernetes community is governed by the Kubernetes Code of Conduct.