Distributed Inference with llm-d: Release Components Version

Updated -



# Released Components Version

RHOAI Version RHAII version llm-d upstream version vLLM Dates
General Availability (GA) RHOAI 3.5 RHAII 3.5 0.9 v0.24 August 27, 2026
General Availability (GA) RHOAI 3.4 RHAII 3.4 0.6 v0.13.0 May 14, 2026
General Availability (GA) RHOAI 3.3 RHAII 3.3 0.4 v0.13.0 March 5, 2026
Technology Preview (TP) RHOAI 2.25 RHAII 3.2.2 0.2 v0.10.1 October 23, 2025

Components level checklist

Component Version
OpenShift 4.20+

API Compatibility

API format Endpoint Status
OpenAI-compatible Chat Completions /v1/chat/completions Supported
OpenAI-compatible Completions /v1/completions Supported
Anthropic-compatible Messages API /v1/messages Supported
OpenAI-compatible Responses API /v1/responses Not supported
Provider-specific APIs - Not supported

Note: Per-request token usage (prompt_tokens, completion_tokens) is returned in the usage field for text inputs.

GA RHOAI 3.5

Supported configuration(s):

OpenShift: 4.20+

Distributed Inference with llm-d — Supported Well-Lit Paths

A well-lit path is a documented, tested, and benchmarked deployment pattern that reduces adoption risk and maintenance cost.

Well-lit path RHOAI 3.5 maturity Support Boundary
Intelligent Inference Scheduling Generally Available Supported for Valid Configurations
P/D Disaggregation Technology Preview Technology Preview - Could be Supported through Support Exception
Wide Expert Parallelism Developer Preview Developer Preview; not recommended for production workloads
KV-cache Offload to CPU Generally Available Supported
KV-cache Offload to a POSIX Filesystem Developer Preview Developer Preview; not recommended for production workloads

NVIDIA: Hardware & Accelerator Matrix

Well-Lit Path Primary Goal Recommended NVIDIA Hardware Networking/Interconnect Requirement Storage
Intelligent Inference Scheduling Route requests to the most appropriate GPU. H100, H200, B200, B300, A100 Standard DC Ethernet (25/100 GbE) Not Applicable
P/D Disaggregation Separate prefill and decode compute stages. H100, H200, B200, B300 HPC Fabric with RDMA
• InfiniBand
• RoCE
Not Applicable
KV-cache Management (Local CPU Offload) Increase throughput by offloading KV cache to CPU RAM. H100, H200, B200, B300, A100 PCIe 5+ Not Applicable
KV-cache Offload to a POSIX Filesystem Extend KV-cache capacity beyond GPU and CPU memory. H100, H200, B200, B300 Storage-network dependent ReadWriteMany volume providing standard POSIX file access
Wide Expert Parallelism Distribute mixture-of-experts models across multiple GPUs. H100, H200, B200, B300 HPC Fabric with RDMA
• InfiniBand
• RoCE
High-speed NVMe SSDs

AMD: Hardware & Accelerator Matrix

Well-Lit Path Primary Goal Recommended AMD Hardware Networking/Interconnect Requirement Storage
Intelligent Inference Scheduling Route requests to the most appropriate GPU. MI300X Standard DC Ethernet (25/100 GbE) Not Applicable
KV-cache Management (Local CPU Offload) Increase throughput by offloading KV cache to CPU RAM. MI300X PCIe 5+ Not Applicable
KV-cache Offload to a POSIX Filesystem Extend KV-cache capacity beyond GPU and CPU memory. MI300X Storage-network dependent ReadWriteMany volume providing standard POSIX file access

Mixed Accelerator Architectures

Mixed accelerator architecture combines different hardware generations from the same vendor within one inference cluster, for example NVIDIA H200 and B200.

Inference architecture Mixed Accelerator Support
Intelligent Inference Scheduling Supported
P/D Disaggregation Not supported
Wide Expert Parallelism Not supported

Approximate and Precise Cache Scoring

Selection criterion Approximate scoring (default) Precise scoring (experimental)
How it works Predicts cache locality from request traffic patterns Uses real-time cache state across the cluster
External dependencies None Requires access to model tokenizer artifacts
Disconnected support Yes No, unless required tokenizer artifacts are mirrored internally and the configuration is validated
Status GA and recommended for most deployments Experimental
Best suited for Standard workloads and benchmarks Workloads with highly variable prefix patterns

Approximate scoring delivers strong cache-hit rates across standard benchmarks and production workloads without additional configuration.

GA RHOAI 3.4

Supported configuration(s):

Note:
- P/D Disaggregation : Technology Preview
- Wide Expert-Parallelism multi-node: Developer Preview
- Wide Expert-Parallelism on Blackwell B200: Not available but can be provided as a Developer Preview
- Multi node on GB200 is not supported

Hardware and Accelerator support per "Well-Lit" paths for distributed inference with llm-d
A well-lit path is a documented, tested, and benchmarked deployment pattern that reduces adoption risk and maintenance cost.

NVIDIA: Hardware & Accelerator Matrix

Well-Lit Path Primary Goal Recommended NVIDIA Hardware Networking/Interconnect Requirement Storage
Intelligent Inference Scheduling Route requests to the most optimal GPU. H100, H200, B200, A100 Standard DC Ethernet (25/100 GbE) Local SSD (NVMe Recommended)
P/D Disaggregation Separate prefill and decode compute stages. H100, H200, B200 HPC Fabric with RDMA
• InfiniBand
• RoCE
Local SSD (NVMe Recommended)
KV Cache Management (Local CPU Offload) Increase throughput by offloading KV cache to CPU RAM. H100, H200, B200, A100 PCIe 5+ Not Applicable
Wide Expert Parallelism (WEP) Distribute MoE models across many GPUs. H100, H200, B200 HPC Fabric with RDMA
• InfiniBand
• RoCE
High-speed NVMe SSDs.

AMD: Hardware & Accelerator Matrix

Well-Lit Path Primary Goal Recommended AMD Hardware Networking/Interconnect Requirement Storage
Intelligent Inference Scheduling Route requests to the most optimal GPU. MI300X Standard DC Ethernet (25/100 GbE) Local SSD (NVMe Recommended)
KV Cache Management (Local CPU Offload) Increase throughput by offloading KV cache to CPU RAM. MI300X PCIe 5+ Not Applicable

Mixed Accelerator Architectures
Mixed accelerator refers to combining different hardware generations from the same vendor (e.g., NVIDIA H200 and B200) within the same inference cluster.

Feature Mixed Accelerator Support
Intelligent Inference Scheduling Supported
P/D Disaggregation Not supported
Wide Expert Parallelism Not supported

Intelligent Inference Scheduling

The inference scheduler balances requests across load and prefix cache locality to optimize throughput and latency. Two cache scoring methods are available:

Approximate (default) Precise (experimental)
How it works Predicts cache locality from request traffic patterns Uses real-time cache state across the cluster
External dependencies None Requires HuggingFace connectivity to download the model tokenizer
Air-gapped support Yes No
Status GA, recommended for most deployments Experimental, GA expected in a future release
Best for Standard workloads and benchmarks Workloads with highly variable prefix patterns

The approximate scorer delivers strong cache hit rates across standard benchmarks and production workloads with no additional configuration.

GA RHOAI 3.3

Supported configuration(s):

Note:
- P/D Disaggregation : Technology Preview
- Wide Expert-Parallelism multi-node: Developer Preview
- Wide Expert-Parallelism on Blackwell B200: Not available but can be provided as a Developer Preview
- Multi node on GB200 is not supported

Hardware and Accelerator support per "Well-Lit" paths for distributed inference with llm-d
A well-lit path is a documented, tested, and benchmarked deployment pattern that reduces adoption risk and maintenance cost.

NVIDIA: Hardware & Accelerator Matrix

Well-Lit Path Primary Goal Recommended NVIDIA Hardware Networking/Interconnect Requirement Storage
Intelligent Inference Scheduling Route requests to the most optimal GPU. H100, H200, B200, A100 Standard DC Ethernet (25/100 GbE) Local SSD (NVMe Recommended)
P/D Disaggregation Separate prefill and decode compute stages. H100, H200, B200 HPC Fabric with RDMA
• InfiniBand
• RoCE
Local SSD (NVMe Recommended)
KV Cache Management (Local CPU Offload) Increase throughput by offloading KV cache to CPU RAM. H100, H200, B200, A100 PCIe 5+ Not Applicable
Wide Expert Parallelism (WEP) Distribute MoE models across many GPUs. H100, H200, B200 HPC Fabric with RDMA
• InfiniBand
• RoCE
High-speed NVMe SSDs.

AMD: Hardware & Accelerator Matrix

Well-Lit Path Primary Goal Recommended AMD Hardware Networking/Interconnect Requirement Storage
Intelligent Inference Scheduling Route requests to the most optimal GPU. MI300X Standard DC Ethernet (25/100 GbE) Local SSD (NVMe Recommended)
KV Cache Management (Local CPU Offload) Increase throughput by offloading KV cache to CPU RAM. MI300X PCIe 5+ Not Applicable

Mixed Accelerator Architectures
Mixed accelerator refers to combining different hardware generations from the same vendor (e.g., NVIDIA H200 and B200) within the same inference cluster.

Feature Mixed Accelerator Support
Intelligent Inference Scheduling Supported
P/D Disaggregation Not supported
Wide Expert Parallelism Not supported

Intelligent Inference Scheduling

The inference scheduler balances requests across load and prefix cache locality to optimize throughput and latency. Two cache scoring methods are available:

Approximate (default) Precise (experimental)
How it works Predicts cache locality from request traffic patterns Uses real-time cache state across the cluster
External dependencies None Requires HuggingFace connectivity to download the model tokenizer
Air-gapped support Yes No
Status GA, recommended for most deployments Experimental, GA expected in a future release
Best for Standard workloads and benchmarks Workloads with highly variable prefix patterns

The approximate scorer delivers strong cache hit rates across standard benchmarks and production workloads with no additional configuration.

Tech Preview - RHOAI 2.25

Supported configuration:

Note: WIDE EP multi-node support is included in this Tech Preview, but it may not function as expected and is not yet stable.

Hardware and Accelerator support per "Well-Lit" paths for distributed inference with llm-d
A well-lit path is a documented, tested, and benchmarked deployment pattern that reduces adoption risk and maintenance cost.

NVIDIA: Hardware & Accelerator Matrix

Well-Lit Path Primary Goal Recommended NVIDIA Hardware Networking/Interconnect Requirement Storage
Intelligent Inference Scheduling Route requests to the most optimal GPU. H100, H200, B200, A100 Standard DC Ethernet (25/100 GbE) Local SSD (NVMe Recommended)
P/D Disaggregation Separate prefill and decode compute stages. H100, H200, B200 HPC Fabric with RDMA
• InfiniBand
• RoCE
Local SSD (NVMe Recommended)
KV Cache Management (Local CPU Offload) Increase throughput by offloading KV cache to CPU RAM. H100, H200, B200, A100 PCIe 5+ Not Applicable
Wide Expert Parallelism (WEP) Distribute MoE models across many GPUs. H100, H200, B200, ~~GB200~~ NVL72 HPC Fabric with RDMA
• InfiniBand
• RoCE
High-speed NVMe SSDs.

AMD: Hardware & Accelerator Matrix

Well-Lit Path Primary Goal Recommended AMD Hardware Networking/Interconnect Requirement Storage
Intelligent Inference Scheduling Route requests to the most optimal GPU. MI300X Standard DC Ethernet (25/100 GbE) Local SSD (NVMe Recommended)
KV Cache Management (Local CPU Offload) Increase throughput by offloading KV cache to CPU RAM. MI300X PCIe 5+ Not Applicable

Mixed Accelerator Architectures
Mixed accelerator refers to combining different hardware generations from the same vendor (e.g., NVIDIA H200 and B200) within the same inference cluster.

Feature Mixed Accelerator Support
Intelligent Inference Scheduling Supported
P/D Disaggregation Not supported
Wide Expert Parallelism Not supported

Intelligent Inference Scheduling

The inference scheduler balances requests across load and prefix cache locality to optimize throughput and latency. Two cache scoring methods are available:

Approximate (default) Precise (experimental)
How it works Predicts cache locality from request traffic patterns Uses real-time cache state across the cluster
External dependencies None Requires HuggingFace connectivity to download the model tokenizer
Air-gapped support Yes No
Status GA, recommended for most deployments Experimental, GA expected in a future release
Best for Standard workloads and benchmarks Workloads with highly variable prefix patterns

The approximate scorer delivers strong cache hit rates across standard benchmarks and production workloads with no additional configuration.

Comments