AMD Inference Microservice deployment guide

AMD Inference Microservice deployment guide#

This guide gives step-by-step instructions to deploy the AMD Inference Microservice (AIM) container for the monai/wholeBrainSeg_Large_UNEST_segmentation model in different environments. Follow these instructions to run the model on AMD accelerators. This guide assumes a MI300X accelerator on the target system.

This AIM has an API of its own, shaped by what the model takes and returns. See Testing your deployment for the endpoints and payloads of this model.

Prerequisites#

  • AMD Instinct™ GPU with ROCm support (e.g., MI300X, MI325X)

  • Docker installed and configured with GPU support

  • A host directory for the model cache, if the model downloads its weights at first start

1. Docker deployment#

1.1 Running the container#

Set AIM_PROFILE_ID to one of the profiles listed in the README. The runtime then uses that profile and does not select a profile from the detected hardware.

mkdir -p /path/to/model-cache

docker run \
  -e AIM_ID=monai/wholeBrainSeg_Large_UNEST_segmentation \
  -e AIM_PROFILE_ID=bentoml-mi300x-fp16-tp1-latency \
  -v /path/to/model-cache:/workspace/model-cache \
  --device=/dev/kfd --device=/dev/dri \
  --group-add video --ipc=host \
  -p 8000:8000 \
  amdenterpriseai/aim-instinct-model-monai-wholebrainseg-large-unest-segmentation:2026.14.0

1.2 Customizing deployment with environment variables#

Customize your deployment with optional environment variables:

docker run \
  -e AIM_ID=monai/wholeBrainSeg_Large_UNEST_segmentation \
  -e AIM_PROFILE_ID=bentoml-mi300x-fp16-tp1-latency \
  -e AIM_CACHE_PATH=/workspace/model-cache \
  -e AIM_LOG_LEVEL=INFO \
  -e AIM_PORT=8080 \
  -v /path/to/model-cache:/workspace/model-cache \
  --device=/dev/kfd --device=/dev/dri \
  --group-add video --ipc=host \
  -p 8080:8080 \
  amdenterpriseai/aim-instinct-model-monai-wholebrainseg-large-unest-segmentation:2026.14.0

Without AIM_PROFILE_ID, the runtime selects a profile from the detected hardware. AIM_ACCELERATOR_COUNT changes the accelerator count that the selection uses. The runtime does not use it when AIM_PROFILE_ID is set.

2. Model caching for production#

This AIM has no separate model download step. If the model downloads its weights, it does so at service startup and keeps them in $AIM_CACHE_PATH, which defaults to /workspace/model-cache.

Mount a persistent volume at that path, as shown in section 1.1. Without the mount, the container downloads the weights again at each start.

If the cache must be at a different path, set AIM_CACHE_PATH and mount the volume at the same path:

docker run \
  -e AIM_ID=monai/wholeBrainSeg_Large_UNEST_segmentation \
  -e AIM_PROFILE_ID=bentoml-mi300x-fp16-tp1-latency \
  -e AIM_CACHE_PATH=/mnt/aim-cache \
  -v /path/to/model-cache:/mnt/aim-cache \
  --device=/dev/kfd --device=/dev/dri \
  --group-add video --ipc=host \
  -p 8000:8000 \
  amdenterpriseai/aim-instinct-model-monai-wholebrainseg-large-unest-segmentation:2026.14.0

The first start is slower than later starts, because it downloads the weights. Size the volume for the model checkpoint, and keep it writable.

3. Kubernetes deployment#

3.1 Deployment#

It is possible to deploy AIM with Kubernetes. This section gives a sample Kubernetes deployment manifest.

Create deployment.yaml with the following content:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: minimal-aim-deployment
  labels:
    app: minimal-aim-deployment
spec:
  progressDeadlineSeconds: 3600
  replicas: 1
  selector:
    matchLabels:
      app: minimal-aim-deployment
  template:
    metadata:
      labels:
        app: minimal-aim-deployment
    spec:
      containers:
        - name: minimal-aim-deployment
          image: amdenterpriseai/aim-instinct-model-monai-wholebrainseg-large-unest-segmentation:2026.14.0
          imagePullPolicy: Always
          env:
            - name: AIM_ID
              value: "monai/wholeBrainSeg_Large_UNEST_segmentation"
            - name: AIM_PROFILE_ID
              value: "bentoml-mi300x-fp16-tp1-latency"
            - name: AIM_CACHE_PATH
              value: "/workspace/model-cache"
            - name: AIM_LOG_LEVEL_ROOT
              value: "INFO"
            - name: AIM_LOG_LEVEL
              value: "INFO"
            - name: AIM_PORT
              value: "8000"
          ports:
            - name: http
              containerPort: 8000
          resources:
            requests:
              memory: "64Gi"
              cpu: "8"
              amd.com/gpu: "1"
            limits:
              memory: "64Gi"
              cpu: "8"
              amd.com/gpu: "1"
          startupProbe:
            httpGet:
              path: /healthz
              port: http
            periodSeconds: 10
            timeoutSeconds: 5
            failureThreshold: 360
          readinessProbe:
            httpGet:
              path: /readyz
              port: http
            periodSeconds: 10
            timeoutSeconds: 5
            failureThreshold: 3
          livenessProbe:
            httpGet:
              path: /healthz
              port: http
            periodSeconds: 10
            timeoutSeconds: 5
            failureThreshold: 30
          volumeMounts:
            - name: model-cache
              mountPath: /workspace/model-cache
            - name: ephemeral-storage
              mountPath: /tmp
            - name: dshm
              mountPath: /dev/shm
      volumes:
        - name: model-cache
          persistentVolumeClaim:
            claimName: minimal-aim-model-cache
        - name: ephemeral-storage
          emptyDir:
            sizeLimit: 256Gi
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 32Gi

The startup probe gives the start up to one hour: 360 checks at 10-second intervals. That covers a first start that downloads weights, as described in section 2. The pod stays out of the service endpoints until this probe succeeds, so no request reaches a model that is still loading. A later start reads the cache and succeeds much sooner.

The readiness probe checks /readyz every 10 seconds. After three failed checks (30 seconds) Kubernetes takes the pod out of the service endpoints, so a worker that fails after a successful start stops receiving requests until it recovers. The liveness probe below is slower on purpose and only restarts the container.

Kubernetes runs the liveness probe only after the startup probe succeeds. The liveness probe then accepts 30 failed checks, which is five minutes, before it restarts the container. This tolerates a model that does more work after the health endpoint answers.

The memory and cpu values above are a starting point. Raise them if the profile uses more than one accelerator, because this AIM starts one worker process per accelerator, and each worker holds its own copy of the model.

Create the cache volume claim in pvc.yaml:

apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: minimal-aim-model-cache
spec:
  accessModes:
    - ReadWriteOnce
  resources:
    requests:
      storage: 100Gi

Create a service configuration in the service.yaml file:

apiVersion: v1
kind: Service
metadata:
  name: minimal-aim-deployment
  labels:
    app: minimal-aim-deployment
spec:
  type: ClusterIP
  ports:
    - name: http
      port: 80
      targetPort: 8000
  selector:
    app: minimal-aim-deployment

Customize the deployment further to fit your needs. Apply it to the Kubernetes cluster with kubectl:

kubectl apply -f . -n <YOUR_K8S_NAMESPACE>

4. Testing your deployment#

To test the deployment, execute an API call. First find the pod, then set up port forwarding.

kubectl port-forward service/minimal-aim-deployment 8000:80 -n <YOUR_K8S_NAMESPACE>

After port forwarding starts, the API is available locally at http://localhost:8000/.

4.1 Health check#

curl http://localhost:8000/healthz

The service answers with HTTP 200 when it is ready to accept requests. The startup probe and the liveness probe in section 3.1 use it.

/healthz answers a GET. An AIM can also give an endpoint that reports the state in more detail. That endpoint takes a POST, and the section below names it if this AIM has one.

4.2 The endpoints#

Method

Path

Purpose

POST

/v1/inference

Segment one brain MRI volume.

POST

/v1/models

Report the served model and its label map.

GET

/healthz

Check whether the service is ready to accept requests.

/v1/inference and /v1/models accept POST only. /healthz accepts GET; see section 4.1.

4.3 Request body#

/v1/inference takes three top-level fields:

Field

Type

Default

Purpose

model

string

the served model

Model identifier.

input

object

required

Holds the input and output paths. See below.

parameters

object

empty

Returned unchanged in the response metadata.

input holds:

Field

Type

Default

Purpose

image_path

string

required

Path of the input NIfTI volume, inside the container.

output_dir

string

$AIM_OUTPUT_PATH (default /workspace/output)

Directory for the segmentation mask, inside the container.

/v1/models takes an empty body. /healthz needs no request body.

The service reads and writes files, not request bodies. image_path must exist in the container filesystem, and the response returns the path of the mask and not the mask itself. Mount your data into the container, as shown in section 4.4.

4.4 Example request#

The service reads and writes the paths in the request inside the container, so the volumes must be reachable there. Add -v /path/to/mri-data:/workspace/data for the input and -v /path/to/output:/workspace/output for the masks to the docker run command of section 1.1. Without the output mount, the masks stay in the container and are lost when it is removed. In the Kubernetes deployment of section 3, add both volumes to the existing volumeMounts and volumes lists, and create the two claims like the cache claim in pvc.yaml:

volumeMounts:
  - name: input-data
    mountPath: /workspace/data
  - name: output-data
    mountPath: /workspace/output
volumes:
  - name: input-data
    persistentVolumeClaim:
      claimName: minimal-aim-input-data
  - name: output-data
    persistentVolumeClaim:
      claimName: minimal-aim-output-data

The examples below use /workspace/data for input and /workspace/output for output, which are the defaults. AIM_INPUT_PATH and AIM_OUTPUT_PATH change them; a path outside those roots is rejected.

The file name in the examples stands for a volume of your own. The mni305 part of it is a reminder: the service accepts only a volume that is already registered to the MNI305 atlas, as section 4.6 describes.

Call the endpoint:

curl http://localhost:8000/v1/inference \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "image_path": "/workspace/data/subject_001_t1_mni305.nii.gz",
      "output_dir": "/workspace/output"
    }
  }'

The same request in Python:

import requests

response = requests.post(
    "http://localhost:8000/v1/inference",
    json={"input": {"image_path": "/workspace/data/subject_001_t1_mni305.nii.gz", "output_dir": "/workspace/output"}},
    timeout=1800,
)

print(response.json()["output"]["output_path"])

4.5 Response#

{
  "model": "monai/wholeBrainSeg_Large_UNEST_segmentation",
  "output": {
    "output_path": "/workspace/output/subject_001_t1_mni305/subject_001_t1_mni305_trans.nii.gz"
  },
  "metadata": {
    "device": "cuda:0",
    "used_amp": true,
    "parameters": {},
    "optimizations": {
      "compile": true,
      "compile_mode": "max-autotune",
      "dynamic": false,
      "autocast": true,
      "gpu_transforms": true
    }
  }
}

The mask is written to <output_dir>/<stem>/<stem>_trans.nii.gz, where <stem> is the input filename without its extensions. Read the file from the volume mounted at /workspace/output, as described in section 4.4. A later request for the same input writes this same path and overwrites the earlier mask. To keep both, give each request its own output_dir under /workspace/output.

/v1/models answers with the served model and the label map. The map has 133 entries: the background and 132 brain regions.

/healthz returns HTTP 200 when the service is ready. It does not return the model details or optimization settings.

4.6 What to know before you deploy#

The input must already be registered to the MNI305 atlas. The service does no resampling, no reorientation and no resizing. It normalizes intensity and nothing else. Register your volume with NiftyReg, ANTs or an equivalent tool before you send it. If the volume is not registered, the mask is wrong.

The model expects a T1-weighted brain MRI. Other modalities give meaningless output.

The service always listens on port 8000. The image sets BENTOML_PORT=8000, and the service does not read AIM_PORT, so the AIM_PORT=8080 example in section 1.2 does not work for this AIM: nothing listens on port 8080. Leave AIM_PORT unset and map the host port you want to 8000, for example -p 8080:8000. In Kubernetes, keep containerPort and targetPort at 8000, as in section 3.

Requests are served one at a time. The service holds a lock around inference, so a second request waits for the first to finish. The profile uses one accelerator.

The request timeout is 1800 seconds. Set a client timeout that is not shorter, or the client stops before the service replies. BRAINSEG_TIMEOUT_SECONDS changes it. When compile is on, as in the bentoml-mi300x-fp16-tp1-latency profile (see section 4.7), the network compiles on the first request and not at startup. That request takes much longer than later ones, so a client timeout sized for later requests can stop it.

The weights are fetched at startup. They are not baked into the image. They are downloaded from MONAI/wholeBrainSeg_Large_UNEST_segmentation on Hugging Face and kept in $AIM_CACHE_PATH. Mount a persistent volume there, as described in section 2, or they are downloaded again on every start.

This model does not need a Hugging Face token.

4.7 Optimization environment variables#

A profile can set these variables in its env_vars. A value that the profile sets overrides the code default below and any -e. The bentoml-mi300x-fp16-tp1-latency profile sets BRAINSEG_COMPILE, BRAINSEG_AUTOCAST and BRAINSEG_GPU_TRANSFORMS to true, which increases the startup time and decreases the inference time. The example response in section 4.5 comes from that profile. To see what another profile sets, run the dry-run command of section 6.3 with AIM_PROFILE_ID set, and read its env_vars.

Variable

Code default

Effect

BRAINSEG_COMPILE

false

Compile the network with torch.compile.

BRAINSEG_COMPILE_MODE

max-autotune

Compile mode used when BRAINSEG_COMPILE is on.

BRAINSEG_DYNAMIC

false

Allow dynamic shapes in the compiled network.

BRAINSEG_AUTOCAST

false

Run inference in mixed precision.

BRAINSEG_GPU_TRANSFORMS

false

Run the preprocessing transforms on the accelerator.

BRAINSEG_TIMEOUT_SECONDS

1800

Whole-request cap.

BRAINSEG_LOG_LEVEL

INFO

Log level of the service.

A variable that the profile does not set can be changed with -e, as in section 1.2. To change a variable that the profile sets, write a custom profile with the changed env_vars, mount its directory at /workspace/aim-runtime/profiles/custom, and set AIM_PROFILE_ID to its filename without the .yaml extension.

5. Advanced deployment scenarios#

5.1 Selecting a different profile#

Each profile of this AIM is pinned with AIM_PROFILE_ID. The profile identifier is the filename of the profile without the .yaml extension. The README lists the available profiles and their accelerator counts.

docker run \
  -e AIM_ID=monai/wholeBrainSeg_Large_UNEST_segmentation \
  -e AIM_PROFILE_ID=<PROFILE_ID> \
  -v /path/to/model-cache:/workspace/model-cache \
  --device=/dev/kfd --device=/dev/dri \
  --group-add video --ipc=host \
  -p 8000:8000 \
  amdenterpriseai/aim-instinct-model-monai-wholebrainseg-large-unest-segmentation:2026.14.0

Section 1.1 describes how the runtime uses AIM_PROFILE_ID. If the profile uses a different number of accelerators than the default profile, change the accelerator count under resources in the section 3.1 manifest to match.

6. Monitoring and troubleshooting#

6.1 Getting help on the commands#

A general help command is available as follows:

docker run \
  amdenterpriseai/aim-instinct-model-monai-wholebrainseg-large-unest-segmentation:2026.14.0 \
  --help

A help command for specific subcommands is also available:

docker run \
  amdenterpriseai/aim-instinct-model-monai-wholebrainseg-large-unest-segmentation:2026.14.0 \
  <subcommand> --help

6.2 Enabling detailed logging#

docker run \
  -e AIM_LOG_LEVEL=DEBUG \
  -e AIM_ID=monai/wholeBrainSeg_Large_UNEST_segmentation \
  -e AIM_PROFILE_ID=bentoml-mi300x-fp16-tp1-latency \
  -v /path/to/model-cache:/workspace/model-cache \
  --device=/dev/kfd --device=/dev/dri \
  --group-add video --ipc=host \
  -p 8000:8000 \
  amdenterpriseai/aim-instinct-model-monai-wholebrainseg-large-unest-segmentation:2026.14.0

6.3 Checking profile selection results#

It is possible to check which profile AIM selects from the given environment variables.

docker run \
  -e AIM_ID=monai/wholeBrainSeg_Large_UNEST_segmentation \
  -e AIM_ACCELERATOR_COUNT=1 \
  -e AIM_PRECISION=fp16 \
  -e AIM_ACCELERATOR_MODEL=MI300X \
  amdenterpriseai/aim-instinct-model-monai-wholebrainseg-large-unest-segmentation:2026.14.0 \
  dry-run

6.4 List available profiles#

docker run \
  amdenterpriseai/aim-instinct-model-monai-wholebrainseg-large-unest-segmentation:2026.14.0 \
  list-profiles

6.5 Common problems#

  • The container restarts during the first start. The manifest in section 3.1 gives the start one hour. If the model needs more, raise failureThreshold on the startup probe. Mount a persistent cache, so that later starts skip the download.

  • The container restarts after it answers on /healthz. The liveness probe accepts five minutes of failed checks. If the model does more work after the health endpoint answers, and that work takes longer, raise failureThreshold on the liveness probe.

  • The container stops with a profile selection error. The error message gives the cause.

    • If AIM_PROFILE_ID is set, the ID matches no profile of this AIM. Compare it with the README or the output of list-profiles (see section 6.4).

    • If the error lists profile IDs, the matching profiles must be selected explicitly. Set AIM_PROFILE_ID to one of them.

    • If no profile is compatible, the error counts the profiles per reason, such as accelerator_mismatch or precision_mismatch. Compare the Detected Accelerator and Accelerator Count lines in the log, and any AIM_PRECISION or AIM_METRIC you set, with the profile table in the README. Or set AIM_PROFILE_ID to a profile from the README (see section 1.1).

  • /healthz does not answer yet. The service binds the port only after the model is loaded. Read the container logs to see how far startup got.

7. Security considerations#

  • Use Kubernetes secrets or environment variables for sensitive credentials

  • Implement appropriate network policies to restrict access to your deployment

  • The model cache holds downloaded model weights. Restrict write access to it

Contents