The model-serving framework is an experimental feature. For updates on the progress of the model-serving framework, or if you want to leave feedback that could help improve the feature, join the discussion in the [Model-serving framework forum](https://forum.opensearch.org/t/feedback-machine-learning-model-serving-framework-experimental-release/11439).
ML Commons allows you to serve custom models and use those models to make inferences. For those who want to run their PyTorch deep learning model inside an OpenSearch cluster, you can upload and run that model with the ML Commons REST API.
This page outlines the steps required to upload a custom model and run it with the ML Commons plugin.
To upload a custom model to OpenSearch, you need to prepare it outside of your OpenSearch cluster. You can use a pretrained model, like one from [Huggingface](https://huggingface.co/), or train a new model in accordance with your needs.
To use a model in OpenSearch, you'll need to export the model into a portable format. As of Version 2.5, OpenSearch only supports the [TorchScript](https://pytorch.org/docs/stable/jit.html) and [ONNX][https://onnx.ai/] formats.
Furthermore, files must be saved as zip files before upload. Therefore, to ensure that ML Commons can upload your model, compress your TorchScript file before uploading. You can download an example file [here](https://github.com/opensearch-project/ml-commons/blob/2.x/ml-algorithms/src/test/resources/org/opensearch/ml/engine/algorithms/text_embedding/all-MiniLM-L6-v2_torchscript_sentence-transformer.zip).
Most deep learning models are more than 100 MB, making it difficult to fit them into a single document. OpenSearch splits the model file into smaller chunks to be stored in a model index. When allocating machine learning (ML) or data nodes for your OpenSearch cluster, make sure you correctly size your ML nodes so that you have enough memory when making ML inferences.
## GPU acceleration
To achieve better performance within the model-serving framework, you can take advantage of GPU acceleration on your ML node. For more information, see [GPU acceleration]({{site.url}}{{site.baseurl}}/ml-commons-plugin/gpu-acceleration/).
`version` | string | The version number of the model. Since OpenSearch does not enforce a specific version schema for models, you can choose any number or format that makes sense for your models. |
`model_format` | string | The portable format of the model file. Currently only supports `TORCH_SCRIPT`. |
[`model_config`](#the-model_config-object) | json object | The model's configuration, including the `model_type`, `embedding_dimension`, and `framework_type`. |
| `model_type` | string | The model type, such as `bert`. For a Huggingface model, the model type is specified in `config.json`. For an example, see the [`all-MiniLM-L6-v2` Huggingface model `config.json`](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/blob/main/config.json#L15).|
| `embedding_dimension` | integer | The dimension of the model-generated dense vector. For a Huggingface model, the dimension is specified in the model card. For example, in the [`all-MiniLM-L6-v2` Huggingface model card](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2), the statement `384 dimensional dense vector space` specifies 384 as the embedding dimension. |
| `framework_type` | string | The framework the model is using. Currently, we support `sentence_transformers` and `huggingface_transformers` frameworks. The `sentence_transformers` model outputs text embeddings directly, so ML Commons does not perform any post processing. For `huggingface_transformers`, ML Commons performs post processing by applying mean pooling to get text embeddings. See the example [`all-MiniLM-L6-v2` Huggingface model](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) for more details. |
| `all_config`_(Optional)_ | string | This field is used for reference purposes. You can specify all model configurations in this field. For example, if you are using a Huggingface model, you can minify the `config.json` file to one line and save its contents in the `all_config` field. Once the model is uploaded, you can use the get model API operation to get all model configurations stored in this field. |
To see the status of your model upload, pass the `task_id` into the [task API]({{site.url}}{{site.baseurl}}/ml-commons-plugin/api#get-task-information).
The load model operation reads the model's chunks from the model index and then creates an instance of the model to load into memory. The bigger the model, the more chunks the model is split into. The more chunks a model index contains, the longer it takes for the model to load into memory.
To load a model, you need the `model_id`. To find the `model_id`, take the `task_id` from the model's upload operations API response and use the `GET _ml/tasks` API.
This example request uses the `task_id` from the upload example.
With the `model_id`, you can now load the model from the model's index in order to deploy the model to ML nodes. The load API reads model chunks from the model index, creates an instance of that model, and saves the model instance in the ML node's cache.
By default, the ML Commons setting `plugins.ml_commons.only_run_on_ml_node` is set to `false`. When `false`, models load on ML nodes first. If no ML nodes exist, models load on data nodes. When running ML models in production, set `plugins.ml_commons.only_run_on_ml_node` to `true` so that models only load on ML nodes.
#### Sample request: Load into any available ML node
If you want to reserve the memory of other ML nodes within your cluster, you can load your model into a specific node(s) by specifying each node's ID in the request body:
With your `task_id` from the load response, you can use the `GET _ml/tasks` API to see the load status of your model. Before a loaded model can be used for inferences, the load task's `state` must be `COMPLETED`.
After the model has been loaded, you can enter the `model_id` into the [predict API]({{site.url}}{{site.baseurl}}/ml-commons-plugin/api#predict) to perform inferences.
If you're done making predictions with your model, use the unload operation to remove the model from your memory cache. The model will remain accessible in the model index.
