# KServe v0.16.0

2025-11-03 · [Original release notes](https://github.com/kserve/kserve/releases/tag/v0.16.0) · [Full summary](https://ratatosk.io/en/releases/kserve/v0.16.0)

Security 2 · Breaking changes 19 · Deprecations 1 · Important fixes 28 · Enhancements 43 · Dependency updates 8 · Other changes 49

## Security (2)

### `RuntimeDefault` seccomp profiles added to KServe controllers

- Affects: Deployments using the KServe controller, LocalModel controller, or LocalModelNode agent manifests, including Helm chart controller deployments (Inferred)

Previously, the shown pod security contexts did not set `seccompProfile`, and the Helm chart's documented controller security context default was `{"runAsNonRoot":true}`. The affected contexts now specify `seccompProfile.type: RuntimeDefault`.

> * Fix pss restricted warnings by @akagami-harsh in https://github.com/kserve/kserve/pull/4327

### Storage-initializer HF token exposure in pod manifests fixed

- Affects: storage-initializer workloads using an HF token stored in a Kubernetes Secret (Inferred)
- Additional action: `HF_TOKEN` is injected using `valueFrom` with a `secretKeyRef` to the Kubernetes Secret instead of embedding its plaintext value in the pod manifest.

The HF token was injected as plaintext in the pod manifest, exposing it to potential leaks.

> * Fix HF Token Vulnerability in Storage Initializer Container by @brettmthompson in https://github.com/kserve/kserve/pull/4677

## Breaking changes (19)

### InferenceService services created without the `default` suffix

- Affects: InferenceService components using the `default` suffix

These components were previously supported with the suffix.

Evidence: “With this PR, the support for compatibility with the "default" suffix has been removed. Any InferenceService component that is still using the "default" suffix will now have a new service created without the suffix.” https://github.com/kserve/kserve/pull/4178

> * chore: remove 'default' suffix compatibility by @sivanantha321 in https://github.com/kserve/kserve/pull/4178

### Default `OpenTelemetryCollector` resource requests and limits added

- Affects: installations using the default `OpenTelemetryCollector` configuration for `InferenceService` autoscaling (Inferred)

The defaults set both CPU request and limit to `1`, and both memory request and limit to `2Gi`.

> * add resource default for otel collector container limit in configmap by @andyi2it in https://github.com/kserve/kserve/pull/4633

### Quick-install script version pins updated

- Affects: fresh kind-cluster installations using `hack/quick_install.sh`

On fresh kind clusters, `hack/quick_install.sh -r` failed amid Istio Helm schema-validation errors, and the KServe `v0.16.0-rc0` release candidate was unavailable. The script now selects Istio `1.27.1` to address schema-validation compatibility and KServe `v0.15.2`, an available, supported release.

> * fix: istio installation failures and update kserve to stable version  by @Ce11an in https://github.com/kserve/kserve/pull/4684

### SSL disabled in LLMInferenceService presets

- Affects: LLMInferenceService deployments using the changed configuration presets

The presets previously enabled SSL with `--enable-ssl-refresh` and certificate and key arguments, and used HTTPS health probes. They now omit those SSL arguments and use HTTP health probes; the PR describes this as temporary while `BackendTLSPolicy` is not implemented.

> * Temporarily disable SSL to unblock e2e tests by @Jooho in https://github.com/kserve/kserve/pull/4731

### `LLMInferenceService` presets set `runAsNonRoot` to false

- Affects: `LLMInferenceService` deployments using the affected configuration presets (Inferred)

Before this release, the presets set `runAsNonRoot` to true.

Evidence: “+ runAsNonRoot: false” https://github.com/kserve/kserve/pull/4731

> * Temporarily disable SSL to unblock e2e tests by @Jooho in https://github.com/kserve/kserve/pull/4731

### Autoscaling metric quantities preserve their input representation

- Affects: InferenceServices using autoscaling metric targets with `value` or `averageValue`, including KEDA compatibility paths. (Inferred)

Previously, JSON serialization could normalize `value` and `averageValue`, so the serialized representation differed from what was supplied. Serialization now uses the original raw input when available while retaining a parsed `resource.Quantity` for validation.

Evidence: “// MarshalJSON implements custom JSON marshaling using the original raw value for KEDA compatibility” https://github.com/kserve/kserve/pull/4699

> * Use a wrapper struct to accept resource.Quantity and keep the original input by @houshengbo in https://github.com/kserve/kserve/pull/4699

### S3 configuration sources and empty-value handling changed

- Affects: S3 storage configured through service-account or Secret annotations, Secret data, or the `inferenceservice-config` ConfigMap

Previously, empty `S3_ENDPOINT` or `S3_USE_HTTPS` annotation values could still be set in the container environment. Now S3 values can also come from Secret data, with annotations taking precedence over Secret data and Secret data over the S3 config; empty strings from any source are omitted from the container environment.

Evidence: “Regardless of where their value is set (annotation, secret data, or s3 config) they will always be set in the container env unless the value is an empty string.” https://github.com/kserve/kserve/pull/4727

> * Add Support for Configuring S3 Storage via Secret Data by @brettmthompson in https://github.com/kserve/kserve/pull/4727

### `llmisvc-resources` chart adds inference service ConfigMap

- Affects: the `llmisvc-resources` Helm chart (Inferred)

The chart no longer includes `gateway-inference-extension.yaml`. Its new `inferenceservice.yaml` template defines the `inferenceservice-config` ConfigMap.

Evidence: “+ name: inferenceservice-config” https://github.com/kserve/kserve/pull/4729

> * Add llmd e2e tests by @andresllh in https://github.com/kserve/kserve/pull/4729

### LLMInferenceService CRDs and webhooks moved to dedicated Kustomize configuration

- Affects: Kustomize-based LLMInferenceService installations (Inferred)

The independent installation can be built with `kustomize build config/overlay/llmisvc`.

> * refactor: separate llmisvc manifests from default kustomization by @Jooho in https://github.com/kserve/kserve/pull/4763

### KServe Python drops support for `Pydantic` v1

- Affects: Python SDK setups using `Pydantic` v1 (Inferred)

Previously, both `Pydantic` v1 and v2 were supported.

Evidence: “Currently both v1 and v2 pydantic is supported, dropped support for v1 for the support overhead.” https://github.com/kserve/kserve/pull/4353

> * chore: drop pydantic v1 support by @sivanantha321 in https://github.com/kserve/kserve/pull/4353

### Python `Model` class no longer exposes `predictor_host` or `PredictorConfig`

- Affects: Python KServe SDK models using `predictor_host` or `PredictorConfig` on the `Model` class (Inferred)

Previously, the Python `Model` class exposed both attributes at model scope. `PredictorConfig` is now retrieved from context, and existing functionality is maintained without requiring it.

Evidence: “Breaking: Remove predictor_host and PredictorConfig from the Model class” https://github.com/kserve/kserve/pull/4526

> * Refactor KServe to use global context for PredictorConfig by @sivanantha321 in https://github.com/kserve/kserve/pull/4526

### Trained-model sharding `Strategy` interface removed

- Affects: Go code that imports or refers to the trained-model sharding package’s `Strategy` interface (Inferred)

The interface declared `GetOrAssignShard`.

> * Remove unused Strategy interface from sharding package by @houshengbo in https://github.com/kserve/kserve/pull/4590

### Creation blocked for standard predictors with a non-empty `name`

- Affects: InferenceServices using standard predictors with a non-empty `name` field

Evidence: “This PR adds validation to catch misconfiguration in the `name` field of standard predictor in creation time.” https://github.com/kserve/kserve/pull/4535

> * disallow `name` field in standard predictor by @HutakiHare in https://github.com/kserve/kserve/pull/4535

### Parallelism CRD formats narrowed to `int32`

- Affects: LLMInferenceService and LLMInferenceServiceConfig resources with `parallelism.pipeline` or `parallelism.tensor` values outside the `int32` range (Inferred)

Evidence: “+ format: int32” https://github.com/kserve/kserve/pull/4596

> * update CRDs for LWS based multi node support by @VedantMahabaleshwarkar in https://github.com/kserve/kserve/pull/4596

### `storage` spec removed from LLMInferenceService CRDs and API types

- Affects: LLMInferenceService or LLMInferenceServiceConfig manifests that set `storage` (Inferred)

The `LLMInferenceService` CRD previously included `storage` for compatibility with non-ModelMesh deployments.

Evidence: “The Storage Spec was added to InferenceServices because this is the format that ModelMesh supports. For non-ModelMesh deployments, KServe added support for compatibility.” https://github.com/kserve/kserve/pull/4622

> * Remove storage spec from LLMIsvc by @israel-hdez in https://github.com/kserve/kserve/pull/4622

### LLMInferenceService validation webhook added

- Affects: LLMInferenceService resources with router routes or workload parallelism configured

It checks router cross-field constraints, including disallowing custom HTTPRoute references alongside an inline route spec, and validates workload parallelism constraints.

> * feat(webhook): introduces Validating Webhook for LLMInferenceService by @VedantMahabaleshwarkar in https://github.com/kserve/kserve/pull/4631

### `LLMInferenceService` LoRA API now uses `LLMModelSpec`

- Affects: setups using the LoRA API in the `LLMInferenceService` CRD

Previously, the LoRA API used the existing `ModelSpec` type.

Evidence: “The LLMInferneceService CRD LoRA API wasn't meant to use the existing ModelSpec but the LLMModelSpec type” https://github.com/kserve/kserve/pull/4663

> * [llmisvc] Improve config merge and update well-known presets by @pierDipi in https://github.com/kserve/kserve/pull/4663

### `EnableDirectPvcVolumeMount` flag removed

- Affects: storage-initializer configurations using `EnableDirectPvcVolumeMount` (Inferred)

The flag had defaulted to true for several releases. The PR describes its legacy false path as effectively dead, although its code and tests still existed.

Evidence: “EnableDirectPvcVolumeMount has been defaulted to true for several releases, making the legacy false path effectively dead.” https://github.com/kserve/kserve/pull/4694

> * deprecate: remove EnableDirectPvcVolumeMount flag by @anurags25 in https://github.com/kserve/kserve/pull/4694

### Validating webhooks added to the `llmisvc-resources` Helm chart

- Affects: installations of the `llmisvc-resources` Helm chart managing `LLMInferenceService` or `LLMInferenceServiceConfig` resources (Inferred)

The validating webhooks cover `CREATE` and `UPDATE` operations.

> * llmisvc dev & kustomize setup, add webhook config to helm chart  by @sivanantha321 in https://github.com/kserve/kserve/pull/4712

## Deprecations (1)

### KServe deploy modes renamed to `Standard` and `KNative`

- Affects: KServe configurations that set deploy-mode values or rely on the default mode
- Removed in: not stated

The default is now `Standard`. Legacy values remain as deprecated fields and are upconverted by the config constructor.

> * refactor: 4602 - Refactor "RawDeployment" to "Standard" and "Serverless" to  "KNative" to clarify usage by @cjohannsen-cloudera in https://github.com/kserve/kserve/pull/4608

## Important fixes (28)

### HuggingFace `TextIteratorStreamer` skips special tokens

- Affects: HuggingFace backend text-generation streaming

With this functionality, streamed responses previously included special tokens such as `<|im_end|>` and `<|endoftext|>`.

> * fix: Update TextIteratorStreamer to skip special tokens by @sivanantha321 in https://github.com/kserve/kserve/pull/4490

### InferenceGraph endpoint resolution honors the ServingRuntime protocol

- Affects: InferenceGraph workloads with an InferenceService using a V2-only ServingRuntime and no `protocolVersion`

Previously, endpoint resolution returned a V1 endpoint in this case, misconfiguring the InferenceGraph router. The workload is now configured with the V2 endpoint.

> * Resolve inference endpoint using runtime protocol when applicable by @israel-hdez in https://github.com/kserve/kserve/pull/4527

### KEDA OpenTelemetry scaling metadata corrected

- Affects: InferenceServices using KEDA with an OpenTelemetry metric trigger (Inferred)

When KServe generated a KEDA OpenTelemetry metric trigger, it populated `clampMin` and `clampMax` with the InferenceService's replica bounds, even though those fields bound the scale itself. KServe-generated trigger metadata now omits both fields.

> * fix: unset clampMax and clampMin, since they are not for replicas by @houshengbo in https://github.com/kserve/kserve/pull/4556

### HTTPRoute reconciliation requeue handling fixed

- Affects: InferenceServices using HTTPRoute ingress reconciliation (Inferred)

When an HTTPRoute was missing or not ready during reconciliation, the reconciler could fail to requeue, and a not-ready route could be silently skipped. It now requeues for routes that are not ready or not found, supporting convergence to the desired state.

> * refactor: Enhance HTTPRoute readiness checks by @sivanantha321 in https://github.com/kserve/kserve/pull/4543

### OpenTelemetryCollector autoscaling includes all specified metrics

- Affects: InferenceServices with multiple OpenTelemetryCollector autoscaling metrics (Inferred)

Previously, only the last specified metric was included. Now all specified metrics are included for autoscaling.

> * Fix: Support multiple metrics in OpenTelemetryCollector for autoscaling by @houshengbo in https://github.com/kserve/kserve/pull/4591

### KEDA autoscaling isolated by namespace and deployment

- Affects: `InferenceService` deployments using KEDA/OpenTelemetry autoscaling (Inferred)

Metrics from different deployments or namespaces could be mixed, causing incorrect autoscaling. The OTelCollector now adds `namespace` and `deployment` labels, which KEDA queries use to filter metrics.

> * Fixed the issue of the same metrics across different deployments under different namespaces by @houshengbo in https://github.com/kserve/kserve/pull/4593

### Hugging Face classification probability handling fixed

- Affects: Hugging Face classification tasks using `return_probabilities=True` (Inferred)

With `return_probabilities=True`, sequence classification could return raw logits instead of probabilities, while fill-mask returned probabilities. Softmax is now applied consistently across task types.

> * feat: Add disable postprocessing option for raw logits by @sivanantha321 in https://github.com/kserve/kserve/pull/4566

### InferenceService `Stopped` condition handling corrected

- Affects: InferenceServices reconciled by the stop/resume controller (Inferred)

A missing `Stopped` condition is initialized to false. On resume, the condition is set to false only if it was previously true.

> * [Followup] Improve the stop/resume Inference Service tests and status by @hdefazio in https://github.com/kserve/kserve/pull/4636

### GPU detection recognizes NVIDIA MIG resources

- Affects: workloads with NVIDIA MIG GPU resource limits

Resource names beginning with `nvidia.com/mig` identify workloads as GPU-enabled.

> * fix: Adds support for NVIDIA MIG GPU resource detection by @sivanantha321 in https://github.com/kserve/kserve/pull/4642

### Progressing and failed states reflected in component deployment status

- Affects: KServe component deployments (Inferred)

> * Last Deployment Status Should Reflect Deployment Status by @HotsauceLee in https://github.com/kserve/kserve/pull/4667

### InferenceService model-copy status updated in all states

- Affects: InferenceServices with models in `Pending`, `Loading`, `FailedToLoad`, or `Loaded` states (Inferred)

Previously, `ModelCopies` was updated only after a model reached `Loaded`, leaving `totalCopies` out of status while it was `Pending`, `Loading`, or `FailedToLoad`. Now, `ModelCopies.TotalCopies` is stored in every model state and counts ready pods.

> * Fix: Update ModelCopies.TotalCopies for all model states  by @hardik-menger in https://github.com/kserve/kserve/pull/4676

### Inference logging prioritizes populated embedded logger specs

- Affects: inference services with inference logging configured through an embedded logger spec and a global logger ConfigMap

Previously, a populated embedded logger spec was ignored and the global ConfigMap was used instead. Now the embedded spec takes precedence, and the global configuration is used when the embedded spec is not populated.

> * 4739 - Fix: blob storage for inference logging recognizes embedded spec by @cjohannsen-cloudera in https://github.com/kserve/kserve/pull/4740

### Inference logging response metadata now includes request metadata headers

- Affects: inference logging requests with metadata headers (Inferred)

Previously, requests carrying metadata headers had those headers omitted from the logged response metadata.

> * 4778 - Fix: apply metadata headers to inference logging response by @cjohannsen-cloudera in https://github.com/kserve/kserve/pull/4779

### Configurable router timeouts for `InferenceGraph`

- Affects: `InferenceGraph` resources needing custom router or service-client timeouts

A short router server timeout could return `EOF` while graph `InferenceService`s were starting, for example during model downloads or Kubernetes node provisioning. Custom values under `InferenceGraph.spec.routerTimeouts` can avoid this; the source example sets `serverRead`, `serverWrite`, and `serverIdle` to `300` and reports a successful response.

> * Allow to set custom timeouts for `InferenceGraph` router by @lifo9 in https://github.com/kserve/kserve/pull/4218

### Standalone `AWS_CA_BUNDLE` paths honored for S3 SSL

- Affects: S3 storage with SSL verification enabled and `AWS_CA_BUNDLE` set to a local CA file (Inferred)

Previously, the path was ignored during an S3 download with SSL verification enabled if `CA_BUNDLE_CONFIGMAP_NAME` was unset.

> * fix: Allow CA bundle path without config map by @fabiendupont in https://github.com/kserve/kserve/pull/4451

### HuggingFaceServer FlashInfer build failure fixed

- Affects: HuggingFaceServer images running vLLM’s `EngineCore` with `FlashInfer`

When `EngineCore` started and runtime JIT compilation ran, the Ninja build could fail with “RuntimeError: Ninja build failed.” The image now includes precompiled sampling kernels, avoiding that runtime build and failure.

> * Add AOT flashinfer build to huggingfaceserver dockerfile to precompil… by @AyushSawant18588 in https://github.com/kserve/kserve/pull/4567

### LLMInferenceService types added to Python package exports

- Affects: Python SDK imports of `LLMInferenceService` or `LLMInferenceServiceConfig` (Inferred)

These types were missing from Python imports.

> * fix(deps): adds missing LLMInferenceService types for python imports by @bartoszmajsak in https://github.com/kserve/kserve/pull/4604

### `ClusterServingRuntime` image digest handling fixed

- Affects: `ClusterServingRuntime` resources using image digests, especially in GPU clusters

In GPU clusters, adding the `-gpu` postfix to an image digest could make the runtime fail. Digest-based image references are now left unchanged by image-tag updating.

> * fix: allow digests in runtimes' images by @tmvfb in https://github.com/kserve/kserve/pull/4653

### LLMInferenceService controller RBAC permissions expanded

- Affects: clusters running the LLMInferenceService controller (Inferred)

The added permissions support enhanced controller functionality and leader election.

> * llmisvc: fix RBAC, templating, and adds quick install script by @sivanantha321 in https://github.com/kserve/kserve/pull/4698

### Nil-pointer panic fixed in InferenceService reconciliation

- Affects: InferenceService deployment, HPA, or KEDA reconciliation (Inferred)

When `componentExt` was nil during these reconciliation paths, InferenceService reconciliation could panic; it no longer does.

> * Fixed the panic nil pointer issue componentExt being nil by @houshengbo in https://github.com/kserve/kserve/pull/4704

### llmisvc controller Dockerfile entrypoint corrected

- Affects: installations that build or run the llmisvc controller image from this Dockerfile (Inferred)

The Dockerfile previously set the entrypoint to `/llmisvc`; it now sets it to `/manager`.

> * Fix incorrect entrypoint in llmisvc Dockerfile by @Jooho in https://github.com/kserve/kserve/pull/4730

### Reduced memory use for Hugging Face downloads

- Affects: Hugging Face model downloads through the storage initializer

Large-model downloads with the latest Hugging Face Hub Python library consistently ran out of memory. Storage initializer containers now receive `HF_HUB_ENABLE_HF_TRANSFER=1`, `HF_XET_HIGH_PERFORMANCE=1`, and `HF_XET_NUM_CONCURRENT_RANGE_GETS=8`; in the reported test, memory use fell from more than 2 GB to less than 0.5 GB while download time stayed consistent.

> * Configure HF Downloads to Lower Memory Usage by @brettmthompson in https://github.com/kserve/kserve/pull/4726

### Default Kustomize controller image reference corrected

- Affects: installations using the default Kustomize manager image patch (Inferred)

The deployment failed with an image-pull back-off when it used `kserve-controller:latest`; it now uses `kserve/kserve-controller:latest`.

> * fix: wrong kserve controller image url by @Jooho in https://github.com/kserve/kserve/pull/4773

### Storage initializer duplicate environment variable fix

- Affects: storage initialization with customized storage containers defining HuggingFace optimization or CA bundle environment variables (Inferred)

Existing matching user-defined environment variables are retained without an automatically appended duplicate. Default HuggingFace optimization variables are added only when absent.

> * Fix the issue of Invalid value for env[1].valueFrom by @houshengbo in https://github.com/kserve/kserve/pull/4765

### XGBoost loading of `.ubj` and `.json` models from modelcars fixed

- Affects: XGBoost server deployments loading `.ubj` or `.json` models from modelcars

These loads failed because the model paths in modelcars are in a virtual filesystem. The server now passes file contents to XGBoost as a `bytearray` rather than having the library read the virtual path.

> * Fix XGBoost .ubj and .json models with modelcar by @balazsprehoda in https://github.com/kserve/kserve/pull/4746

### CA bundle injection preserves existing environment variables

- Affects: LLMInferenceService workloads using CA bundle injection (Inferred)

When a CA bundle ConfigMap-name or volume-mount-path environment variable already exists, injection leaves it in place rather than adding another; otherwise, it adds the configured or default value.

> * Fixed the CA bundle env var logic by @houshengbo in https://github.com/kserve/kserve/pull/4777

### HuggingFace mean pooling runtime errors fixed

- Affects: HuggingFace runtime deployments whose models use mean pooling for sentence embeddings, including CPU-only and GPU-enabled deployments

When `attention_mask` and `token_embeddings` were on different devices or had mismatched data types, mean pooling could raise runtime errors. It now aligns the mask with the embeddings' device and dtype, handles a missing mask, and clamps the denominator to prevent division by zero, supporting inference across CPU and GPU.

> * huggingfaceserver: fix mean_pooling device/dtype mismatch and add regression test by @JohnsonWang1015 in https://github.com/kserve/kserve/pull/4744

### Optional prediction-to-instance mapping for `InferenceGraph` steps

- Affects: `InferenceGraph` steps chaining V1 Inference Protocol services

When an upstream service returned output under `predictions` but the downstream service expected `instances`, straightforward chaining was prevented. With `mapPredictionsToInstances` set to `true`, the router maps upstream `predictions` to downstream `instances`; without that setting, the mapping is not enabled.

> * Allow mapping predictions to instances in sequential inference graphs by @tmbochenski in https://github.com/kserve/kserve/pull/4738

## Enhancements (43)

- Per-runtime disable settings for built-in serving runtimes (To use: Set each runtime's `disabled` setting to `true` to disable it; it defaults to `false`.)
- KServe storage module released as a separate `kserve-storage` package
- Stop and resume Raw Deployment InferenceServices (To use: Set `serving.kserve.io/stop` to `'true'` to stop a service and to `'false'` to resume it.)
- Original model source URI injected into user containers (To use: Remote storage enabled)
- `LLMInferenceService` and `LLMInferenceServiceConfig` APIs and CRDs added
- Transformer stop and resume support (To use: `serving.kserve.io/stop` annotation)
- OCI storage validation enabled for multi-node/multi-GPU InferenceServices
- Opt-in S3 storage for model request and response logs
- Auto-update opt-out for ready InferenceServices
- Explainer stop and resume support (To use: `constants.StopAnnotationKey` on the InferenceService, set to `true` to stop or `false` to resume)
- Storage initializer resource requests and limits configurable in the KServe Helm chart
- Initial `LLMInferenceService` controller scaffold and Helm charts
- Strategic merging for `LLMInferenceServiceSpec`
- KEDA external metric source authentication configuration
- `utils` collection helpers moved to `collections.go`
- Hugging Face server raw-logit output (To use: The `--return_raw_logits` option, which is supported for `sequence_classification`, `fill_mask`, and `token_classification` and is mutually exclusive with `return_probabilities`.)
- Stop and resume for serverless InferenceGraphs (To use: Set the `serving.kserve.io/stop` annotation to `"true"` to stop the graph or `"false"` to resume it.)
- LWS-based multi-node fields added to `LLMInferenceService` and `LLMInferenceServiceConfig` CRDs
- LLMInferenceServiceConfig base presets added
- llm-d scheduler reconciliation for LLM load balancing
- KEDA autoscaling behavior configuration
- Stop and resume raw-deployment-mode InferenceGraphs (To use: Set the InferenceGraph annotation `serving.kserve.io/stop: "true"` to stop it, or `serving.kserve.io/stop: "false"` to resume it.)
- Single- and multi-node `LLMInferenceServices` workload reconcilers added
- Route reconciliation for `LLMInferenceServices`
- Some `LLMInferenceService` controller logic added
- Updated `LLMInferenceServiceConfig` well-known presets
- `LLMInferenceService` configuration merging handles missing schedulers and external `InferencePool` objects
- Generic CRUD support for cluster-scoped resources and nil owners
- KServe Storage moved into a separate module
- Configurable rollout strategies for raw and `Standard` deployments (To use: Configure a component’s `deploymentStrategy` or the ConfigMap’s `deploymentRolloutStrategy` or `defaultRollout`; ConfigMap defaults apply only when `defaultDeploymentMode` is `"Standard"`.)
- Published LLMInferenceService controller image and Helm chart
- Quick-install script for LLMInferenceService dependencies and components
- Configurable OpenTelemetry and autoscaler settings in the `kserve-resources` Helm chart (To use: The `kserve-resources` Helm chart with the relevant `kserve.opentelemetryCollector` or `kserve.autoscaler` values set)
- Time-series forecasting endpoint and task integration
- LLMInferenceService Kustomize deployment configuration
- Multiple storage locations for inference artifacts (To use: Kubernetes init-container feature flag for Knative deployments)
- Cert Manager installation added to the LLMInferenceService quick install script
- CA bundles injected into LLMInferenceService S3 storage-init containers (To use: CA-bundle configuration defined in a service account, secret annotations, or the `inferenceservice-config` configmap)
- LLMInferenceService Kueue metadata propagation
- GCS and Azure inference logging storage support
- Modular setup scripts for KServe dependencies
- KServe v0.16.0 install manifests and Helm charts (To use: Helm `--version v0.16.0`)
- New Helm chart for minimal LLMInferenceService CRDs (To use: Install the `llmisvc-crd-minimal` Helm chart at version `v0.16.0` to deploy the minimal LLMInferenceService CRDs.)

## Dependency updates (8)

8 more, none mentioning security

- `* Upgrade Torch to v2.6.0 everywhere by @ashahba in https://github.com/kserve/kserve/pull/4450`
- `* upgrade vllm to v0.9.0 and Torch to v2.7.0 by @ashahba in https://github.com/kserve/kserve/pull/4501`
- `* upgrade vllm to v0.9.0 and Torch to v2.7.0 by @ashahba in https://github.com/kserve/kserve/pull/4501`
- `* Upgrade vLLM to v0.9.0.1 by @ashahba in https://github.com/kserve/kserve/pull/4507`
- `* Upgrade vllm to v0.9.2 by @AyushSawant18588 in https://github.com/kserve/kserve/pull/4586`
- `* Protobuf version bump by @karolpustelnik in https://github.com/kserve/kserve/pull/4634`
- `* Update the kserve-storage module to the latest version by @spolti in https://github.com/kserve/kserve/pull/4754`
- `* bump up istio version to 1.27.1 by @Jooho in https://github.com/kserve/kserve/pull/4775`

## Other changes (49)

- Jooho added as an `OWNERS` approver
- Test CRD file renamed to `test/crds/serving.kserve.io_all_crds.yaml`
- HuggingFace CPU dependency locks now avoid updating other dependencies
- HuggingFace request trace logging format fixed
- Code-generation script pins generator binaries to `CODEGEN_VERSION`
- BentoML sample code and manifest updated
- EnvTest CRD lookup resolves the project root from `go.mod`
- Invalid field removed from the OpenShift cert-manager operator guide manifest
- Pull request coverage checks and comparison reports added
- Python dependency installation switched from Poetry to `uv`
- Code simplified using generics
- HuggingFaceServer `PYTHONPATH` now follows `WORKSPACE_DIR`
- Missing code-coverage configuration file added
- Downloader tests switched to local `httptest` server URLs
- Missing `pytest-asyncio` session-scope marker fixed in a test
- `LLMInferenceService` injects missing GIE pool selectors and EPP service accounts
- Corrected parallelism information in the e2e test script
- Faster CI
- Repository typos fixed
- `LLMInferenceServiceConfig` template validation added
- Quickstart link fixed
- CI skips workflows for Markdown-only changes
- README links updated to the KServe website
- Snyk scan SARIF file upload fixed
- Graph test image tag defaults to `latest`
- HTML escaping in API comments fixes website documentation syntax errors
- KEDA and OpenTelemetry reconciliation accounts for API-server defaults
- KEDA pod-metric `targetValue` no longer uses fixed six-decimal formatting
- Kafka sample documentation path updated
- Repository prepared for the `0.16.0-rc0` release
- In-memory `GatewayClass` replaced in test setup
- Role-attachable service accounts for `LLMInferenceService` workloads
- LLMInferenceService router reconciliation status update
- Autoscaling test duration and crashes fixed
- LLMInferenceService credential reconciliation no longer logs missing-account errors
- LLMInferenceService controller image now includes `manager`
- LLMInferenceService Helm templates refactored for compatibility
- Disk-space cleanup in Docker publishing workflows
- Star History section added to the `README`
- Project README now mentions CNCF
- KServe `README` overview and features section revised
- Controller image build and push commands use `KO_DOCKER_REPO`
- End-to-end tests and GitHub Actions workflow for `LLMInferenceService`
- Preparation for the 0.16.0-rc1 release
- Model sidecar UID setting in the `kserve-resources` Helm chart
- Python test workflow recreates a broken cached virtual environment
- Raw-controller tests refactored
- Updated `LLMInferenceService` sample configurations and README
- KServe Storage PyPI publishing workflow uses `KSERVE_STORAGE_PYPI_TOKEN`

---

Summaries are produced by AI, so they can be wrong, and they are provided without warranty.
