Model Caching

Large language models can take minutes to download. Without caching, every pod downloads the full model independently, wasting bandwidth and delaying startup. Dynamo supports a simple shared-storage path and a ModelExpress path for faster weight distribution across larger clusters.

Option 1: PVC + Download Job (Recommended)

The simplest approach: create a shared PVC, run a one-time Job to download the model, then mount the PVC in your DynamoGraphDeployment.

This is the pattern used by all Dynamo recipes today.

Step 1: Create a Shared PVC

1
apiVersion: v1
2
kind: PersistentVolumeClaim
3
metadata:
4
  name: model-cache
5
spec:
6
  accessModes:
7
    - ReadWriteMany
8
  resources:
9
    requests:
10
      storage: 100Gi

Step 2: Download the model

1
apiVersion: batch/v1
2
kind: Job
3
metadata:
4
  name: model-download
5
spec:
6
  template:
7
    spec:
8
      restartPolicy: Never
9
      containers:
10
        - name: downloader
11
          image: python:3.12-slim
12
          command: ["sh", "-c"]
13
          args:
14
            - |
15
              pip install huggingface_hub hf_transfer
16
              HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download \
17
                $MODEL_NAME --revision $MODEL_REVISION
18
          env:
19
            - name: MODEL_NAME
20
              value: "Qwen/Qwen3-0.6B"
21
            - name: MODEL_REVISION
22
              value: "main"
23
            - name: HF_HOME
24
              value: /cache/huggingface
25
          envFrom:
26
            - secretRef:
27
                name: hf-token-secret
28
          volumeMounts:
29
            - name: model-cache
30
              mountPath: /cache/huggingface
31
      volumes:
32
        - name: model-cache
33
          persistentVolumeClaim:
34
            claimName: model-cache

Find the Snapshot Path

After the Job completes, the model is stored in HuggingFace’s cache layout:

hub/models--<org>--<model>/snapshots/<commit-hash>/

For example, meta-llama/Llama-3.1-70B-Instruct becomes:

hub/models--meta-llama--Llama-3.1-70B-Instruct/snapshots/9d3b8e0f71f8c1e0f9b7c2a3d4e5f6a7b8c9d0e1/

To find the exact commit hash after the download Job completes:

$
kubectl run find-snapshot --rm -it --image=busybox --restart=Never \
>
  --overrides='{
>
    "spec": {
>
      "volumes": [{"name": "c", "persistentVolumeClaim": {"claimName": "model-cache"}}],
>
      "containers": [{
>
        "name": "f", "image": "busybox",
>
        "command": ["find", "/c/hub", "-mindepth", "3", "-maxdepth", "3", "-type", "d"],
>
        "volumeMounts": [{"name": "c", "mountPath": "/c"}]
>
      }]
>
    }
>
  }'

Alternatively, look up the commit hash on the HuggingFace Hub model page under Files and versions.

You need this path for the pvcModelPath field in a DGDR spec (see Model Deployment Guide — Model Caching).

Step 3: Mount in DynamoGraphDeployment

1
apiVersion: nvidia.com/v1alpha1
2
kind: DynamoGraphDeployment
3
metadata:
4
  name: my-deployment
5
spec:
6
  pvcs:
7
    - create: false
8
      name: model-cache
9
  services:
10
    VllmWorker:
11
      volumeMounts:
12
        - name: model-cache
13
          mountPoint: /home/dynamo/.cache/huggingface

All VllmWorker pods that mount model-cache now read from the shared cache, avoiding per-pod worker downloads. If you also want the frontend to reuse tokenizer and config files, mount the same PVC there too.

Compilation Cache

For vLLM, you can also cache compiled artifacts (CUDA graphs, etc.) with a second PVC:

1
spec:
2
  pvcs:
3
    - create: false
4
      name: model-cache
5
    - create: false
6
      name: compilation-cache
7
  services:
8
    VllmWorker:
9
      volumeMounts:
10
        - name: model-cache
11
          mountPoint: /home/dynamo/.cache/huggingface
12
        - name: compilation-cache
13
          mountPoint: /home/dynamo/.cache/vllm

Option 2: ModelExpress (P2P Distribution)

ModelExpress is a model weight distribution service that integrates with vLLM’s weight loading pipeline. It can publish model weights from one worker and let later workers pull those tensors from GPU memory over NIXL/RDMA instead of repeating a full storage download.

ModelExpress can also use ModelStreamer as a loading strategy. ModelStreamer streams safetensors directly from object storage or a local filesystem path into GPU memory through the runai-model-streamer package. In that setup, the first worker can stream from storage and then publish ModelExpress metadata so later workers can use the P2P path.

Use this path when startup time or fleet-wide model rollout time matters more than the simplicity of a shared PVC.

How It Works

1. A ModelExpress server runs in the cluster and stores metadata for available sources.
2. vLLM workers use the ModelExpress loader (--load-format mx on newer ModelExpress images, or mx-source / mx-target on older split-loader images).
3. If a compatible source is already serving the model, a new worker pulls model tensors from that source over NIXL/RDMA.
4. If no source is available, the worker falls back to storage. With a shared filesystem (RWX PVC, NFS, hostPath), the worker reads directly from the server’s cache. Without a shared filesystem, set MODEL_EXPRESS_NO_SHARED_STORAGE=1 so the client streams files from the server over gRPC; see Streaming Without Shared Storage below. When MX_MODEL_URI is set, ModelStreamer can stream safetensors from S3, GCS, Azure Blob Storage, or a local path.
5. The Kubernetes operator can inject MODEL_EXPRESS_URL into all Dynamo pods from the platform modelExpressURL setting.

What To Configure

Setup

Install with Dynamo Platform:

$
helm install dynamo-platform dynamo-platform-${RELEASE_VERSION}.tgz \
>
  --namespace ${NAMESPACE} \
>
  --set "dynamo-operator.modelExpressURL=http://model-express-server.model-express.svc.cluster.local:8080"

Configure workers to use ModelExpress:

1
services:
2
  VllmWorker:
3
    extraPodSpec:
4
      mainContainer:
5
        image: <vllm-runtime-image-with-modelexpress>
6
        command: ["python3", "-m", "dynamo.vllm"]
7
        args:
8
          - --model
9
          - meta-llama/Llama-3.1-70B-Instruct
10
          - --load-format
11
          - mx
12
          - --model-express-url
13
          - http://model-express-server.model-express.svc.cluster.local:8080
14
        env:
15
          - name: VLLM_PLUGINS
16
            value: modelexpress

When MODEL_EXPRESS_URL is configured in the operator, it is automatically injected as an environment variable into all component pods. Passing --model-express-url explicitly is still useful in examples because the worker validates that a server URL is available when using the older mx-source / mx-target load formats.

Streaming Without Shared Storage

If the ModelExpress server’s cache is on a non-shared volume (e.g. a ReadWriteOnce PVC, a cross-namespace deployment, or any topology where worker pods cannot mount the same filesystem as the server), the default shared-storage mode fails: the server reports the model as downloaded and returns its own local path, the worker cannot read that path from inside its own pod, and the load silently falls back to a direct HuggingFace download — defeating the point of running ModelExpress.

Set MODEL_EXPRESS_NO_SHARED_STORAGE=1 on every worker pod to switch the ModelExpress client into gRPC streaming mode. The server then sends model files to the client over the existing gRPC channel and the worker writes them to its own local cache.

1
services:
2
  VllmWorker:
3
    extraPodSpec:
4
      mainContainer:
5
        image: <vllm-runtime-image-with-modelexpress>
6
        command: ["python3", "-m", "dynamo.vllm"]
7
        args:
8
          - --model
9
          - meta-llama/Llama-3.1-70B-Instruct
10
          - --load-format
11
          - mx
12
        env:
13
          - name: VLLM_PLUGINS
14
            value: modelexpress
15
          - name: MODEL_EXPRESS_NO_SHARED_STORAGE
16
            value: "1"

MODEL_EXPRESS_URL is injected automatically by the operator (dynamo-operator.modelExpressURL); you do not need to set it explicitly here. No volume mount for the ModelExpress cache is required on worker pods in this mode.

Use this path when:

• The server runs with an RWO PVC, or in a different namespace from the workers.
• The cluster has no RDMA / InfiniBand fabric available, so P2P over NIXL is not an option.
• You want ModelExpress to act as a centralized download-and-cache server (one HuggingFace pull, fan out over gRPC to many workers) without standing up object storage and MX_MODEL_URI.

Shared-filesystem mode is still faster when available, so prefer an RWX PVC mounted on both the server and the workers when the storage class supports it. See the ModelExpress storage access modes documentation for the full trade-off and tuning knobs (chunk size, etc.).

ModelStreamer From Object Storage

Set MX_MODEL_URI when the first worker should stream safetensors directly from storage instead of reading a PVC or relying on a prior source worker.

1
services:
2
  VllmWorker:
3
    extraPodSpec:
4
      mainContainer:
5
        image: <vllm-runtime-image-with-modelexpress-and-modelstreamer>
6
        command: ["python3", "-m", "dynamo.vllm"]
7
        args:
8
          - --model
9
          - meta-llama/Llama-3.1-70B-Instruct
10
          - --load-format
11
          - mx
12
          - --model-express-url
13
          - http://model-express-server.model-express.svc.cluster.local:8080
14
        env:
15
          - name: VLLM_PLUGINS
16
            value: modelexpress
17
          - name: MX_MODEL_URI
18
            value: s3://my-model-bucket/meta-llama/Llama-3.1-70B-Instruct
19
          - name: RUNAI_STREAMER_CONCURRENCY
20
            value: "8"

ModelStreamer relies on the underlying cloud SDK credentials:

Credentials are consumed by the storage SDKs in the worker pod. They do not flow through the ModelExpress server.

Relationship To Shadow Engine Failover

ModelExpress and ModelStreamer are model loading and distribution paths. They are not required for Shadow Engine Failover, and enabling them does not create standby engines.

Use Shadow Engine Failover only when you specifically need an active/shadow recovery topology backed by GPU Memory Service (GMS), DRA, and a backend load format such as --load-format gms. Keep the ModelExpress / ModelStreamer configuration separate unless you have validated a combined workflow for your runtime image and cluster.

When to Use ModelExpress

See Also

• Managing Models with DynamoModel — declarative model management CRD
• Detailed Installation Guide — Helm chart configuration including ModelExpress
• Shadow Engine Failover — GMS-backed active/shadow engine recovery, separate from model distribution
• ModelExpress deployment guide — server, P2P, and ModelStreamer configuration
• LoRA Adapters — dynamic adapter loading (separate from base model caching)