Build Mega Service of MultimodalQnA on Gaudi

This document outlines the deployment process for a MultimodalQnA application utilizing the GenAIComps microservice pipeline on Intel Gaudi server. The steps include Docker image creation, container deployment via Docker Compose, and service execution to integrate microservices such as multimodal_embedding that employs BridgeTower model as embedding model, multimodal_retriever, lvm, and multimodal-data-prep. We will publish the Docker images to Docker Hub soon, it will simplify the deployment process for this service.

Setup Environment Variables

Since the compose.yaml will consume some environment variables, you need to setup them in advance as below.

Export the value of the public IP address of your Gaudi server to the host_ip environment variable

Change the External_Public_IP below with the actual IPV4 value

export host_ip="External_Public_IP"

Append the value of the public IP address to the no_proxy list

export your_no_proxy=${your_no_proxy},"External_Public_IP"

export no_proxy=${your_no_proxy}
export http_proxy=${your_http_proxy}
export https_proxy=${your_http_proxy}
export EMBEDDER_PORT=6006
export MMEI_EMBEDDING_ENDPOINT="http://${host_ip}:$EMBEDDER_PORT/v1/encode"
export MM_EMBEDDING_PORT_MICROSERVICE=6000
export REDIS_URL="redis://${host_ip}:6379"
export REDIS_HOST=${host_ip}
export INDEX_NAME="mm-rag-redis"
export LLAVA_SERVER_PORT=8399
export LVM_ENDPOINT="http://${host_ip}:8399"
export EMBEDDING_MODEL_ID="BridgeTower/bridgetower-large-itm-mlm-itc"
export LVM_MODEL_ID="llava-hf/llava-v1.6-vicuna-13b-hf"
export WHISPER_MODEL="base"
export MM_EMBEDDING_SERVICE_HOST_IP=${host_ip}
export MM_RETRIEVER_SERVICE_HOST_IP=${host_ip}
export LVM_SERVICE_HOST_IP=${host_ip}
export MEGA_SERVICE_HOST_IP=${host_ip}
export BACKEND_SERVICE_ENDPOINT="http://${host_ip}:8888/v1/multimodalqna"
export DATAPREP_INGEST_SERVICE_ENDPOINT="http://${host_ip}:6007/v1/ingest_with_text"
export DATAPREP_GEN_TRANSCRIPT_SERVICE_ENDPOINT="http://${host_ip}:6007/v1/generate_transcripts"
export DATAPREP_GEN_CAPTION_SERVICE_ENDPOINT="http://${host_ip}:6007/v1/generate_captions"
export DATAPREP_GET_FILE_ENDPOINT="http://${host_ip}:6007/v1/dataprep/get_files"
export DATAPREP_DELETE_FILE_ENDPOINT="http://${host_ip}:6007/v1/dataprep/delete_files"

Note: Please replace with host_ip with you external IP address, do not use localhost.

🚀 Build Docker Images

First of all, you need to build Docker Images locally and install the python package of it.

1. Build embedding-multimodal-bridgetower Image

Build embedding-multimodal-bridgetower docker image

git clone https://github.com/opea-project/GenAIComps.git
cd GenAIComps
docker build --no-cache -t opea/embedding-multimodal-bridgetower:latest --build-arg EMBEDDER_PORT=$EMBEDDER_PORT --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/embeddings/multimodal/bridgetower/Dockerfile .

Build embedding-multimodal microservice image

docker build --no-cache -t opea/embedding-multimodal:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/embeddings/multimodal/multimodal_langchain/Dockerfile .

2. Build retriever-multimodal-redis Image

docker build --no-cache -t opea/retriever-multimodal-redis:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/retrievers/multimodal/redis/langchain/Dockerfile .

3. Build LVM Images

Build TGI Gaudi image

docker pull ghcr.io/huggingface/tgi-gaudi:2.0.6

Build lvm-tgi microservice image

docker build --no-cache -t opea/lvm-tgi:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/lvms/tgi-llava/Dockerfile .

4. Build dataprep-multimodal-redis Image

docker build --no-cache -t opea/dataprep-multimodal-redis:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/dataprep/multimodal/redis/langchain/Dockerfile .

5. Build MegaService Docker Image

To construct the Mega Service, we utilize the GenAIComps microservice pipeline within the multimodalqna.py Python script. Build MegaService Docker image via below command:

git clone https://github.com/opea-project/GenAIExamples.git
cd GenAIExamples/MultimodalQnA
docker build --no-cache -t opea/multimodalqna:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f Dockerfile .

6. Build UI Docker Image

Build frontend Docker image via below command:

cd  GenAIExamples/MultimodalQnA/ui/
docker build --no-cache -t opea/multimodalqna-ui:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f ./docker/Dockerfile .

Then run the command docker images, you will have the following 8 Docker Images:

  1. opea/dataprep-multimodal-redis:latest

  2. opea/lvm-tgi:latest

  3. ghcr.io/huggingface/tgi-gaudi:2.0.6

  4. opea/retriever-multimodal-redis:latest

  5. opea/embedding-multimodal:latest

  6. opea/embedding-multimodal-bridgetower:latest

  7. opea/multimodalqna:latest

  8. opea/multimodalqna-ui:latest

🚀 Start Microservices

Required Models

By default, the multimodal-embedding and LVM models are set to a default value as listed below:

Service

Model

embedding-multimodal

BridgeTower/bridgetower-large-itm-mlm-gaudi

LVM

llava-hf/llava-v1.6-vicuna-13b-hf

Start all the services Docker Containers

Before running the docker compose command, you need to be in the folder that has the docker compose yaml file

cd GenAIExamples/MultimodalQnA/docker_compose/intel/hpu/gaudi/
docker compose -f compose.yaml up -d

Validate Microservices

  1. embedding-multimodal-bridgetower

curl http://${host_ip}:${EMBEDDER_PORT}/v1/encode \
     -X POST \
     -H "Content-Type:application/json" \
     -d '{"text":"This is example"}'

curl http://${host_ip}:${EMBEDDER_PORT}/v1/encode \
     -X POST \
     -H "Content-Type:application/json" \
     -d '{"text":"This is example", "img_b64_str": "iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mP8/5+hnoEIwDiqkL4KAcT9GO0U4BxoAAAAAElFTkSuQmCC"}'

  1. embedding-multimodal

curl http://${host_ip}:$MM_EMBEDDING_PORT_MICROSERVICE/v1/embeddings \
    -X POST \
    -H "Content-Type: application/json" \
    -d '{"text" : "This is some sample text."}'

curl http://${host_ip}:$MM_EMBEDDING_PORT_MICROSERVICE/v1/embeddings \
    -X POST \
    -H "Content-Type: application/json" \
    -d '{"text": {"text" : "This is some sample text."}, "image" : {"url": "https://github.com/docarray/docarray/blob/main/tests/toydata/image-data/apple.png?raw=true"}}'

  1. retriever-multimodal-redis

export your_embedding=$(python3 -c "import random; embedding = [random.uniform(-1, 1) for _ in range(512)]; print(embedding)")
curl http://${host_ip}:7000/v1/multimodal_retrieval \
    -X POST \
    -H "Content-Type: application/json" \
    -d "{\"text\":\"test\",\"embedding\":${your_embedding}}"

  1. TGI LLaVA Gaudi Server

curl http://${host_ip}:${LLAVA_SERVER_PORT}/generate \
    -X POST \
    -d '{"inputs":"![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png)What is this a picture of?\n\n","parameters":{"max_new_tokens":16, "seed": 42}}' \
    -H 'Content-Type: application/json'

  1. lvm-tgi

curl http://${host_ip}:9399/v1/lvm \
    -X POST \
    -H 'Content-Type: application/json' \
    -d '{"retrieved_docs": [], "initial_query": "What is this?", "top_n": 1, "metadata": [{"b64_img_str": "iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mP8/5+hnoEIwDiqkL4KAcT9GO0U4BxoAAAAAElFTkSuQmCC", "transcript_for_inference": "yellow image", "video_id": "8c7461df-b373-4a00-8696-9a2234359fe0", "time_of_frame_ms":"37000000", "source_video":"WeAreGoingOnBullrun_8c7461df-b373-4a00-8696-9a2234359fe0.mp4"}], "chat_template":"The caption of the image is: '\''{context}'\''. {question}"}'

curl http://${host_ip}:9399/v1/lvm  \
    -X POST \
    -H 'Content-Type: application/json' \
    -d '{"image": "iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mP8/5+hnoEIwDiqkL4KAcT9GO0U4BxoAAAAAElFTkSuQmCC", "prompt":"What is this?"}'

Also, validate LVM TGI Gaudi Server with empty retrieval results

curl http://${host_ip}:9399/v1/lvm \
    -X POST \
    -H 'Content-Type: application/json' \
    -d '{"retrieved_docs": [], "initial_query": "What is this?", "top_n": 1, "metadata": [], "chat_template":"The caption of the image is: '\''{context}'\''. {question}"}'

  1. Multimodal Dataprep Microservice

Download a sample video, image, and audio file and create a caption

export video_fn="WeAreGoingOnBullrun.mp4"
wget http://commondatastorage.googleapis.com/gtv-videos-bucket/sample/WeAreGoingOnBullrun.mp4 -O ${video_fn}

export image_fn="apple.png"
wget https://github.com/docarray/docarray/blob/main/tests/toydata/image-data/apple.png?raw=true -O ${image_fn}

export caption_fn="apple.txt"
echo "This is an apple."  > ${caption_fn}

export audio_fn="AudioSample.wav"
wget https://github.com/intel/intel-extension-for-transformers/raw/main/intel_extension_for_transformers/neural_chat/assets/audio/sample.wav -O ${audio_fn}

Test dataprep microservice with generating transcript. This command updates a knowledge base by uploading a local video .mp4 and an audio .wav file.

curl --silent --write-out "HTTPSTATUS:%{http_code}" \
    ${DATAPREP_GEN_TRANSCRIPT_SERVICE_ENDPOINT} \
    -H 'Content-Type: multipart/form-data' \
    -X POST \
    -F "files=@./${video_fn}" \
    -F "files=@./${audio_fn}"

Also, test dataprep microservice with generating an image caption using lvm-tgi

curl --silent --write-out "HTTPSTATUS:%{http_code}" \
    ${DATAPREP_GEN_CAPTION_SERVICE_ENDPOINT} \
    -H 'Content-Type: multipart/form-data' \
    -X POST -F "files=@./${image_fn}"

Now, test the microservice with posting a custom caption along with an image

curl --silent --write-out "HTTPSTATUS:%{http_code}" \
    ${DATAPREP_INGEST_SERVICE_ENDPOINT} \
    -H 'Content-Type: multipart/form-data' \
    -X POST -F "files=@./${image_fn}" -F "files=@./${caption_fn}"

Also, you are able to get the list of all files that you uploaded:

curl -X POST \
    -H "Content-Type: application/json" \
    ${DATAPREP_GET_FILE_ENDPOINT}

Then you will get the response python-style LIST like this. Notice the name of each uploaded file e.g., videoname.mp4 will become videoname_uuid.mp4 where uuid is a unique ID for each uploaded file. The same files that are uploaded twice will have different uuid.

[
    "WeAreGoingOnBullrun_7ac553a1-116c-40a2-9fc5-deccbb89b507.mp4",
    "WeAreGoingOnBullrun_6d13cf26-8ba2-4026-a3a9-ab2e5eb73a29.mp4",
    "apple_fcade6e6-11a5-44a2-833a-3e534cbe4419.png",
    "AudioSample_976a85a6-dc3e-43ab-966c-9d81beef780c.wav
]

To delete all uploaded files along with data indexed with $INDEX_NAME in REDIS.

curl -X POST \
    -H "Content-Type: application/json" \
    ${DATAPREP_DELETE_FILE_ENDPOINT}

  1. MegaService

curl http://${host_ip}:8888/v1/multimodalqna \
    -H "Content-Type: application/json" \
    -X POST \
    -d '{"messages": "What is the revenue of Nike in 2023?"}'

curl http://${host_ip}:8888/v1/multimodalqna \
	-H "Content-Type: application/json" \
	-d '{"messages": [{"role": "user", "content": [{"type": "text", "text": "hello, "}, {"type": "image_url", "image_url": {"url": "https://www.ilankelman.org/stopsigns/australia.jpg"}}]}, {"role": "assistant", "content": "opea project! "}, {"role": "user", "content": "chao, "}], "max_tokens": 10}'