Model Caching

Model Caching#

Model caching pre-downloads model artifacts to shared persistent volumes, reducing startup time and bandwidth usage across service replicas and restarts.

v1alpha2

Examples on this page use aim.eai.amd.com/v1alpha2. Caching is keyed by the resolved profile (via AIMProfileCache) instead of the v1alpha1 template (AIMTemplateCache). The spec.caching.mode field and its semantics are unchanged across versions. For the legacy template-based cache, see Legacy AIMService.

Caching Modes#

Control caching behavior with spec.caching.mode:

Mode

Behavior

Shared

Reuses shared cache assets across services that resolve to the same profile. This is the default.

Dedicated

Creates service-owned dedicated cache assets, isolated from other services.

 
apiVersion: aim.eai.amd.com/v1alpha2
kind: AIMService
metadata:
  name: qwen-chat
  annotations:
    aim.eai.amd.com/reconciler-pipeline: profile
spec:
  model:
    name: qwen-qwen3-32b
  caching:
    mode: Shared

Note

The caching mode is immutable after creation. Legacy values Always, Auto, and Never are accepted for backward compatibility (Always/Auto map to Shared, Never maps to Dedicated).

How Caching Works#

When caching is active, AIM Engine creates a hierarchy of resources:

  1. AIMProfileCache (v1alpha2) — Groups all model artifacts for a specific resolved profile on a shared PVC. (v1alpha1 used AIMTemplateCache, keyed by template.)

  2. AIMArtifact — Manages the download of individual model sources

  3. PVC + Download Job — The actual storage and download execution

When AIMProfile.spec.caching.enabled is set, the shared profile cache is owned by the profile, so multiple services using that profile reuse it. A dedicated cache is owned by its service and is garbage-collected with the service.

A profile-owned cache resolves the namespace or cluster RuntimeConfig named default; AIMProfile.spec.caching has no runtimeConfigName selector. Consequently, selecting a different RuntimeConfig on an AIMService does not retarget that shared cache. Service-owned caches use the service’s selected RuntimeConfig.

Download Protocols#

For HuggingFace models (hf:// sources), AIM Engine tries download protocols in sequence:

Protocol

Description

XET

XetHub protocol (fastest for large models)

HF_TRANSFER

HuggingFace Transfer (optimized multi-part download)

HTTP

Standard HTTP download (slowest, most compatible)

The default order is XET,HF_TRANSFER. If a protocol fails, the downloader cleans up incomplete files and tries the next one.

Override the protocol order via runtime configuration:

apiVersion: aim.eai.amd.com/v1alpha1
kind: AIMRuntimeConfig
metadata:
  name: default
  namespace: ml-team
spec:
  env:
    - name: AIM_DOWNLOADER_PROTOCOL
      value: "HF_TRANSFER,HTTP"

Storage Sizing#

AIM Engine automatically sizes PVCs based on discovered model sizes plus a headroom percentage (default 10%). Configure the headroom via cluster runtime config:

apiVersion: aim.eai.amd.com/v1alpha1
kind: AIMClusterRuntimeConfig
metadata:
  name: default
spec:
  storage:
    pvcHeadroomPercent: 15
    defaultStorageClassName: longhorn

Repositories with required subdirectories#

When no download filter is configured, AIM Engine excludes files in subdirectories (equivalent to exclude: ["*/*"]). This is suitable for many flat model repositories, but repository layouts such as Diffusers commonly store required transformer, VAE, text-encoder, scheduler, and tokenizer files under nested directories.

Set an explicitly empty filter to download all files:

apiVersion: aim.eai.amd.com/v1alpha1
kind: AIMRuntimeConfig
metadata:
  name: diffusers-model
  namespace: ml-team
spec:
  storage:
    downloadFilter: {}

For a service-owned cache, reference that RuntimeConfig from the relevant AIMService:

spec:
  runtimeConfigName: diffusers-model

For a profile-owned shared cache, configure the download filter on the namespace or cluster RuntimeConfig named default instead. Avoid changing the cluster-wide default unless every affected model needs nested files: downloading them can increase transfer time and storage usage for unrelated models.

This setting is based on repository layout, not inference engine. For example, a Diffusers model served by vLLM-Omni needs it, while a flat repository served by the same engine may not.

AIMArtifact.spec.downloadFilter is immutable. If an artifact was created with the wrong filter, delete and recreate that artifact after changing the runtime configuration.

Download Verification#

After each download completes, AIM Engine automatically verifies that all expected files are present on disk and persisted to storage. If verification fails, the download job retries with a clean state. No configuration is required — verification runs by default for all HuggingFace downloads.

For details on how verification works, see Download Verification.

Monitoring Cache Status#

Check the status of profile caches and artifacts:

# List profile caches (v1alpha2)
kubectl get aimprofilecache -n <namespace>

# List artifacts
kubectl get aimartifact -n <namespace>

# Check artifact download progress
kubectl get aimartifact <name> -n <namespace> -o jsonpath='{.status}' | jq

For services still managed by the v1alpha1 template pipeline, list caches with kubectl get aimtemplatecache -n <namespace> instead.

Next Steps#

Contents