live wire
▸AI · Red Hat documents usage-based admission fair sharing for Kueue 1.4 on OpenShiftRed Hat Developer▸AI: Red Hat maps governed firewall changes from ServiceNow through Ansible and two human approval gatesRed Hat Developer▸CLUSTER MGMT · ACM 2.17 makes Submariner 0.24 GA with Important-rated fixesRed Hat Errata▸PLATFORM · Red Hat makes on-premises Lightspeed recommendations GA for Satellite 6.18Red Hat Errata▸SECURITY · Red Hat Hardened Images updates Tomcat 10 for nine authentication, access-control and DoS flawsRed Hat Errata▸AI · Open Data Hub 3.6.0 EA1 bundles Trainer, MLflow and llm-d componentsOpen Data Hub▸AI · Speculators 0.6.0 adds P-EAGLE parallel drafting for vLLM speculative decodingRed Hat Developer▸SECURITY · OpenShift 4.17.57 fixes seven Go and TLS CVEs in an Important-rated updateRed Hat Errata▸AI · Red Hat benchmarks local LLM guardrails with EvalHub, exposing regex accuracy and latency trade-offsRed Hat Developer▸AI · Red Hat maps silent tool-call failures across agentic pipelinesRed Hat▸API · Kuadrant 1.5.3 adds GRPCRoute policies and developer-portal API-key workflowsKuadrant▸AI · (Aug 25) IBM releases Apache-2.0 Granite 4.2 reasoning models in 3B, 8B and 30B sizesIBM Research▸JAVA · Red Hat build of Quarkus 3.33.3.SP1 fixes 13 CVEs in an Important-rated updateRed Hat errata▸AI · vLLM moves Kimi K2 RL weight sync across 384 H100s in 7.53 seconds (Aug 22)vLLM▸AI · Red Hat documents usage-based admission fair sharing for Kueue 1.4 on OpenShiftRed Hat Developer▸AI: Red Hat maps governed firewall changes from ServiceNow through Ansible and two human approval gatesRed Hat Developer▸CLUSTER MGMT · ACM 2.17 makes Submariner 0.24 GA with Important-rated fixesRed Hat Errata▸PLATFORM · Red Hat makes on-premises Lightspeed recommendations GA for Satellite 6.18Red Hat Errata▸SECURITY · Red Hat Hardened Images updates Tomcat 10 for nine authentication, access-control and DoS flawsRed Hat Errata▸AI · Open Data Hub 3.6.0 EA1 bundles Trainer, MLflow and llm-d componentsOpen Data Hub▸AI · Speculators 0.6.0 adds P-EAGLE parallel drafting for vLLM speculative decodingRed Hat Developer▸SECURITY · OpenShift 4.17.57 fixes seven Go and TLS CVEs in an Important-rated updateRed Hat Errata▸AI · Red Hat benchmarks local LLM guardrails with EvalHub, exposing regex accuracy and latency trade-offsRed Hat Developer▸AI · Red Hat maps silent tool-call failures across agentic pipelinesRed Hat▸API · Kuadrant 1.5.3 adds GRPCRoute policies and developer-portal API-key workflowsKuadrant▸AI · (Aug 25) IBM releases Apache-2.0 Granite 4.2 reasoning models in 3B, 8B and 30B sizesIBM Research▸JAVA · Red Hat build of Quarkus 3.33.3.SP1 fixes 13 CVEs in an Important-rated updateRed Hat errata▸AI · vLLM moves Kimi K2 RL weight sync across 384 H100s in 7.53 seconds (Aug 22)vLLM
upstreambeat.ai
releaseAI

vLLM 0.28 moves KV-cache offload to disk and expands its Rust serving path

The release adds tiered cache management, multimodal gRPC inference and broad accelerator work while removing several deprecated interfaces.

Old vLLM cache path versus new disk tier and Rust gRPC serving path.
Side by side: what changed
By The Release Desk· Aug 26, 2026

vLLM 0.28 is available with a wider serving surface and a set of compatibility changes that operators should review before moving existing inference deployments.

What changed

The v0.28.0 release, published August 26, adds disk as a secondary tier for KV-cache offloading, exposes out-of-tree secondary-tier managers, and reports tiering metrics. The release also extends the Rust frontend with multimodal image inference over gRPC, explicit data-parallel routing and reinforcement-learning lifecycle controls.

Model and hardware work spans Kimi-K3, DeepSeek V4, NVIDIA, AMD ROCm, Intel XPU, CPU and IBM architectures. vLLM also raises the default maximum batched-token count from 8,192 to 16,384 and enables prefix caching by default for Mamba models.

Who it affects

Teams operating large-context or disaggregated inference can use the new cache tiers and connector work to trade local accelerator memory against host or disk capacity. Platform teams exposing vLLM through custom frontends gain a more capable gRPC and Rust path.

The upgrade is not drop-in for every deployment. Bitsandbytes support moved to an out-of-tree plugin, Transformers moved to 5.15.0, runtime KV-scale calculation through calculate_kv_scales was removed, and override_attention_dtype was removed.

What to do

Test existing model runners, quantization plugins and client integrations against 0.28 before changing production images. Operators using KV offload should validate tiering behavior and renamed metrics in staging; deployments relying on removed options need configuration changes before rollout.

sources

Filed by The Release Desk. Corrections: desk@upstreambeat.ai · Our standards →

comments · 0

    Comments are moderated before they appear. Your email is used once to confirm it is you — never shown, never sold. Corrections and questions get an answer from the desk when we have one.

    more from the desk