thomassuedbroecker / caikit-custom-model-protofiles

This project shows how to add a custom model to caikit and use the proto-files with GRCURL.

Caikit custom model profiles

---

----- UNDER CONSTRUCTION -----

1. Objective

The objective of the project is to add a custom model to caikit==v0.4.1 and use grpcurl to invoke a text generation command.

The project contains a client/server example for using GRPC.

2. Environment setup

• Clone the project

git clone https://github.com/thomassuedbroecker/caikit-custom-model-protofiles.git

• Create virtual environment

cd code

python3.10 -m venv caikit-env-3.10
source ./caikit-env-3.10/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install -r requirements.txt
python3 -m pip install grpcio
python3 -m pip install grpcio-tools
python3 -m pip install grpcio-health-checking
python3 -m pip install grpcio-reflection
python3 -m pip install prometheus_client
python3 -m pip install py_grpc_prometheus
python3 -m pip install caikit.runtime.service_factory
#python3 -m caikit.runtime.dump_services protos

3. Start the server/client example

• Simplified dependences of the client/server architecture of the example

• Start the server in the first terminal

source ./caikit-env-3.10/bin/activate
python start_runtime.py

Output:

...
2023-08-06T11:28:10.609698 [2PCVR:DBUG] Message name: TrainingStatusResponse
2023-08-06T11:28:10.609997 [2PCVR:DBUG] Creating FileDescriptorProto
2023-08-06T11:28:10.610045 [2PCVR:DBUG] Adding Descriptors to DescriptorPool
{"channel": "SERVR-GRPC", "exception": null, "level": "info", "message": "Serving prometheus metrics on port 8086", "num_indent": 0, "thread_id": 140704560412224, "timestamp": "2023-08-06T11:28:10.623832"}
...

• Start the client in the second terminal

source ./caikit-env-3.10/bin/activate
python client.py

3. Create protos

Execute the following command in the third terminal. The command will create a folder with the proto files on your local computer.

source ./caikit-env-3.10/bin/activate
export CONFIG_FILES=text_generation/config.yml
python3 -m caikit.runtime.dump_services protos

Use grpcurl

• List the services

grpcurl -import-path ./protos -proto textgenerationservice.proto list

Example output:

caikit.runtime.TextGeneration.TextGenerationService

• List the services with the grpc-server

grpcurl -import-path ./protos -proto textgenerationservice.proto -plaintext localhost:8085 list

Example output:

caikit.runtime.TextGeneration.TextGenerationService

• Describe the service

grpcurl -import-path ./protos -proto textgenerationservice.proto -plaintext localhost:8085 describe caikit.runtime.TextGeneration.TextGenerationService

Example output:

service TextGenerationService {
  rpc HfModulePredict ( .caikit.runtime.TextGeneration.HfModuleRequest ) returns ( .text_generation.data_model.TextOutput );
}

grpcurl -import-path ./protos -proto hfmodulerequest.proto -plaintext localhost:8085 describe caikit.runtime.TextGeneration.HfModuleRequest

Example output:

message HfModuleRequest {
  //-- fields --
  .text_generation.data_model.TextInput text_input = 1;
}

• Invoke a method of the service

MM_MODEL_ID=text_generation
grpcurl -d '{"text_input":{"text":"This is"}}' -H "mm-model-id:$MM_MODEL_ID" -import-path ./protos -proto textgenerationservice.proto -plaintext localhost:8085 caikit.runtime.TextGeneration.TextGenerationService/HfModulePredict

Container

• Build a container image and execute a container image

docker build -f Dockerfile -t caikit-server-proto-export:v1 .
docker run -it caikit-server-proto-export:v1

• Verify the created protos

Navigate inside the container to the protos folder.

cd protos
ls

• Example output

hfmodulerequest.proto                textoutput.proto
modelpointer.proto                   traininginforequest.proto
producerid.proto                     traininginforesponse.proto
producerpriority.proto               trainingjob.proto
textgenerationservice.proto          trainingmanagement.proto
textgenerationtrainingservice.proto  trainingstatus.proto
textinput.proto