kserve/docs/samples/v1beta1/torchserve/v1/bloom

TorchServe example with Huggingface BLOOM model

In this example we will show how to serve Large Huggingface models with TorchServe on KServe.

Model archive file creation

Clone pytorch/serve repository, navigate to examples/Huggingface_Largemodels and follow the steps for creating the MAR file including serialized model and other dependent files.

The above Torchserve example works on shard version of Huggingface models.

For sharding the Huggingface models you can use the following script, and then compress the model

from transformers import AutoModelForCausalLM, AutoTokenizer
model_name="bigscience/bloomz-7b1"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
model.save_pretrained("model/"+model_name, max_shard_size="5GB")
tokenizer.save_pretrained("model/"+model_name)

Create NVMe Persistent Volume

Use SSH to connect to the worker nodes and prepare the NVMe drives for Kubernetes, as follows.

Run the lsblk command on each worker node to lists the available disks.

NAME          MAJ:MIN RM   SIZE RO TYPE MOUNTPOINT
nvme1n1       259:0    0 116.4G  0 disk 
nvme0n1       259:1    0    80G  0 disk 
├─nvme0n1p1   259:2    0    80G  0 part /
└─nvme0n1p128 259:3    0     1M  0 part 

Output of the command line that shows the result of running the lsblk command.

$ sudo mkfs.xfs /dev/nvme1n1

$ sudo mkdir -p /mnt/data/vol1
$ sudo chmod -R 777 /mnt     
$ sudo mount /dev/nvme1n1 /mnt/data/vol1

Permanently mount the device:

$ sudo blkid /dev/nvme1n1

To get it to mount every time, add the following line to the /etc/fstab file:

UUID=nvme_UUID      /mnt/data/vol1   xfs    defaults,nofail    0    2

Clone the local provisioner repository:

$ git clone https://github.com/kubernetes-sigs/sig-storage-local-static-provisioner.git

Create a StorageClass yaml file storageclass.yaml

kind: StorageClass
apiVersion: storage.k8s.io/v1
metadata:
  name: fast-disks
provisioner: kubernetes.io/no-provisioner
volumeBindingMode: WaitForFirstConsumer

$ kubectl apply -f storageclass.yaml

Create Local Persistent Volumes for Kubernetes

• Change hostDir to the mount path

cd sig-storage-local-static-provisioner
helm template ./helm/provisioner > ./provisioner/deployment/kubernetes/provisioner_generated.yaml

kubectl apply -f ./deployment/kubernetes/provisioner_generated.yaml

Output of the command line that shows the result of running the kubectl get pods command.

NAME                                         READY   STATUS    RESTARTS   AGE
kserve-controller-manager-5c5c4d8c89-lrzbd   2/2     Running   0          4d2h
local-nvme-pv-provisioner-vwxgt              1/1     Running   0          16m

$ kubectl get pv

NAME                                       CAPACITY   ACCESS MODES   RECLAIM POLICY   STATUS      CLAIM                     STORAGECLASS   REASON   AGE
local-pv-2a85b8ac                          116Gi      RWO            Delete           Bound       kserve-test/model-cache   fast-disks              4d3h

Create PVC and Mount the Model.

kubectl apply -f pvc-pod.yaml

Refer: Bloom Model Example

Move the config.properties and MAR file to PVC in the below structure.

|_model-store
  |_bloom-560m.mar
|_config
  |_config.properties

kubectl exec -it pv-pod -- mkdir /pv/config
kubectl exec -it pv-pod -- mkdir /pv/model-store

kubectl cp config.properties pv-pod:/pv/config
kubectl cp bloom-560m.mar -it pv-pod:/pv/config

Create the InferenceService

Apply the CRD

kubectl apply -f bloom-560m.yaml

Expected Output

$inferenceservice.serving.kserve.io/torchserve-bloom-560m created

Run a prediction

The first step is to determine the ingress IP and ports and set INGRESS_HOST and INGRESS_PORT

MODEL_NAME=BLOOMSeqClassification
ISVC_NAME=torchserve-bloom-560m
SERVICE_HOSTNAME=$(kubectl get inferenceservice ${ISVC_NAME} -n <namespace> -o jsonpath='{.status.url}' | cut -d "/" -f 3)

curl -v -H "Host: ${SERVICE_HOSTNAME}" http://${INGRESS_HOST}:${INGRESS_PORT}/v1/models/${MODEL_NAME}:predict -d ./sample_text.txt

Expected Output

*   Trying 44.239.20.204...
* Connected to a881f5a8c676a41edbccdb0a394a80d6-2069247558.us-west-2.elb.amazonaws.com (44.239.20.204) port 80 (#0)
> PUT /v1/models/BLOOMSeqClassification:predict HTTP/1.1
> Host: torchserve-bloom-560m.kserve-test.example.com
> User-Agent: curl/7.47.0
> Accept: */*
> Content-Length: 79
> Expect: 100-continue
>
< HTTP/1.1 100 Continue
* We are completely uploaded and fine
< HTTP/1.1 200 OK
< cache-control: no-cache; no-store, must-revalidate, private
< content-length: 8
< date: Wed, 04 Nov 2020 10:54:49 GMT
< expires: Thu, 01 Jan 1970 00:00:00 UTC
< pragma: no-cache
< x-request-id: 4b54d3ac-185f-444c-b344-b8a785fdeb50
< x-envoy-upstream-service-time: 2085
< server: istio-envoy
<
* Connection #0 to host torchserve-bloom-560m.kserve-test.example.com left intact
Accepted

Note For larger models use A100 80g GPU instances.