Skip to content

Latest commit

 

History

History
169 lines (114 loc) · 3.99 KB

FasterTransformer_HuggingFace_Bert.md

File metadata and controls

169 lines (114 loc) · 3.99 KB
Raw

Faster Transformer HuggingFace Bert example in Kubernetes Torchserve.

Overview

This document demonstrates, running fast transformers HuggingFace BERT example with Torchserve in kubernetes setup.

Refer: FasterTransformer_HuggingFace_Bert

Using Azure AKS Cluster

AKS Cluster setup

Using AWS EKS Cluster

EKS Cluster setup

Using Google GKE Cluster

GKE Cluster setup

Once the cluster and the PVCs are ready, we can generate MAR file.

Generate Mar file

Follow steps from here to generate MAR file

Copy Mar file from container to local path

docker cp <container-id>:/workspace/serve/examples/FasterTransformer_HuggingFace_Bert/BERTSeqClassification.mar ./BERTSeqClassification.mar

Create config.properties

inference_address=http://0.0.0.0:8080
management_address=http://0.0.0.0:8081
metrics_address=http://0.0.0.0:8082
NUM_WORKERS=1
number_of_gpu=1
install_py_dep_per_model=true
number_of_netty_threads=32
job_queue_size=1000
model_store=/home/model-server/shared/model-store
model_snapshot={"name":"startup.cfg","modelCount":1,"models":{"bert":{"1.0":{"defaultVersion":true,"marName":"BERTSeqClassification.mar","minWorkers":2,"maxWorkers":3,"batchSize":1,"maxBatchDelay":100,"responseTimeout":120}}}}

Copy Mar file and config.properties to PVC

kubectl exec --tty pod/model-store-pod -- mkdir /pv/model-store/
kubectl cp BERTSeqClassification.mar model-store-pod:/pv/model-store/BERTSeqClassification.mar

kubectl exec --tty pod/model-store-pod -- mkdir /pv/config/
kubectl cp config.properties model-store-pod:/pv/config/config.properties

Build Torchserve image

  1. Clone Torchserve Repo
git clone https://github.com/pytorch/serve.git
cd serve/docker
  1. Modify Python and Pip paths in Dockerfile as below
sed -i 's#/usr/bin/python3#/opt/conda/bin/python3#g' Dockerfile
sed -i 's#/usr/local/bin/pip3#/opt/conda/bin/pip3#g' Dockerfile
  1. Change GPU check in Dockerfile for nvcr.io image
sed -i 's#grep -q "cuda:"#grep -q "nvidia:"#g' Dockerfile
  1. Add transformers==2.5.1 to Dockerfile
sed -i 's#pip install --no-cache-dir captum torchtext torchserve torch-model-archiver#& transformers==2.5.1#g' Dockerfile

  1. Build image

    Refer: NGC deep learning framework container

DOCKER_BUILDKIT=1 docker build -file Dockerfile -t <image-name> --build-arg BASE_IMAGE=nvcr.io/nvidia/pytorch:20.12-py3 --build-arg CUDA_VERSION=cu102 .
  1. Push image
docker push <image-name>

Install Torchserve

  1. Navigate to kubernetes TS Helm package folder
cd ../kubernetes/Helm
  1. Modify values.yaml with image and memory
torchserve_image: <image build in previous step>

namespace: torchserve

torchserve:
  management_port: 8081
  inference_port: 8080
  metrics_port: 8082
  pvd_mount: /home/model-server/shared/
  n_gpu: 1
  n_cpu: 4
  memory_limit: 32Gi
  memory_request: 32Gi

deployment:
  replicas: 1

persitant_volume:
  size: 1Gi
  1. Install TS
helm install torchserve .
  1. Check TS installation
Kubectl get pods -n default
Kubectl logs <pod-name> -n default

Run Inference

  1. Start a shell session into the TS pod
kubectl exec -it <pod-name> -- bash
  1. Create input file

Sample_text_captum_input.txt

{
  "text": "Bloomberg has decided to publish a new report on the global economy.",
  "target": 1
}
  1. Run inference
curl -X POST http://127.0.0.1:8080/predictions/bert -T ../Huggingface_Transformers/Seq_classification_artifacts/sample_text_captum_input.txt