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 when setting the host_ip value (do not use localhost).

export host_ip="External_Public_IP"

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

export no_proxy=${no_proxy},${host_ip}

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 REDIS_DB_PORT=6379
export REDIS_INSIGHTS_PORT=8001
export REDIS_URL="redis://${host_ip}:${REDIS_DB_PORT}"
export REDIS_HOST=${host_ip}
export INDEX_NAME="mm-rag-redis"
export WHISPER_PORT=7066
export WHISPER_SERVER_ENDPOINT="http://${host_ip}:${WHISPER_PORT}/v1/asr"
export MAX_IMAGES=1
export WHISPER_MODEL="base"
export DATAPREP_MMR_PORT=6007
export DATAPREP_INGEST_SERVICE_ENDPOINT="http://${host_ip}:${DATAPREP_MMR_PORT}/v1/dataprep/ingest"
export DATAPREP_GEN_TRANSCRIPT_SERVICE_ENDPOINT="http://${host_ip}:${DATAPREP_MMR_PORT}/v1/dataprep/generate_transcripts"
export DATAPREP_GEN_CAPTION_SERVICE_ENDPOINT="http://${host_ip}:${DATAPREP_MMR_PORT}/v1/dataprep/generate_captions"
export DATAPREP_GET_FILE_ENDPOINT="http://${host_ip}:${DATAPREP_MMR_PORT}/v1/dataprep/get"
export DATAPREP_DELETE_FILE_ENDPOINT="http://${host_ip}:${DATAPREP_MMR_PORT}/v1/dataprep/delete"
export EMM_BRIDGETOWER_PORT=6006
export EMBEDDING_MODEL_ID="BridgeTower/bridgetower-large-itm-mlm-itc"
export BRIDGE_TOWER_EMBEDDING=true
export MMEI_EMBEDDING_ENDPOINT="http://${host_ip}:$EMM_BRIDGETOWER_PORT"
export MM_EMBEDDING_PORT_MICROSERVICE=6000
export REDIS_RETRIEVER_PORT=7000
export LVM_PORT=9399
export LLAVA_SERVER_PORT=8399
export TGI_GAUDI_PORT="${LLAVA_SERVER_PORT}:80"
export LVM_MODEL_ID="llava-hf/llava-v1.6-vicuna-13b-hf"
export LVM_ENDPOINT="http://${host_ip}:${LLAVA_SERVER_PORT}"
export MEGA_SERVICE_PORT=8888
export BACKEND_SERVICE_ENDPOINT="http://${host_ip}:${MEGA_SERVICE_PORT}/v1/multimodalqna"
export UI_PORT=5173
export UI_TIMEOUT=240

Note: The MAX_IMAGES environment variable is used to specify the maximum number of images that will be sent from the LVM service to the LLaVA server. If an image list longer than MAX_IMAGES is sent to the LVM server, a shortened image list will be sent to the LLaVA service. If the image list needs to be shortened, the most recent images (the ones at the end of the list) are prioritized to send to the LLaVA service. Some LLaVA models have not been trained with multiple images and may lead to inaccurate results. If MAX_IMAGES is not set, it will default to 1.

🚀 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=$EMM_BRIDGETOWER_PORT --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/third_parties/bridgetower/src/Dockerfile .

Build embedding microservice image

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

2. Build retriever-multimodal-redis Image

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

3. Build LVM Images

Build TGI Gaudi image

docker pull ghcr.io/huggingface/tgi-gaudi:2.3.1

Build lvm microservice image

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

4. Build dataprep-multimodal-redis Image

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

5. Build Whisper Server Image

Build whisper server image

docker build --no-cache -t opea/whisper:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/third_parties/whisper/src/Dockerfile .

6. Build TTS Server Image

Build TTS server image

docker build --no-cache -t opea/speecht5:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/third_parties/speecht5/src/Dockerfile .

7. 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 .

8. 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 11 Docker Images:

  1. opea/dataprep:latest

  2. opea/lvm:latest

  3. ghcr.io/huggingface/tgi-gaudi:2.3.1

  4. opea/retriever:latest

  5. opea/whisper:latest

  6. opea/speech5:latest

  7. opea/redis-vector-db

  8. opea/embedding:latest

  9. opea/embedding-multimodal-bridgetower:latest

  10. opea/multimodalqna:latest

  11. 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

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

Alternatively, you can run docker compose with compose_milvus.yaml to use the Milvus vector database:

export MILVUS_HOST=${host_ip}
export MILVUS_PORT=19530
export MILVUS_RETRIEVER_PORT=7000
export COLLECTION_NAME=mm_rag_milvus
cd GenAIExamples/MultimodalQnA/docker_compose/intel/hpu/gaudi/
docker compose -f compose_milvus.yaml up -d

Validate Microservices

  1. embedding-multimodal-bridgetower

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

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

  1. embedding

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}:${REDIS_RETRIEVER_PORT}/v1/retrieval \
    -X POST \
    -H "Content-Type: application/json" \
    -d "{\"text\":\"test\",\"embedding\":${your_embedding}}"

  1. whisper

curl ${WHISPER_SERVER_ENDPOINT} \
    -X POST \
    -H "Content-Type: application/json" \
    -d '{"audio" : "UklGRigAAABXQVZFZm10IBIAAAABAAEARKwAAIhYAQACABAAAABkYXRhAgAAAAEA"}'

  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. tts

curl ${TTS_ENDPOINT} \
  -X POST \
  -d '{"text": "Who are you?"}' \
  -H 'Content-Type: application/json'

  1. lvm

curl http://${host_ip}:${LVM_PORT}/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://${host_ip}:${LVM_PORT}/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://${host_ip}:${LVM_PORT}/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 (.mp4), image (.png, .gif, .jpg), pdf, and audio file (.wav, .mp3) 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 pdf_fn="nke-10k-2023.pdf"
wget https://raw.githubusercontent.com/opea-project/GenAIComps/v1.3/comps/third_parties/pathway/src/data/nke-10k-2023.pdf -O ${pdf_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 or .mp3 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

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 and a PDF containing images and text. The image caption can be provided as a text (.txt) or as spoken audio (.wav or .mp3).

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=@./${caption_fn}" \
    -F "files=@./${pdf_fn}"

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

curl -X POST \
    -H "Content-Type: application/json" \
    -d '{"file_path": "all"}' \
    ${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",
    "nke-10k-2023_28000757-5533-4b1b-89fe-7c0a1b7e2cd0.pdf",
    "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" \
    -d '{"file_path": "all"}' \
    ${DATAPREP_DELETE_FILE_ENDPOINT}

  1. MegaService

Test the MegaService with a text query:

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

Test the MegaService with an audio query:

curl http://${host_ip}:${MEGA_SERVICE_PORT}/v1/multimodalqna  \
    -H "Content-Type: application/json"  \
    -d '{"messages": [{"role": "user", "content": [{"type": "audio", "audio": "UklGRigAAABXQVZFZm10IBIAAAABAAEARKwAAIhYAQACABAAAABkYXRhAgAAAAEA"}]}]}'

Test the MegaService with a text and image query:

curl http://${host_ip}:${MEGA_SERVICE_PORT}/v1/multimodalqna \
    -H "Content-Type: application/json" \
    -d  '{"messages": [{"role": "user", "content": [{"type": "text", "text": "Green bananas in a tree"}, {"type": "image_url", "image_url": {"url": "http://images.cocodataset.org/test-stuff2017/000000004248.jpg"}}]}]}'

Test the MegaService with a back and forth conversation between the user and assistant including a text to speech response from the assistant using "modalities": ["text", "audio"]':

curl http://${host_ip}:${MEGA_SERVICE_PORT}/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, "modalities": ["text", "audio"]}'