Skip to content

Predict on a Triton InferenceService with TorchScript model

While Python is a suitable and preferred language for many scenarios requiring dynamism and ease of iteration, there are equally many situations where precisely these properties of Python are unfavorable. One environment in which the latter often applies is production – the land of low latencies and strict deployment requirements. For production scenarios, C++ is very often the language of choice, The following example will outline the path PyTorch provides to go from an existing Python model to a serialized representation that can be loaded and executed purely from C++ like Triton Inference Server, with no dependency on Python.

Setup

  1. Skip tag resolution for nvcr.io which requires auth to resolve triton inference server image digest
    kubectl patch cm config-deployment --patch '{"data":{"registriesSkippingTagResolving":"nvcr.io"}}' -n knative-serving
    
  2. Increase progress deadline since pulling triton image and big bert model may longer than default timeout for 120s, this setting requires knative 0.15.0+
    kubectl patch cm config-deployment --patch '{"data":{"progressDeadline": "600s"}}' -n knative-serving
    

Export as Torchscript Model

A PyTorch model’s journey from Python to C++ is enabled by Torch Script, a representation of a PyTorch model that can be understood, compiled and serialized by the Torch Script compiler. If you are starting out from an existing PyTorch model written in the vanilla eager API, you must first convert your model to Torch Script.

Convert the above model via Tracing and serialize the script module to a file

import torch
# Use torch.jit.trace to generate a torch.jit.ScriptModule via tracing.
example = torch.rand(1, 3, 32, 32)
traced_script_module = torch.jit.trace(net, example)
traced_script_module.save("model.pt")

Store your trained model on GCS in a Model Repository

Once the model is exported as Torchscript model file, the next step is to upload the model to a GCS bucket. Triton supports loading multiple models so it expects a model repository which follows a required layout in the bucket.

<model-repository-path>/
  <model-name>/
    [config.pbtxt]
    [<output-labels-file> ...]
    <version>/
      <model-definition-file>
    <version>/
      <model-definition-file>
    ...
  <model-name>/
    [config.pbtxt]
    [<output-labels-file> ...]
    <version>/
      <model-definition-file>
    <version>/
      <model-definition-file>
For example in your model repository bucket gs://kfserving-examples/models/torchscript, the layout can be
torchscript/
  cifar/
    config.pbtxt
    1/
      model.pt
The config.pbtxt defines a model configuration that provides the required and optional information for the model. A minimal model configuration must specify name, platform, max_batch_size, input, and output. Due to the absence of names for inputs and outputs in a TorchScript model, the name attribute of both the inputs and outputs in the configuration must follow a specific naming convention i.e. “__”. Where can be any string and refers to the position of the corresponding input/output. This means if there are two inputs and two outputs they must be named as: INPUT__0, INPUT__1 and OUTPUT__0, OUTPUT__1 such that INPUT__0 refers to first input and INPUT__1 refers to the second input, etc.
name: "cifar"
platform: "pytorch_libtorch"
max_batch_size: 1
input [
  {
    name: "INPUT__0"
    data_type: TYPE_FP32
    dims: [3,32,32]
  }
]
output [
  {
    name: "OUTPUT__0"
    data_type: TYPE_FP32
    dims: [10]
  }
]

instance_group [
    {
        count: 1
        kind: KIND_CPU
    }
]

To schedule the model on GPU you would need to change the instance_group with GPU kind

instance_group [
    {
        count: 1
        kind: KIND_GPU
    }
]

Inference with HTTP endpoint

Create the InferenceService

Create the inference service yaml with the above specified model repository uri.

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: torchscript-cifar10
spec:
  predictor:
    triton:
      storageUri: gs://kfserving-examples/models/torchscript
      runtimeVersion: 20.10-py3
      env:
      - name: OMP_NUM_THREADS
        value: "1"

Warning

Setting OMP_NUM_THREADS env is critical for performance, OMP_NUM_THREADS is commonly used in numpy, PyTorch, and Tensorflow to perform multi-threaded linear algebra. We want one thread per worker instead of many threads per worker to avoid contention.

kubectl apply -f torchscript.yaml

Expected Output

$ inferenceservice.serving.kserve.io/torchscript-cifar10 created

Run a prediction with curl

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

The latest Triton Inference Server already switched to use KServe prediction V2 protocol, so the input request needs to follow the V2 schema with the specified data type, shape.

MODEL_NAME=cifar10
INPUT_PATH=@./input.json
SERVICE_HOSTNAME=$(kubectl get inferenceservice torchscript-cifar10 -o jsonpath='{.status.url}' | cut -d "/" -f 3)
curl -v -X -H "Host: ${SERVICE_HOSTNAME}" POST http://${INGRESS_HOST}:${INGRESS_PORT}/v2/models/$MODEL_NAME/infer -d $INPUT_PATH
Expected Output
* Connected to torchscript-cifar.default.svc.cluster.local (10.51.242.87) port 80 (#0)
> POST /v2/models/cifar10/infer HTTP/1.1
> Host: torchscript-cifar.default.svc.cluster.local
> User-Agent: curl/7.47.0
> Accept: */*
> Content-Length: 110765
> Content-Type: application/x-www-form-urlencoded
> Expect: 100-continue
>
< HTTP/1.1 100 Continue
* We are completely uploaded and fine
< HTTP/1.1 200 OK
< content-length: 315
< content-type: application/json
< date: Sun, 11 Oct 2020 21:26:51 GMT
< x-envoy-upstream-service-time: 8
< server: istio-envoy
<
* Connection #0 to host torchscript-cifar.default.svc.cluster.local left intact
{"model_name":"cifar10","model_version":"1","outputs":[{"name":"OUTPUT__0","datatype":"FP32","shape":[1,10],"data":[-2.0964810848236086,-0.13700756430625916,-0.5095657706260681,2.795621395111084,-0.5605481863021851,1.9934231042861939,1.1288187503814698,-1.4043136835098267,0.6004879474639893,-2.1237082481384279]}]}

Run a performance test

QPS rate --rate can be changed in the perf.yaml.

kubectl create -f perf.yaml

Requests      [total, rate, throughput]         6000, 100.02, 100.01
Duration      [total, attack, wait]             59.995s, 59.99s, 4.961ms
Latencies     [min, mean, 50, 90, 95, 99, max]  4.222ms, 5.7ms, 5.548ms, 6.384ms, 6.743ms, 9.286ms, 25.85ms
Bytes In      [total, mean]                     1890000, 315.00
Bytes Out     [total, mean]                     665874000, 110979.00
Success       [ratio]                           100.00%
Status Codes  [code:count]                      200:6000
Error Set:

Inference with gRPC endpoint

Create the InferenceService

Create the inference service yaml and expose the gRPC port, currently only one port is allowed to expose either HTTP or gRPC port and by default HTTP port is exposed.

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: torchscript-cifar10
spec:
  predictor:
    triton:
      storageUri: gs://kfserving-examples/models/torchscript
      runtimeVersion: 20.10-py3
      ports:
      - containerPort: 9000
        name: h2c
        protocol: TCP
      env:
      - name: OMP_NUM_THREADS
        value: "1"

Apply the gRPC InferenceService yaml and then you can call the model with tritonclient python library after InferenceService is ready.

kubectl apply -f torchscript_grpc.yaml

Add Transformer to the InferenceService

Triton Inference Server expects tensors as input data, often times a pre-processing step is required before making the prediction call when the user is sending in request with raw input format. Transformer component can be specified on InferenceService spec for user implemented pre/post processing code. User is responsible to create a python class which extends from KServe KFModel base class which implements preprocess handler to transform raw input format to tensor format according to V2 prediction protocol, postprocess handle is to convert raw prediction response to a more user friendly response.

Implement pre/post processing functions

import kserve
from typing import List, Dict
from PIL import Image
import torchvision.transforms as transforms
import logging
import io
import numpy as np
import base64

logging.basicConfig(level=kserve.constants.KSERVE_LOGLEVEL)

transform = transforms.Compose(
        [transforms.ToTensor(),
         transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

def image_transform(instance):
    byte_array = base64.b64decode(instance['image_bytes']['b64'])
    image = Image.open(io.BytesIO(byte_array))
    a = np.asarray(image)
    im = Image.fromarray(a)
    res = transform(im)
    logging.info(res)
    return res.tolist()


class ImageTransformer(kserve.KFModel):
    def __init__(self, name: str, predictor_host: str):
        super().__init__(name)
        self.predictor_host = predictor_host

    def preprocess(self, inputs: Dict) -> Dict:
        return {
           'inputs': [
               {
                 'name': 'INPUT__0',
                 'shape': [1, 3, 32, 32],
                 'datatype': "FP32",
                 'data': [image_transform(instance) for instance in inputs['instances']]
               }
            ]
        }

    def postprocess(self, results: Dict) -> Dict:
        # Here we reshape the data because triton always returns the flatten 1D array as json if not explicitly requesting binary
        # since we are not using the triton python client library which takes care of the reshape it is up to user to reshape the returned tensor.
        return {output["name"] : np.array(output["data"]).reshape(output["shape"]) for output in results["outputs"]}
Please find the code example here.

Build Transformer docker image

docker build -t $DOCKER_USER/image-transformer-v2:latest -f transformer.Dockerfile . --rm

Create the InferenceService with Transformer

Please use the YAML file to create the InferenceService, which adds the image transformer component with the docker image built from above.

apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
  name: torch-transfomer
spec:
  predictor:
    triton:
      storageUri: gs://kfserving-examples/models/torchscript
      runtimeVersion: 20.10-py3
      env:
      - name: OMP_NUM_THREADS
        value: "1"
  transformer:
    containers:
    - image: kfserving/image-transformer-v2:latest
      name: kserve-container
      command:
      - "python"
      - "-m"
      - "image_transformer_v2"
      args:
      - --model_name
      - cifar10
      - --protocol
      - v2

kubectl apply -f torch_transformer.yaml

Expected Output

$ inferenceservice.serving.kserve.io/torch-transfomer created

Run a prediction with curl

The transformer does not enforce a specific schema like predictor but the general recommendation is to send in as a list of object(dict): "instances": <value>|<list-of-objects>

{
  "instances": [
    {
      "image_bytes": { "b64": "aW1hZ2UgYnl0ZXM=" },
      "caption": "seaside"
    },
    {
      "image_bytes": { "b64": "YXdlc29tZSBpbWFnZSBieXRlcw==" },
      "caption": "mountains"
    }
  ]
}

SERVICE_NAME=torch-transfomer
MODEL_NAME=cifar10
INPUT_PATH=@./image.json

SERVICE_HOSTNAME=$(kubectl get inferenceservice $SERVICE_NAME -o jsonpath='{.status.url}' | cut -d "/" -f 3)

curl -v -X POST -H "Host: ${SERVICE_HOSTNAME}" http://${INGRESS_HOST}:${INGRESS_PORT}/v1/models/$MODEL_NAME:predict -d $INPUT_PATH

Expected Output

> POST /v2/models/cifar:predict HTTP/2
> user-agent: curl/7.71.1
> accept: */*
> content-length: 3422
> content-type: application/x-www-form-urlencoded
>
* We are completely uploaded and fine
* TLSv1.3 (IN), TLS handshake, Newsession Ticket (4):
* TLSv1.3 (IN), TLS handshake, Newsession Ticket (4):
* old SSL session ID is stale, removing
* Connection state changed (MAX_CONCURRENT_STREAMS == 4294967295)!
< HTTP/2 200
< content-length: 338
< content-type: application/json; charset=UTF-8
< date: Thu, 08 Oct 2020 13:15:14 GMT
< server: istio-envoy
< x-envoy-upstream-service-time: 52
<
{"model_name": "cifar", "model_version": "1", "outputs": [{"name": "OUTPUT__0", "datatype": "FP32", "shape": [1, 10], "data": [-0.7299326062202454, -2.186835289001465, -0.029627874493598938, 2.3753483295440674, -0.3476247489452362, 1.3253062963485718, 0.5721136927604675, 0.049311548471450806, -0.3691796362400055, -1.0804035663604736]}]}

Back to top