Text Embeddings Inference on Habana Gaudi

Table of contents

• Get started
• Supported Models
  • Text Embeddings
  • Sequence Classification and Re-Ranking
• How to Use
  • Using Re-rankers models
  • Using Sequence Classification models
  • Using SPLADE pooling

Get started

To use 🤗 text-embeddings-inference on Habana Gaudi/Gaudi2, follow these steps:

1. Pull the official Docker image with:

   docker pull ghcr.io/huggingface/tei-gaudi:latest

Note

Alternatively, you can build the Docker image using Dockerfile-hpu located in this folder with:

docker build -f Dockerfile-hpu -t tei_gaudi .

2. Launch a local server instance on 1 Gaudi card:

   model=BAAI/bge-large-en-v1.5
   volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
   
   docker run -p 8080:80 -v $volume:/data --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e MAX_WARMUP_SEQUENCE_LENGTH=512 --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tei-gaudi:latest --model-id $model --pooling cls

   For models within the Transformers library that need remote code to run customized implementations, please set the environment variable -e TRUST_REMOTE_CODE=TRUE within docker run command line. Here is an example:

   model="Alibaba-NLP/gte-large-en-v1.5"
   volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
   
   docker run -p 8080:80 -v $volume:/data --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e MAX_WARMUP_SEQUENCE_LENGTH=512 -e TRUST_REMOTE_CODE=TRUE --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tei-gaudi:latest --model-id $model --pooling cls
   

3. You can then send a request:

    curl 127.0.0.1:8080/embed \
        -X POST \
        -d '{"inputs":"What is Deep Learning?"}' \
        -H 'Content-Type: application/json'

For more information and documentation about Text Embeddings Inference, checkout README of the original repo.

Supported Models

Text Embeddings

tei-gaudi currently supports Nomic, BERT, CamemBERT, XLM-RoBERTa models with absolute positions, JinaBERT model with Alibi positions and Mistral, Alibaba GTE and Qwen2 models with Rope positions.

Below are some examples of our validated models:

Architecture	Pooling	Models
BERT	Cls/Mean/Last token	BAAI/bge-large-en-v1.5 sentence-transformers/all-MiniLM-L6-v2 sentence-transformers/all-MiniLM-L12-v2 sentence-transformers/multi-qa-MiniLM-L6-cos-v1 sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 sentence-transformers/paraphrase-MiniLM-L3-v2
BERT	Splade	naver/efficient-splade-VI-BT-large-query
MPNet	Cls/Mean/Last token	sentence-transformers/all-mpnet-base-v2 sentence-transformers/paraphrase-multilingual-mpnet-base-v2 sentence-transformers/multi-qa-mpnet-base-dot-v1
ALBERT	Cls/Mean/Last token	sentence-transformers/paraphrase-albert-small-v2
Mistral	Cls/Mean/Last token	intfloat/e5-mistral-7b-instruct Salesforce/SFR-Embedding-2_R
GTE	Cls/Mean/Last token	Alibaba-NLP/gte-large-en-v1.5
JinaBERT	Cls/Mean/Last token	jinaai/jina-embeddings-v2-base-en

Sequence Classification and Re-Ranking

tei-gaudi currently supports CamemBERT, and XLM-RoBERTa Sequence Classification models with absolute positions.

Below are some examples of the currently supported models:

Task	Model Type	Model ID
Re-Ranking	XLM-RoBERTa	BAAI/bge-reranker-large
Re-Ranking	XLM-RoBERTa	BAAI/bge-reranker-base
Sentiment Analysis	RoBERTa	SamLowe/roberta-base-go_emotions

How to Use

Using Re-rankers models

model=BAAI/bge-reranker-large
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 -v $volume:/data --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e MAX_WARMUP_SEQUENCE_LENGTH=512 --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tei-gaudi:latest --model-id $model

And then you can rank the similarity between a query and a list of texts with:

curl 127.0.0.1:8080/rerank \
    -X POST \
    -d '{"query":"What is Deep Learning?", "texts": ["Deep Learning is not...", "Deep learning is..."]}' \
    -H 'Content-Type: application/json'

Using Sequence Classification models

You can also use classic Sequence Classification models like SamLowe/roberta-base-go_emotions:

model=SamLowe/roberta-base-go_emotions
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 -v $volume:/data --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e MAX_WARMUP_SEQUENCE_LENGTH=512 --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tei-gaudi:latest --model-id $model

Once you have deployed the model you can use the predict endpoint to get the emotions most associated with an input:

curl 127.0.0.1:8080/predict \
    -X POST \
    -d '{"inputs":"I like you."}' \
    -H 'Content-Type: application/json'

Using SPLADE pooling

You can choose to activate SPLADE pooling for Bert and Distilbert MaskedLM architectures:

docker build -f Dockerfile-hpu -t tei_gaudi .
model=naver/efficient-splade-VI-BT-large-query
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

docker run -p 8080:80 -v $volume:/data --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e MAX_WARMUP_SEQUENCE_LENGTH=512 --cap-add=sys_nice --ipc=host tei_gaudi --model-id $model --pooling splade

Once you have deployed the model you can use the /embed_sparse endpoint to get the sparse embedding:

curl 127.0.0.1:8080/embed_sparse \
    -X POST \
    -d '{"inputs":"I like you."}' \
    -H 'Content-Type: application/json'

The license to use TEI on Habana Gaudi is the one of TEI: https://github.com/huggingface/text-embeddings-inference/blob/main/LICENSE

Please reach out to [email protected] if you have any question.