Cogstack ModelServe (CMS) is a model-serving and model-governance system created for a range of CogStack NLP tasks. Targeting language models with NER and entity linking capabilities, CMS provides a one-stop shop for serving and fine-tuning models, training lifecycle management, as well as monitoring and end-to-end observability.
A virtual environment is highly recommended prior to installation. To install the dependencies, run the following command:
pip install -e .
Currently, CMS offers both HTTP endpoints for running NLP-related jobs and a command-line interface for administering model management tasks. Multi-language SDKs are under development for providing increased accessibility.
- SNOMED MedCAT Model
- ICD-10 MedCAT Model
- UMLS MedCAT Model
- De-ID MedCAT Model (AnonCAT)
- De-ID Transformers Model
- All-in-One Doc
You can use the following commands to explore available CLI options (see in full docs):
python app/cli/cli.py --help
python app/cli/cli.py serve --help
CMS runs the NLP model packaged in a single ZIP file. To download the GA models, please follow the instructions. Contact Cogstack if you are interested in trying out Alpha release such as the de-identification model.
Default configuration properties can be found in ./app/envs/.envs and further tweaked if needed.
To serve NLP models from your system environment, run the following command:
python app/cli/cli.py serve --model-type <model-type> --model-path PATH/TO/MODEL_PACKAGE.zip --host 127.0.0.1 --port 8000
Then the API docs similar to the ones shown above will be accessible at http://127.0.0.1:8000/docs.
The following table summarises the servable model types with their respective output concepts:
| model-type | docker-service | output-spans |
|---|---|---|
| medcat_snomed | medcat-snomed | labelled with SNOMED concepts |
| medcat_icd10 | medcat-icd10 | labelled with ICD-10 concepts |
| medcat_umls | medcat-umls | labelled with UMLS concepts |
| medcat_deid (anoncat) | medcat-deid | labelled with latest PII concepts |
| transformers_deid | de-identification | labelled with PII concepts |
Default configuration properties can be found and customised in ./docker/<MODEL-TYPE>/.envs
To serve NLP models through a container, run the following commands:
export MODEL_PACKAGE_FULL_PATH=<PATH/TO/MODEL_PACKAGE.zip>
export CMS_UID=$(id -u $USER)
export CMS_GID=$(id -g $USER)
docker compose -f docker-compose.yml up -d <model-service>
Then the API docs will be accessible at localhost on the mapped port specified in docker-compose.yml. The container runs
as a cms non-root user configured during the image build. Ensure the model package file is owned by the currently
logged-in user to avoid permission-related errors. If the file ownership is altered, you will need to rebuild the image.
In addition to the core services such as serving, training and evaluation, CMS provides several ready-to-use components to help users make these production-ready. Their presence and default configuration can be customised to what suits your own needs. The diagram below illustrates the interactions between core and auxiliary services:
CMS has embedded a mlflow stack to provide a pipeline of model lifecycle management, covering tracking of training/fine-tuning,
model registration, versioning on new models and so on.
export MLFLOW_DB_USERNAME=<MLFLOW_DB_USERNAME>
export MLFLOW_DB_PASSWORD=<MLFLOW_DB_PASSWORD>
export AWS_ACCESS_KEY_ID=<AWS_ACCESS_KEY_ID>
export AWS_SECRET_ACCESS_KEY=<AWS_SECRET_ACCESS_KEY>
docker compose -f docker-compose-mlflow.yml up -d
The user authentication feature is optional and turned off by default. If you wish to enable it, consult the Management documentation.
All models are registered in a so-called model registry, which can be run either on a local machine or on a remote HTTP server. To work with the model registry, the prerequisites include setting up the following environment variables:
export AWS_ACCESS_KEY_ID=<AWS_ACCESS_KEY_ID>
export AWS_SECRET_ACCESS_KEY=<AWS_SECRET_ACCESS_KEY>
export MLFLOW_TRACKING_URI=<MLFLOW_TRACKING_URI>
After that, you can run ModelServe with a registered model. For example, with minio as the model storage backend,
python app/cli/cli.py serve --model-type medcat_icd10 --mlflow-model-uri s3://cms-model-bucket/EXPERIMENT_ID/RUN_ID/artifacts/REGISTERED_MODEL_NAME --host 0.0.0.0 --port 8000
You may want to push the package of a pretrained model to the registry before serving. For example, with minio being the model storage backend,
python app/cli/cli.py register --model-type medcat_snomed --model-path PATH/TO/MODEL_PACKAGE.zip --model-name REGISTERED_MODEL_NAME
You can also serve a registered model through standard MLflow APIs instead of ModelServe APIs:
mlflow models serve -m s3://cms-model-bucket/EXPERIMENT_ID/RUN_ID/artifacts/REGISTERED_MODEL_NAME -h 0.0.0.0 -p 8001
Then the /invocations endpoint will be up and running for model scoring. For example:
curl http://127.0.0.1:8001/invocations \
-H 'Content-Type: application/json' \
-d '{"dataframe_split": { "columns": ["name", "text"], "data": [["DOC", "TEXT"], ["ANOTHER_DOC", "ANOTHER_TEXT"]]}}'
By default, CMS exposes most services via HTTP APIs. Their monitoring can be performed by running the mon stack:
export GRAFANA_ADMIN_USER=<GRAFANA_ADMIN_USER>
export GRAFANA_ADMIN_PASSWORD=<GRAFANA_ADMIN_PASSWORD>
docker compose -f docker-compose-mon.yml up -d
CMS provides a log stack to run services used for centralised logging and log analysis:
export GRAYLOG_PASSWORD_SECRET=<GRAYLOG_PASSWORD_SECRET>
export GRAYLOG_ROOT_PASSWORD_SHA2=<GRAYLOG_ROOT_PASSWORD_SHA2>
docker compose -f docker-compose-log.yml up -d
A reverse proxying service is available inside the proxy stack:
docker compose -f docker-compose-proxy.yml up -d
An optional user database can be spun up if authentication is needed:
export AUTH_DB_USERNAME=<AUTH_DB_USERNAME>
export AUTH_DB_PASSWORD=<AUTH_DB_PASSWORD>
docker compose -f docker-compose-auth.yml up -d
Before running ModelServe, define extra environment variables to enable the built-in token-based authentication and hook it up with the database by following this Instruction.
You can send your texts to the CMS stream endpoint and receive NLP results as a stream. To that end, start CMS as a streamable service by running:
python app/cli/cli.py serve --streamable --model-type <model-type> --model-path PATH/TO/MODEL_PACKAGE.zip --host 127.0.0.1 --port 8000
Currently, JSON Lines is supported for formatting request and response bodies. For example, the following request:
curl -X 'POST' 'http://127.0.0.1:8000/stream/process' \
-H 'Content-Type: application/x-ndjson' \
--data-binary $'{"name": "DOC", "text": "TEXT"}\n{"name": "ANOTHER_DOC", "text": "ANOTHER_TEXT"}'
will result in a response like {"doc_name": "DOC", "start": INT, "end": INT, "label_name": "STR", "label_id": "STR", ...}\n...
You can also "chat" with the running model using the /stream/ws endpoint. For example:
<form action="" onsubmit="send_doc(event)">
<input type="text" id="cms-input" autocomplete="off"/>
<button>Send</button>
</form>
<ul id="cms-output"></ul>
<script>
var ws = new WebSocket("ws://localhost:8000/stream/ws");
ws.onmessage = function(event) {
document.getElementById("cms-output").appendChild(
Object.assign(document.createElement('li'), { textContent: event.data })
);
};
function send_doc(event) {
ws.send(document.getElementById("cms-input").value);
event.preventDefault();
};
</script>