Release v0.8.1

LLM-D v0.8.0 Component Summary

Themes: solidify CI coverage & project operations, expand accelerator coverage, graduate multimodal/batch/flow-control to production, introduce initial RL support.

Component	Version	Previous Version	Type	Notes
llm-d/llm-d-router-endpoint-picker	v0.9.0	v0.8.0	Image + Helm Chart	Core EPP image (renamed from llm-d-inference-scheduler)
llm-d/llm-d-router-disagg-sidecar	v0.9.0	v0.8.0	Image	P/D routing sidecar (renamed from llm-d-routing-sidecar)
llm-d/llm-d-uds-tokenizer	vllm-v0.23.0	vllm-v0.19.1	Image	Tokenizer sidecar aligned with vLLM version
llm-d/llm-d-kv-cache	v0.9.0	v0.8.0	Library	HMA support, storage events, multi-tier offloading
llm-d/llm-d-inference-sim	v0.9.2	v0.8.2	Image	Multimodal support, Mooncake bootstrap, configurable latency
llm-d/llm-d-cuda	v0.8.0	v0.7.0	Image	vLLM v0.23.0, CUDA 13.0.2
llm-d/llm-d-aws (EFA)	v0.8.0	v0.7.0	Image
llm-d/llm-d-hpu	v0.8.0	v0.7.0	Image
llm-d/llm-d-kv-cache/llmd_fs_backend_connector	v0.23	v0.19.1	Wheel installed in llm-d	Migrated to vLLM 0.23.0 offload API
llm-d/llm-d-benchmark	v0.7.0	v0.6.8.1	Image	Benchmark workload launcher
llm-d/llm-d-workload-variant-autoscaler	v0.8.0	v0.7.0	Helm Chart + Image	CRD migration to llm-d.ai API group, improved observability
vllm-project/vllm	v0.23.0	v0.19.1	Wheel installed in llm-d	Confirmed by @Gregory-Pereira and @tessapham
kubernetes-sigs/gateway-api-inference-extension	v1.5.0	v1.5.0	Helm Chart	Charts now published from llm-d-router OCI registry

Upstream vLLM Images (replacing llm-d-built images)

Per PR #1791, the following platforms now use upstream vLLM images directly instead of llm-d-built custom images:

Platform	New Image	Tag	Previous llm-d Image
GPU (CUDA)	vllm/vllm-openai	v0.23.0	ghcr.io/llm-d/llm-d-cuda (still available for advanced builds)
ROCm (AMD)	vllm/vllm-openai-rocm		ghcr.io/llm-d/llm-d-rocm
CPU	ghcr.io/llm-d/llm-d-cpu	v0.7.0	Same image, version bump
XPU (Intel)	vllm/vllm-openai	v0.23.0	ghcr.io/llm-d/llm-d-xpu

---

Infrastructure Changes

Component	Version	Previous Version	Notes
Gateway API	v1.5.1	v1.5.1	No change
Istio	1.29.4	1.29.1	Patch update
kgateway (agentgateway)	v2.3.3	v2.2.1

---

Deprecated / Removed Components

Component	Status	Replaced By
llm-d/llm-d-inference-scheduler	Renamed	ghcr.io/llm-d/llm-d-router-endpoint-picker
llm-d/llm-d-routing-sidecar	Renamed	ghcr.io/llm-d/llm-d-router-disagg-sidecar
llm-d/llm-d-cuda (debug)	Removed	N/A
llm-d/llm-d-cuda-gb200	Removed	N/A
llm-d/llm-d-xpu	Removed	vllm/vllm-openai
llm-d/llm-d-rocm	Removed	vllm/vllm-openai-rocm

---

New Capabilities in This Release

Capability	Component(s)	Status
Multi-modal serving (production)	llm-d-router, llm-d-inference-sim	Graduated
Batch gateway (production)	llm-d-router	Graduated
Flow control (production)	llm-d-router	Graduated
Non-Kubernetes mode (RL/Slurm)	llm-d-router (FileDiscovery plugin)	New
Responses API support	llm-d-router	New
Multi-tier KV offloading (CPU → storage)	llm-d-kv-cache	New
HMA (Heterogeneous Memory Allocation) support	llm-d-kv-cache, fs-connector	New
DP-Aware scheduling	llm-d-router	Graduating
Mooncake connector	llm-d-kv-cache	New
Predicted latency scheduling	llm-d-router	New
Agentic workload routing	llm-d-router	New
TPU nightly tests	CI/CD	New

What's Changed

• Point to patch v0.8.1 by @maugustosilva in commit d6ffb5c
• Updated the branch to clone in the guides to release-0.8 branch by @ahg-g in #1958
• Pin inference-perf version by @diegocastanibm in #1946

Full Changelog: v0.8.0...v0.8.1