Custom local models

Introduced 2.9

To use a custom model locally, you can upload it to the OpenSearch cluster.

Model support

As of OpenSearch 2.6, OpenSearch supports local text embedding models.

As of OpenSearch 2.11, OpenSearch supports local sparse encoding models.

As of OpenSearch 2.12, OpenSearch supports local cross-encoder models.

As of OpenSearch 2.13, OpenSearch supports local question answering models.

Running local models on the CentOS 7 operating system is not supported. Moreover, not all local models can run on all hardware and operating systems.

Preparing a model

For all the models, you must provide a tokenizer JSON file within the model zip file.

For sparse encoding models, make sure your output format is {"output":<sparse_vector>} so that ML Commons can post-process the sparse vector.

If you fine-tune a sparse model on your own dataset, you may also want to use your own sparse tokenizer model. It is preferable to provide your own IDF JSON file in the tokenizer model zip file because this increases query performance when you use the tokenizer model in the query. Alternatively, you can use an OpenSearch-provided generic IDF from MSMARCO. If the IDF file is not provided, the default weight of each token is set to 1, which may influence sparse neural search performance.

Model format

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 and ONNX formats.

You must save the model file as zip before uploading it to OpenSearch. To ensure that ML Commons can upload your model, compress your TorchScript file before uploading. For an example, download a TorchScript model file.

Additionally, you must calculate a SHA256 checksum for the model zip file that you’ll need to provide when registering the model. For example, on UNIX, use the following command to obtain the checksum:

  1. shasum -a 256 sentence-transformers_paraphrase-mpnet-base-v2-1.0.0-onnx.zip

Model size

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 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.

Prerequisites

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 Hugging Face, or train a new model in accordance with your needs.

Cluster settings

This example uses a simple setup with no dedicated ML nodes and allows running a model on a non-ML node.

On clusters with dedicated ML nodes, specify "only_run_on_ml_node": "true" for improved performance. For more information, see ML Commons cluster settings.

To ensure that this basic local setup works, specify the following cluster settings:

  1. PUT _cluster/settings
  2. {
  3.   "persistent": {
  4.     "plugins": {
  5.       "ml_commons": {
  6.         "allow_registering_model_via_url": "true",
  7.         "only_run_on_ml_node": "false",
  8.         "model_access_control_enabled": "true",
  9.         "native_memory_threshold": "99"
  10.       }
  11.     }
  12.   }
  13. }

copy

Step 1: Register a model group

To register a model, you have the following options:

  • You can use model_group_id to register a model version to an existing model group.
  • If you do not use model_group_id, ML Commons creates a model with a new model group.

To register a model group, send the following request:

  1. POST /_plugins/_ml/model_groups/_register
  2. {
  3.   "name": "local_model_group",
  4.   "description": "A model group for local models"
  5. }

copy

The response contains the model group ID that you’ll use to register a model to this model group:

  1. {
  2.  "model_group_id": "wlcnb4kBJ1eYAeTMHlV6",
  3.  "status": "CREATED"
  4. }

To learn more about model groups, see Model access control.

Step 2: Register a local model

To register a local model to the model group created in step 1, send a Register Model API request. For descriptions of Register Model API parameters, see Register a model.

The function_name corresponds to the model type. For text embedding models, set this parameter to TEXT_EMBEDDING. For sparse encoding models, set this parameter to SPARSE_ENCODING or SPARSE_TOKENIZE. For cross-encoder models, set this parameter to TEXT_SIMILARITY. For question answering models, set this parameter to QUESTION_ANSWERING. In this example, set function_name to TEXT_EMBEDDING because you’re registering a text embedding model.

Provide the model group ID from step 1 and send the following request:

  1. POST /_plugins/_ml/models/_register
  2. {
  3.   "name": "huggingface/sentence-transformers/msmarco-distilbert-base-tas-b",
  4.   "version": "1.0.1",
  5.   "model_group_id": "wlcnb4kBJ1eYAeTMHlV6",
  6.   "description": "This is a port of the DistilBert TAS-B Model to sentence-transformers model: It maps sentences & paragraphs to a 768 dimensional dense vector space and is optimized for the task of semantic search.",
  7.   "function_name": "TEXT_EMBEDDING",
  8.   "model_format": "TORCH_SCRIPT",
  9.   "model_content_size_in_bytes": 266352827,
  10.   "model_content_hash_value": "acdc81b652b83121f914c5912ae27c0fca8fabf270e6f191ace6979a19830413",
  11.   "model_config": {
  12.     "model_type": "distilbert",
  13.     "embedding_dimension": 768,
  14.     "framework_type": "sentence_transformers",
  15.     "all_config": "{\"_name_or_path\":\"old_models/msmarco-distilbert-base-tas-b/0_Transformer\",\"activation\":\"gelu\",\"architectures\":[\"DistilBertModel\"],\"attention_dropout\":0.1,\"dim\":768,\"dropout\":0.1,\"hidden_dim\":3072,\"initializer_range\":0.02,\"max_position_embeddings\":512,\"model_type\":\"distilbert\",\"n_heads\":12,\"n_layers\":6,\"pad_token_id\":0,\"qa_dropout\":0.1,\"seq_classif_dropout\":0.2,\"sinusoidal_pos_embds\":false,\"tie_weights_\":true,\"transformers_version\":\"4.7.0\",\"vocab_size\":30522}"
  16.   },
  17.   "created_time": 1676073973126,
  18.   "url": "https://artifacts.opensearch.org/models/ml-models/huggingface/sentence-transformers/msmarco-distilbert-base-tas-b/1.0.1/torch_script/sentence-transformers_msmarco-distilbert-base-tas-b-1.0.1-torch_script.zip"
  19. }

copy

Note that in OpenSearch Dashboards, wrapping the all_config field contents in triple quotes (""") automatically escapes quotation marks within the field and provides better readability:

  1. POST /_plugins/_ml/models/_register
  2. {
  3.   "name": "huggingface/sentence-transformers/msmarco-distilbert-base-tas-b",
  4.   "version": "1.0.1",
  5.   "model_group_id": "wlcnb4kBJ1eYAeTMHlV6",
  6.   "description": "This is a port of the DistilBert TAS-B Model to sentence-transformers model: It maps sentences & paragraphs to a 768 dimensional dense vector space and is optimized for the task of semantic search.",
  7.   "function_name": "TEXT_EMBEDDING",
  8.   "model_format": "TORCH_SCRIPT",
  9.   "model_content_size_in_bytes": 266352827,
  10.   "model_content_hash_value": "acdc81b652b83121f914c5912ae27c0fca8fabf270e6f191ace6979a19830413",
  11.   "model_config": {
  12.     "model_type": "distilbert",
  13.     "embedding_dimension": 768,
  14.     "framework_type": "sentence_transformers",
  15.     "all_config": """{"_name_or_path":"old_models/msmarco-distilbert-base-tas-b/0_Transformer","activation":"gelu","architectures":["DistilBertModel"],"attention_dropout":0.1,"dim":768,"dropout":0.1,"hidden_dim":3072,"initializer_range":0.02,"max_position_embeddings":512,"model_type":"distilbert","n_heads":12,"n_layers":6,"pad_token_id":0,"qa_dropout":0.1,"seq_classif_dropout":0.2,"sinusoidal_pos_embds":false,"tie_weights_":true,"transformers_version":"4.7.0","vocab_size":30522}"""
  16.   },
  17.   "created_time": 1676073973126,
  18.   "url": "https://artifacts.opensearch.org/models/ml-models/huggingface/sentence-transformers/msmarco-distilbert-base-tas-b/1.0.1/torch_script/sentence-transformers_msmarco-distilbert-base-tas-b-1.0.1-torch_script.zip"
  19. }

copy

OpenSearch returns the task ID of the register operation:

  1. {
  2.   "task_id": "cVeMb4kBJ1eYAeTMFFgj",
  3.   "status": "CREATED"
  4. }

To check the status of the operation, provide the task ID to the Get task:

  1. GET /_plugins/_ml/tasks/cVeMb4kBJ1eYAeTMFFgj

copy

When the operation is complete, the state changes to COMPLETED:

  1. {
  2.   "model_id": "cleMb4kBJ1eYAeTMFFg4",
  3.   "task_type": "REGISTER_MODEL",
  4.   "function_name": "TEXT_EMBEDDING",
  5.   "state": "COMPLETED",
  6.   "worker_node": [
  7.     "XPcXLV7RQoi5m8NI_jEOVQ"
  8.   ],
  9.   "create_time": 1689793598499,
  10.   "last_update_time": 1689793598530,
  11.   "is_async": false
  12. }

Take note of the returned model_id because you’ll need it to deploy the model.

Step 3: Deploy the model

The deploy 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 and longer it takes for the model to load into memory.

To deploy the registered model, provide its model ID from step 3 in the following request:

  1. POST /_plugins/_ml/models/cleMb4kBJ1eYAeTMFFg4/_deploy

copy

The response contains the task ID that you can use to check the status of the deploy operation:

  1. {
  2.   "task_id": "vVePb4kBJ1eYAeTM7ljG",
  3.   "status": "CREATED"
  4. }

As in the previous step, check the status of the operation by calling the Tasks API:

  1. GET /_plugins/_ml/tasks/vVePb4kBJ1eYAeTM7ljG

copy

When the operation is complete, the state changes to COMPLETED:

  1. {
  2.   "model_id": "cleMb4kBJ1eYAeTMFFg4",
  3.   "task_type": "DEPLOY_MODEL",
  4.   "function_name": "TEXT_EMBEDDING",
  5.   "state": "COMPLETED",
  6.   "worker_node": [
  7.     "n-72khvBTBi3bnIIR8FTTw"
  8.   ],
  9.   "create_time": 1689793851077,
  10.   "last_update_time": 1689793851101,
  11.   "is_async": true
  12. }

If a cluster or node is restarted, then you need to redeploy the model. To learn how to set up automatic redeployment, see Enable auto redeploy.

Step 4 (Optional): Test the model

Use the Predict API to test the model.

For a text embedding model, send the following request:

  1. POST /_plugins/_ml/_predict/text_embedding/cleMb4kBJ1eYAeTMFFg4
  2. {
  3.   "text_docs":[ "today is sunny"],
  4.   "return_number": true,
  5.   "target_response": ["sentence_embedding"]
  6. }

copy

The response contains text embeddings for the provided sentence:

  1. {
  2.   "inference_results" : [
  3.     {
  4.       "output" : [
  5.         {
  6.           "name" : "sentence_embedding",
  7.           "data_type" : "FLOAT32",
  8.           "shape" : [
  9.             768
  10.           ],
  11.           "data" : [
  12.             0.25517133,
  13.             -0.28009856,
  14.             0.48519906,
  15.             ...
  16.           ]
  17.         }
  18.       ]
  19.     }
  20.   ]
  21. }

For a sparse encoding model, send the following request:

  1. POST /_plugins/_ml/_predict/sparse_encoding/cleMb4kBJ1eYAeTMFFg4
  2. {
  3.   "text_docs":[ "today is sunny"]
  4. }

copy

The response contains the tokens and weights:

  1. {
  2.   "inference_results": [
  3.     {
  4.       "output": [
  5.         {
  6.           "name": "output",
  7.           "dataAsMap": {
  8.             "response": [
  9.               {
  10.                 "saturday": 0.48336542,
  11.                 "week": 0.1034762,
  12.                 "mood": 0.09698499,
  13.                 "sunshine": 0.5738209,
  14.                 "bright": 0.1756877,
  15.                 ...
  16.               }
  17.           }
  18.         }
  19.     }
  20. }

To learn how to use the model for vector search, see Using an ML model for neural search.

Question answering models

A question answering model extracts the answer to a question from a given context. ML Commons supports context in text format.

To register a question answering model, send a request in the following format. Specify the function_name as QUESTION_ANSWERING:

  1. POST /_plugins/_ml/models/_register
  2. {
  3.     "name": "question_answering",
  4.     "version": "1.0.0",
  5.     "function_name": "QUESTION_ANSWERING",
  6.     "description": "test model",
  7.     "model_format": "TORCH_SCRIPT",
  8.     "model_group_id": "lN4AP40BKolAMNtR4KJ5",
  9.     "model_content_hash_value": "e837c8fc05fd58a6e2e8383b319257f9c3859dfb3edc89b26badfaf8a4405ff6",
  10.     "model_config": { 
  11.         "model_type": "bert",
  12.         "framework_type": "huggingface_transformers"
  13.     },
  14.     "url": "https://github.com/opensearch-project/ml-commons/blob/main/ml-algorithms/src/test/resources/org/opensearch/ml/engine/algorithms/question_answering/question_answering_pt.zip?raw=true"
  15. }

copy

Then send a request to deploy the model:

  1. POST _plugins/_ml/models/<model_id>/_deploy

copy

To test a question answering model, send the following request. It requires a question and the relevant context from which the answer will be generated:

  1. POST /_plugins/_ml/_predict/question_answering/<model_id>
  2. {
  3.   "question": "Where do I live?"
  4.   "context": "My name is John. I live in New York"
  5. }

copy

The response provides the answer based on the context:

  1. {
  2.   "inference_results": [
  3.     {
  4.       "output": [
  5.         {
  6.           "result": "New York"
  7.         }
  8.     }
  9. }