llm-d

Distributed LLM inference stack with cache-aware routing over vLLM model servers.

Provider: default
Kind: ML inference
Category: Compute
Official docs: https://llm-d.ai/docs/architecture

Properties · 3 set by this component

property	type	values	required	kind default	this component
accelerator	enum	none \| gpu \| asic	no	none	▸ gpu
acceleratorModel · advanced	enum	nvidia-h100 \| nvidia-h200 \| nvidia-b200 \| nvidia-l4 \| nvidia-a100 \| amd-mi300x \| aws-trainium \| aws-inferentia \| google-tpu \| intel-gaudi \| groq-lpu \| cerebras-wse	no	—	inherited
cacheAwareRouting · component property	boolean	—	no	true	inherited
embeddingSyncMinutes · advanced	integer	—	no	—	inherited
inputs	enum	live \| historical \| both	no	live	▸ live
instances	integer	—	no	—	inherited
managed	boolean	—	no	false	▸ false
modelServer · component property	enum	vllm \| sglang	no	vllm	inherited
prefillDecodeDisaggregation · component property	boolean	—	no	false	inherited
snapshotToObjectStore · advanced	boolean	—	no	false	inherited
training	enum	batch \| online	no	batch	inherited

Rules that apply · 1

Every rule whose selector covers this component's kind or its category. Whether a rule fires depends on the design: its properties, its edges and its groups.

Stateful node without zone redundancyerror

stateful-node-without-zone-redundancy

A stateful compute node pinned to exactly one availability zone disappears with that zone, and its state goes with it. Nodes inside a cell are exempt, because there the cell is deliberately single-zone and redundancy comes from the other cells.

— all of these hold:
— the node's `stateless` is false
— the node sits inside exactly 1 availability-zone group
— this does not hold:
— the node sits inside at least one cell group

Use it in a design

nodes:
  - { id: "llm-d", component: "llm-d" }

Other ML inference products · 37

SageMaker endpointsagemaker-endpointawsml-inference Azure ML endpointazure-ml-endpointazureml-inference Workers AIcloudflare-workers-aicloudflareml-inference Anyscale Endpointsanyscale-endpointsml-inference Bananabanana-devml-inference Basetenbasetenml-inference Beambeam-cloudml-inference BentoMLbentomlml-inference Hugging Face Inference Endpointshuggingface-inferenceml-inference Infinityinfinity-embeddingsml-inference Janjanml-inference KServekserveml-inference llama.cpp serverllama-cpp-serverml-inference LM Studiolm-studioml-inference LMDeploylmdeployml-inference LocalAIlocalaiml-inference Generic ML inferenceml-inferenceml-inference MLC LLMmlc-llmml-inference Modalmodalml-inference Mystic AImystic-aiml-inference NVIDIA NIM (self-hosted)nvidia-nimml-inference OctoAIoctoaiml-inference Ollamaollamaml-inference Ray Serveray-serveml-inference Seldon Coreseldon-coreml-inference SGLangsglangml-inference TensorFlow Servingtensorflow-servingml-inference TensorRT-LLMtensorrt-llmml-inference Text Embeddings Inferencetext-embeddings-inferenceml-inference Text Generation Inferencetext-generation-inferenceml-inference TorchServetorchserveml-inference NVIDIA Tritontritonml-inference vLLMvllmml-inference Xinferencexinferenceml-inference Vertex AI endpointvertex-ai-endpointgcpml-inference OCI Data Science Model Deploymentoci-data-science-model-deploymentociml-inference OCI Generative AI Dedicated AI Clusteroci-generative-ai-clusterociml-inference

Source

1# yaml-language-server: $schema=../../schema/component.schema.json
2# llm-d: distributed LLM inference serving stack for Kubernetes. Per the llm-d README it runs
3# model servers such as vLLM behind an inference-aware scheduler, adding KV-cache aware routing and
4# optional prefill/decode disaggregation across accelerators.
5schemaVersion: "1"
6type: "component"
7id: "llm-d"
8kind: "ml-inference"
9provider: "default"
10docs: "https://llm-d.ai/docs/architecture"
11defaults:
12  managed: false
13  accelerator: "gpu"
14  inputs: "live"
15properties:
16  modelServer:
17    type: "enum"
18    values: ["vllm", "sglang"]
19    default: "vllm"
20  prefillDecodeDisaggregation:
21    type: "boolean"
22    default: false
23  cacheAwareRouting:
24    type: "boolean"
25    default: true

The same page as markdown, for an agent: /catalogue/components/llm-d.md