Open Inference Protocol with nightly build not working

[harshita-meena]

🐛 Describe the bug

While trying to run load tests with latest merged changes on v2 Open inference protocol, I noticed that the example for mnist does not work in preprocessing step. https://github.com/pytorch/serve/pull/2609/files

Error logs

The server side showed error

Installation instructions

ARG VERSION=latest-cpu
ARG IMAGE_NAME=pytorch/torchserve-nightly

from $IMAGE_NAME:$VERSION

USER root

RUN apt-get -y update
RUN apt-get install -y curl vim

# Installation steps to download model from GCP
# Downloading gcloud package
RUN curl https://dl.google.com/dl/cloudsdk/release/google-cloud-sdk.tar.gz > /tmp/google-cloud-sdk.tar.gz

# Installing the package
RUN mkdir -p /usr/local/gcloud \
  && tar -C /usr/local/gcloud -xvf /tmp/google-cloud-sdk.tar.gz \
  && /usr/local/gcloud/google-cloud-sdk/install.sh

# Adding the package path to local
ENV PATH $PATH:/usr/local/gcloud/google-cloud-sdk/bin

ENV TS_OPEN_INFERENCE_PROTOCOL oip

RUN pip install protobuf googleapis-common-protos grpcio loguru

COPY config.properties /home/model-server/config.properties
COPY mnist.mar /home/model-server/model-store/

copied model from gs://kfserving-examples/models/torchserve/image_classifier/v2/model-store/mnist.mar Built the docker file using docker build -f Dockerfile -t metadata . and brought it up locally Ran ghz load test tool with

ghz  --proto serve/frontend/server/src/main/resources/proto/open_inference_grpc.proto   --call org.pytorch.serve.grpc.openinference.GRPCInferenceService/ModelInfer --duration 300s --rps 1 --insecure localhost:79 -D ./serve/kubernetes/kserve/kf_request_json/v2/mnist/mnist_v2_tensor_grpc.json

Model Packaing

Used an existing packaged model mnist.mar at gs://kfserving-examples/models/torchserve/image_classifier/v2

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 enable_metrics_api=true model_metrics_auto_detect=true metrics_mode=prometheus number_of_netty_threads=32 job_queue_size=1000 enable_envvars_config=true model_store=/home/model-server/model-store load_models=mnist.mar workflow_store=/home/model-server/wf-store

Versions

---

Environment headers

Torchserve branch:

**Warning: torchserve not installed .. torch-model-archiver==0.9.0

Python version: 3.7 (64-bit runtime) Python executable: /Users/hmeena/development/ml-platform-control-planes/venv/bin/python

Versions of relevant python libraries: numpy==1.21.6 requests==2.31.0 requests-oauthlib==1.3.1 torch-model-archiver==0.9.0 wheel==0.41.0 Warning: torch not present .. Warning: torchtext not present .. Warning: torchvision not present .. Warning: torchaudio not present ..

Java Version:

OS: Mac OSX 11.7.8 (x86_64) GCC version: N/A Clang version: 12.0.0 (clang-1200.0.32.29) CMake version: version 3.23.2

Versions of npm installed packages: **Warning: newman, newman-reporter-html markdown-link-check not installed...

Repro instructions

same as installation instruction

Possible Solution

I am unsure of how well the OIP is working with Torchserve at the moment. I tried a small ranker example and it fails in the post processing step where the worker crashes completely, it is not able to send response as ModelInferResponse.

[agunapal]

Hi @harshita-meena Thanks for reporting the issue.

Do you mind trying this script https://github.com/pytorch/serve/blob/master/kubernetes/kserve/tests/scripts/test_mnist.sh

We are running this nightly. https://github.com/pytorch/serve/actions/workflows/kserve_cpu_tests.yml

[harshita-meena]

Currently I am trying to setup the tests but they will probably not fail because the images used in oip kserve yamls are custom ones (http and grpc) and both do not refer to nightly ones though are part of the test_mnist.sh at lines 189 and 215.

[harshita-meena]

Still struggling to get the tests running, if you can approve the workflow for this PR. The primary reason I am trying to get this working is because I wanted to use Open Inference Protocol for a non-kserve deployment. Everything works till the worker dies after the post processing step. I was heavily relying on this because OIP provides a great generic metadata/inference API. If this doesn't work I will use the inference.proto instead.

[agunapal]

Hi @harshita-meena Thanks for the details. Checking with kserve regarding this. Will update

[harshita-meena]

You can reproduce the worker died issue, if you build the dockerfile part of this issue with config properties and create a new mnist.mar with a slightly modified handler for OIP specific requests. (attached a zip)

torch-model-archiver --model-name mnist --version 1.0 --serialized-file mnist_cnn.pt --model-file mnist.py --handler mnist_handler.py -r requirements.txt

mnist.zip

ghz --proto serve/frontend/server/src/main/resources/proto/open_inference_grpc.proto --call org.pytorch.serve.grpc.openinference.GRPCInferenceService/ModelInfer --duration 300s --rps 1 --insecure localhost:79 -D ./serve/kubernetes/kserve/kf_request_json/v2/mnist/mnist_v2_tensor_grpc.json

[agunapal]

Hi @harshita-meena Thanks! This is a new feature and there might be bugs. Will update when I repro it

[harshita-meena]

Hi @agunapal I was wondering if you identified the reason for the issue or got a chance to discuss with kserve.

[harshita-meena]

I figured out the solution and will reply back with it soon. Thankyou!

[agunapal]

Hi @harshita-meena I was able to repro the issue with the steps you shared. Please feel free to send a PR if you have identified the problem

[harshita-meena]

It is how the OIP expects the response, I was sending only dict or only list but if I send list of dicts containing parameters specific to OIP response, it will give successful prediction.

    def postprocess(self, data):
        """The post process of MNIST converts the predicted output response to a label.

        Args:
            data (list): The predicted output from the Inference with probabilities is passed
            to the post-process function
        Returns:
            list : A list of dictionary with predictons and explanations are returned.
        """

        return [{
            "model_name":"mnist",
            "model_version":"N/A",
            "id":"N/A",
            "outputs":[{"name":"output-0","datatype":"FLOAT64","shape":["1"],"data":[data.argmax(1).flatten().tolist()]}]
        }]

mnist_handler_oip.py.zip

[harshita-meena]

Just saw your message @agunapal, if it helps I can submit a PR with only handler specific to OIP.

[agunapal]

Hi @harshita-meena The error you posted and the one I see is in pre-processing? So, how is it related to post-processing.

Also, I'm wondering how do we address this backward compatibility breaking change for post-processing

[harshita-meena]

Apologies if my stream of errors confused you about the actual issue.

My main goal is to get an inference working with gRPC using Open Inference Protocol in a basic deployment not using kserve.

The pre-processing error is because I was using the old handler when i first started this issue that didn't extract the request following Open Inference Protocol but using the old inference.proto. The second error I posted is after I finally resolved pre-processing but I was not able to figure out the post processing step. Finally yesterday I was able to figure out how the post processing step should look.

Overall if your question is how can we prevent the worker from crashing, it will be in the OIP server logic in GRPCJob.java.

Posting my findings from yesterday The issue was in postprocessing before a response is sent, assuming the prediction is for a batch of 1 and has model predicted value 0.

• If i send a dictionary with output set as 0, handler layer does not allow me to send a response that is not list
• if i send a list with output [0], server layer needs dict in the response and json serialization fails

But if i send a response as a list of dictionary, the parsing logic goes through so OIP response can process it. [{"model_name":....."outputs":[{"name":"output-0","datatype":"INT64","shape":["1"],"data":[0]}] We will have to send each and every parameter part of OIP response (even the ones that are optional in the protocol) otherwise Pytorch Worker will crash in parsing.

[agunapal]

Thanks for the detailed findings. cc @lxning