Skip to content
QuenchWorks

vllm

Chart · Machine learning & AI · standard · v0.0.16

digest pinnedcosign signedSPDX SBOMSLSA provenanceamd64 · arm64Rebuilt 2026-10-22

vLLM, the high-throughput LLM inference and serving engine with an OpenAI-compatible API. CPU build — installs vLLM's prebuilt CPU wheel (torch+cpu, no source compile, no CUDA) into a venv on a hardened Wolfi python-3.12 base, 0-CVE across ~140 Python packages. The user supplies the model; the chart runs it as a StatefulSet with a persistent model cache.

Version

0.0.16 · latest0.0.150.0.14

The latest line lives at the base page; older lines have their own page so you can pin and verify exactly that version. Only the most recent releases are listed — an older chart pins an older image digest.

Deployed image digest

sha256:229f56ef3e552a58412f4bf447ec8972c742f195d4188ecd4316a7f2323705f0

Chart OCI version

oci://ghcr.io/quenchworks/charts/vllm:0.0.16

The chart pins its image by this signed digest, so you never track it yourself. Signatures, SBOM, and provenance attach to the same digest.

Signed
cosign keyless
SBOM
SPDX, on image
Provenance
SLSA build
Architectures
amd64, arm64
Runs as
nonroot (uid 1001)
Root filesystem
read-only
Image size
1002.6 MB

Security report (Trivy)

D· 0/10017 fixable · rebuild clears them

Vulnerability detail

vllm 0.30.0 · 17 CVE
CVESeverityPackageInstalledFixed inTitle
CVE-2026-114053HIGHvllm0.30.0+cpu0.31.0vLLM: Request-controlled mm_processor_kwargs.code_revision allows remote code execution
CVE-2026-114055HIGHvllm0.30.0+cpu0.31.0vllm: vLLM: Denial of Service via unvalidated multimodal processor arguments
CVE-2026-115912HIGHvllm0.30.0+cpu0.32.0vllm: vllm: Arbitrary code execution via forged shared-memory handles
CVE-2026-115915HIGHvllm0.30.0+cpu0.32.0vllm: vllm: Arbitrary code execution via unrestricted image format parsing
CVE-2026-114050MEDIUMvllm0.30.0+cpu0.31.0vllm: vllm: Denial of Service via multimodal cache desynchronization
CVE-2026-115911MEDIUMvllm0.30.0+cpu0.32.0vllm: vLLM: Denial of Service via unbounded chat template rendering
CVE-2026-115913MEDIUMvllm0.30.0+cpu0.32.0vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
CVE-2026-115914MEDIUMvllm0.30.0+cpu0.32.0vllm: vllm: Denial of service via crafted image input in OCR models
CVE-2026-115916MEDIUMvllm0.30.0+cpu0.32.0vllm: vLLM: Denial of Service via out-of-bounds prompt token ID
CVE-2026-115918MEDIUMvllm0.30.0+cpu0.32.0vllm: vLLM: Denial of service via untrusted chat template processing
CVE-2026-94624MEDIUMvllm0.30.0+cpu0.32.0vllm: vLLM: Denial of Service via unbounded P2P KV offloading sessions
GHSA-c8m5-c952-6cm4MEDIUMrustls0.23.450.23.46, 0.24.0-dev.2Rustls: Session Debug formatting exposes TLS master secrets and resumption PSKs
CVE-2026-115917LOWvllm0.30.0+cpu0.32.0vllm: vLLM: Information disclosure via permissive CORS configuration
GHSA-h2v8-394m-p7vvLOWrustls0.23.450.23.46, 0.24.0-dev.2Rustls: Brotli certificate decompression can allocate an attacker-controlled amount of memory up to 1 GiB before authentication.
GHSA-vp88-798v-h67fLOWrustls0.23.450.23.46, 0.24.0-dev.2rustls: ECH rejection can expose buffered application data to authenticated outer server before `ech_required` alert
GHSA-w44r-93j7-qv28LOWrustls0.23.450.23.46, 0.24.0-dev.2rustls: An unauthenticated ServerHello poisons the per-name key-exchange record, so later ClientHellos carry only a classical X25519 share
GHSA-xmj3-9gf7-h52wLOWsetuptools83.0.084.1.0setuptools: Windows path escape via namespace_packages.txt during legacy wheel installation Description
0
Critical
4
High
8
Medium
5
Low
0
Unknown

Security report (Trivy) · image vllm 0.30.0

Install the chart

Deploy to Kubernetes with hardened defaults. The chart pins its image by signed digest, so you never track it yourself.

Install (latest)

helm install my-vllm oci://ghcr.io/quenchworks/charts/vllm --version 0.0.16

Deploys image (digest-pinned)

Chart version
0.0.16
App version
0.30.0
Chart license
Apache-2.0
App license
Apache-2.0
Service port
8000
Signed
cosign (keyless)
Values schema
yes
Last published
2026-10-26
Values keys
  • image: repository digest pullPolicy
  • nameOverride
  • model
  • extraArgs
  • resources: requests limits
  • persistence: enabled size accessModes annotations selector existingClaim
  • service: type port
  • podLabels
  • podAnnotations
  • nodeSelector
  • affinity
  • tolerations
  • topologySpreadConstraints
  • priorityClassName
  • schedulerName
  • terminationGracePeriodSeconds
  • updateStrategy
  • extraEnvVars
  • extraEnvVarsCM
  • extraEnvVarsSecret
  • extraVolumes
  • extraVolumeMounts
  • initContainers
  • sidecars
  • lifecycleHooks
  • command
  • args
  • podSecurityContext
  • containerSecurityContext
  • livenessProbe
  • readinessProbe
  • startupProbe
  • customLivenessProbe
  • customReadinessProbe
  • customStartupProbe
  • serviceAccount: create name annotations
  • rbac: create
  • networkPolicy: enabled allowExternal
  • podDisruptionBudget: enabled minAvailable
  • ingress: enabled className annotations servicePort hosts tls

Verify the chart

cosign verify ghcr.io/quenchworks/charts/vllm:0.0.16 \
  --certificate-identity-regexp 'https://github.com/quenchworks/.+' \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com

Transparency

The chart publishes its attestations on GitHub and the image it deploys carries its own on the same digest, publicly verifiable with the commands above. Both log to the Sigstore transparency log (Rekor), which cosign verify checks for you.

View the imageView on ArtifactHubChart source

Upstream project: https://docs.vllm.ai