Online Inference Serving

Introduction
diagram
CoreWeave Cloud allows for easy serving of machine learning models. The models can be sourced from a range of storage backends, including Amazon S3 and a CoreWeave Storage Volume. After deployment the inference engine auto scales the containers based on demand to swiftly fulfill user requests and scales down as load decreases to not waste GPU resources. Allocating new resources and scaling up a container can be as fastas 20 seconds for the 6B GPT-J model. The quick autoscale allows a significantly more responsive service than depending on scaling of hypervisor backed instances of other cloud providers.
Well supported Open Source tools back the inference stack. Knative Serving acts as the serverless runtime, managing auto scaling, revision control and canary deployments. Kubeflow Serving provides an easy to use interface, via Kubernetes resource definitions to deploy models without having to worry about correctly configuring the underlying framework (ie. Tensorflow). The examples in this repository are tailored specifically for common use cases, and there are many more examples in the KFServing repostiory that can be used directly in your namespace.
Autoscaling
Autoscaling is enabled by default for all Inference Services. The autoscaling parameters have been pre-configured for GPU based workloads, where a large dataset usually needs to be loaded into GPU VRAM before serving can begin.
KNative Parameter
Value
Description
tick-interval
15s
Auto scaling decision are made every 15 seconds
stable-window
180s
The time period average concurrency is measured over in stable mode
container-concurrency-target-percentage
85%
Scale to keep an average headroom of 15% of available concurrent request to accommodate for bursts
max-scale-up-rate
20
Scale up at a maximum of 20x of current capacity or 1 container (whichever is larger) per 15 seconds
max-scale-down-rate
1.1
Scale down at a maximum of 10% of current capacity or 1 container (whichever is larger) per 15 seconds
scale-to-zero-pod-retention-period
30m
If no requests have been received for 30 minutes, a service will be scaled to zero and not use any resources. This behavior can be disabled by setting minReplicas to 1 in the service spec
scale-down-delay
60s
Containers are only scaled down if scaled-down has been requested over a 60s period. This is to avoid thrashing.
If concurrent requests exceeds the current scaled-for request volume by 200% during a period of 24 seconds, the autoscaler enters "panic mode" and starts scaling containers faster than the normal 128s stable window. Some of these settings, such as stable window can be modified using annotations on the InferenceService. Refer to the KNative documentation for more information.
Scale To Zero
Inference Services with long periods of idle time can automatically be scaled to zero. When scaled down, the Inference Service will consume no resources and incur no billing. As soon as a new request comes in, a Pod will be instantiated and the request will be served. For small models, this can be as quick as five seconds. For larger models, spin up times can be 30 to 60 seconds. Model loading times are highly dependent on the code responsible of loading the model into the GPU. Many popular PyTorch frameworks do not optimize for optimal loading times.
To enable Scale To Zero, simply set minReplicas to 0. By default, scale down will happen in 0 to 30 minutes after the last request was served. To set the max scale down delay lower, add the annotation autoscaling.knative.dev/scaleToZeroPodRetentionPeriod: "10m" to the InferenceService. By default, this is set to 30 minutes.
Billing
For on demand customers, billing is done on a per minute basis when containers are running. Scale To Zero allows rarely used models to occur no costs, while still be available to receive requests. Please note that when an InferenceService is scaled to zero due to being idle, it usually takes 30-60 seconds depending on model size until the first API request is answered. If this is unacceptable, we recommend setting minReplicas to 1.
Model Storage
Models can be served directly from Amazon S3, which is practical for testing and multi cloud deployment. For faster container launch times in a production environment it is recommended to cache the model in a ReadWriteMany persistent volume on CoreWeave storage. The model can be written to a PVC from any container, such as SSH, Jupyter or FileBrowser deployed via CoreWeave Apps.
For best performance, small models (less than 1GB) can safely be included in the docker image. For larger models, loading from a storage PVC is strongly recommended. Use NVMe backed storage volumes for optimal loading speeds.
GPU Selectors & Affinities
You will want to use affinities, to specify what GPU type (or CPU, in case of CPU only Inference) your InferenceService should be scheduled on.
1
spec:
2
  predictor:
3
    minReplicas: 0 
4
    maxReplicas: 3
5
    containerConcurrency: 1
6
    containers:
7
    ...
8
      resources:
9
        requests:
10
          cpu: 1
11
          memory: 6Gi
12
        limits:
13
          cpu: 3
14
          memory: 16Gi
15
          nvidia.com/gpu: 1
16
    affinity:
17
      nodeAffinity:
18
        requiredDuringSchedulingIgnoredDuringExecution:
19
          nodeSelectorTerms:
20
          - matchExpressions:
21
            - key: gpu.nvidia.com/class
22
              operator: In
23
              values:
24
              - Quadro_RTX_5000
Copied!

Highly Available Storage using Samba-AD and AD DFS

Next - Inference

One Click Models

Last modified 24d ago

Copy link

Contents

GPU Selectors & Affinities