[ChatQnA] Switch to vLLM as default llm backend on Gaudi (#1404)

Switching from TGI to vLLM as the default LLM serving backend on Gaudi for the ChatQnA example to enhance the perf. 

#1213
Signed-off-by: Wang, Kai Lawrence <kai.lawrence.wang@intel.com>

Author: wangkl2

ChatQnA/README.md

@@ -202,7 +202,7 @@ Gaudi default compose.yaml
 | Embedding | Langchain | Xeon | 6000 | /v1/embeddings |
 | Retriever | Langchain, Redis | Xeon | 7000 | /v1/retrieval |
 | Reranking | Langchain, TEI | Gaudi | 8000 | /v1/reranking |
-| LLM | Langchain, TGI | Gaudi | 9000 | /v1/chat/completions |
+| LLM | Langchain, vLLM | Gaudi | 9000 | /v1/chat/completions |
 | Dataprep | Redis, Langchain | Xeon | 6007 | /v1/dataprep |
 
 ### Required Models
@@ -266,16 +266,21 @@ Refer to the [Intel Technology enabling for Openshift readme](https://github.com
 
 ### Check Service Status
 
-Before consuming ChatQnA Service, make sure the TGI/vLLM service is ready (which takes up to 2 minutes to start).
+Before consuming ChatQnA Service, make sure the vLLM/TGI service is ready, which takes some time.
 
 ```bash
+# vLLM example
+docker logs vllm-gaudi-server 2>&1 | grep complete
 # TGI example
-docker logs tgi-service | grep Connected
+docker logs tgi-gaudi-server | grep Connected
 ```
 
-Consume ChatQnA service until you get the TGI response like below.
+Consume ChatQnA service until you get the response like below.
 
 ```log
+# vLLM
+INFO: Application startup complete.
+# TGI
 2024-09-03T02:47:53.402023Z  INFO text_generation_router::server: router/src/server.rs:2311: Connected
 ```
 

ChatQnA/docker_compose/intel/hpu/gaudi/README.md

@@ -1,6 +1,8 @@
 # Build MegaService of ChatQnA on Gaudi
 
-This document outlines the deployment process for a ChatQnA application utilizing the [GenAIComps](https://github.com/opea-project/GenAIComps.git) 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 embedding, retriever, rerank, and llm. We will publish the Docker images to Docker Hub, it will simplify the deployment process for this service.
+This document outlines the deployment process for a ChatQnA application utilizing the [GenAIComps](https://github.com/opea-project/GenAIComps.git) 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 `embedding`, `retriever`, `rerank`, and `llm`.
+
+The default pipeline deploys with vLLM as the LLM serving component and leverages rerank component. It also provides options of not using rerank in the pipeline, leveraging guardrails, or using TGI backend for LLM microservice, please refer to [start-all-the-services-docker-containers](#start-all-the-services-docker-containers) section in this page.
 
 Quick Start:
 
@@ -184,15 +186,18 @@ By default, the embedding, reranking and LLM models are set to a default value a
 
 Change the `xxx_MODEL_ID` below for your needs.
 
-For users in China who are unable to download models directly from Huggingface, you can use [ModelScope](https://www.modelscope.cn/models) or a Huggingface mirror to download models. TGI can load the models either online or offline as described below:
+For users in China who are unable to download models directly from Huggingface, you can use [ModelScope](https://www.modelscope.cn/models) or a Huggingface mirror to download models. The vLLM/TGI can load the models either online or offline as described below:
 
 1. Online
 
    ```bash
    export HF_TOKEN=${your_hf_token}
    export HF_ENDPOINT="https://hf-mirror.com"
    model_name="Intel/neural-chat-7b-v3-3"
-   docker run -p 8008:80 -v ./data:/data --name tgi-service -e HF_ENDPOINT=$HF_ENDPOINT -e http_proxy=$http_proxy -e https_proxy=$https_proxy --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e HUGGING_FACE_HUB_TOKEN=$HF_TOKEN -e ENABLE_HPU_GRAPH=true -e LIMIT_HPU_GRAPH=true -e USE_FLASH_ATTENTION=true -e FLASH_ATTENTION_RECOMPUTE=true --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tgi-gaudi:2.0.6 --model-id $model_name --max-input-tokens 1024 --max-total-tokens 2048
+   # Start vLLM LLM Service
+   docker run -p 8007:80 -v ./data:/data --name vllm-gaudi-server -e HF_ENDPOINT=$HF_ENDPOINT -e http_proxy=$http_proxy -e https_proxy=$https_proxy --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e HUGGING_FACE_HUB_TOKEN=$HF_TOKEN -e VLLM_TORCH_PROFILER_DIR="/mnt" --cap-add=sys_nice --ipc=host opea/vllm-gaudi:latest --model $model_name --tensor-parallel-size 1 --host 0.0.0.0 --port 80 --block-size 128 --max-num-seqs 256 --max-seq_len-to-capture 2048
+   # Start TGI LLM Service
+   docker run -p 8005:80 -v ./data:/data --name tgi-gaudi-server -e HF_ENDPOINT=$HF_ENDPOINT -e http_proxy=$http_proxy -e https_proxy=$https_proxy --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e HUGGING_FACE_HUB_TOKEN=$HF_TOKEN -e ENABLE_HPU_GRAPH=true -e LIMIT_HPU_GRAPH=true -e USE_FLASH_ATTENTION=true -e FLASH_ATTENTION_RECOMPUTE=true --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tgi-gaudi:2.0.6 --model-id $model_name --max-input-tokens 1024 --max-total-tokens 2048
    ```
 
 2. Offline
@@ -201,12 +206,15 @@ For users in China who are unable to download models directly from Huggingface,
 
    - Click on `Download this model` button, and choose one way to download the model to your local path `/path/to/model`.
 
-   - Run the following command to start TGI service.
+   - Run the following command to start the LLM service.
 
      ```bash
      export HF_TOKEN=${your_hf_token}
      export model_path="/path/to/model"
-     docker run -p 8008:80 -v $model_path:/data --name tgi_service --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e HUGGING_FACE_HUB_TOKEN=$HF_TOKEN -e ENABLE_HPU_GRAPH=true -e LIMIT_HPU_GRAPH=true -e USE_FLASH_ATTENTION=true -e FLASH_ATTENTION_RECOMPUTE=true --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tgi-gaudi:2.0.6 --model-id /data --max-input-tokens 1024 --max-total-tokens 2048
+     # Start vLLM LLM Service
+     docker run -p 8007:80 -v $model_path:/data --name vllm-gaudi-server --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e HUGGING_FACE_HUB_TOKEN=$HF_TOKEN -e VLLM_TORCH_PROFILER_DIR="/mnt" --cap-add=sys_nice --ipc=host opea/vllm-gaudi:latest --model /data --tensor-parallel-size 1 --host 0.0.0.0 --port 80 --block-size 128 --max-num-seqs 256 --max-seq_len-to-capture 2048
+     # Start TGI LLM Service
+     docker run -p 8005:80 -v $model_path:/data --name tgi-gaudi-server --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none -e HUGGING_FACE_HUB_TOKEN=$HF_TOKEN -e ENABLE_HPU_GRAPH=true -e LIMIT_HPU_GRAPH=true -e USE_FLASH_ATTENTION=true -e FLASH_ATTENTION_RECOMPUTE=true --cap-add=sys_nice --ipc=host ghcr.io/huggingface/tgi-gaudi:2.0.6 --model-id /data --max-input-tokens 1024 --max-total-tokens 2048
      ```
 
 ### Setup Environment Variables
@@ -242,7 +250,7 @@ For users in China who are unable to download models directly from Huggingface,
 cd GenAIExamples/ChatQnA/docker_compose/intel/hpu/gaudi/
 ```
 
-If use tgi for llm backend.
+If use vLLM as the LLM serving backend.
 
 ```bash
 # Start ChatQnA with Rerank Pipeline
@@ -251,10 +259,10 @@ docker compose -f compose.yaml up -d
 docker compose -f compose_without_rerank.yaml up -d
 ```
 
-If use vllm for llm backend.
+If use TGI as the LLM serving backend.
 
 ```bash
-docker compose -f compose_vllm.yaml up -d
+docker compose -f compose_tgi.yaml up -d
 ```
 
 If you want to enable guardrails microservice in the pipeline, please follow the below command instead:
@@ -309,35 +317,40 @@ For validation details, please refer to [how-to-validate_service](./how_to_valid
 
 4. LLM backend Service
 
-   In first startup, this service will take more time to download the model files. After it's finished, the service will be ready.
+   In the first startup, this service will take more time to download, load and warm up the model. After it's finished, the service will be ready.
 
    Try the command below to check whether the LLM serving is ready.
 
    ```bash
-   docker logs tgi-gaudi-server | grep Connected
+   # vLLM service
+   docker logs vllm-gaudi-server 2>&1 | grep complete
+   # If the service is ready, you will get the response like below.
+   INFO:     Application startup complete.
    ```
 
+   ```bash
+   # TGI service
+   docker logs tgi-gaudi-server | grep Connected
    If the service is ready, you will get the response like below.
-
-   ```
    2024-09-03T02:47:53.402023Z  INFO text_generation_router::server: router/src/server.rs:2311: Connected
    ```
 
    Then try the `cURL` command below to validate services.
 
    ```bash
-   # TGI service
-   curl http://${host_ip}:8005/v1/chat/completions \
+   # vLLM Service
+   curl http://${host_ip}:8007/v1/chat/completions \
      -X POST \
      -d '{"model": ${LLM_MODEL_ID}, "messages": [{"role": "user", "content": "What is Deep Learning?"}], "max_tokens":17}' \
      -H 'Content-Type: application/json'
    ```
 
    ```bash
-   # vLLM Service
-   curl http://${host_ip}:8007/v1/chat/completions \
-     -H "Content-Type: application/json" \
-     -d '{"model": ${LLM_MODEL_ID}, "messages": [{"role": "user", "content": "What is Deep Learning?"}]}'
+   # TGI service
+   curl http://${host_ip}:8005/v1/chat/completions \
+     -X POST \
+     -d '{"model": ${LLM_MODEL_ID}, "messages": [{"role": "user", "content": "What is Deep Learning?"}], "max_tokens":17}' \
+     -H 'Content-Type: application/json'
    ```
 
 5. MegaService

ChatQnA/docker_compose/intel/hpu/gaudi/compose.yaml

@@ -25,7 +25,6 @@ services:
       INDEX_NAME: ${INDEX_NAME}
       TEI_ENDPOINT: http://tei-embedding-service:80
       HUGGINGFACEHUB_API_TOKEN: ${HUGGINGFACEHUB_API_TOKEN}
-      TELEMETRY_ENDPOINT: ${TELEMETRY_ENDPOINT}
   tei-embedding-service:
     image: ghcr.io/huggingface/text-embeddings-inference:cpu-1.5
     container_name: tei-embedding-gaudi-server
@@ -38,7 +37,7 @@ services:
       no_proxy: ${no_proxy}
       http_proxy: ${http_proxy}
       https_proxy: ${https_proxy}
-    command: --model-id ${EMBEDDING_MODEL_ID} --auto-truncate --otlp-endpoint $OTEL_EXPORTER_OTLP_TRACES_ENDPOINT
+    command: --model-id ${EMBEDDING_MODEL_ID} --auto-truncate
   retriever:
     image: ${REGISTRY:-opea}/retriever:${TAG:-latest}
     container_name: retriever-redis-server
@@ -56,9 +55,6 @@ services:
       INDEX_NAME: ${INDEX_NAME}
       TEI_EMBEDDING_ENDPOINT: http://tei-embedding-service:80
       HUGGINGFACEHUB_API_TOKEN: ${HUGGINGFACEHUB_API_TOKEN}
-      TELEMETRY_ENDPOINT: ${TELEMETRY_ENDPOINT}
-      LOGFLAG: ${LOGFLAG}
-      RETRIEVER_COMPONENT_NAME: "OPEA_RETRIEVER_REDIS"
     restart: unless-stopped
   tei-reranking-service:
     image: ghcr.io/huggingface/tei-gaudi:1.5.0
@@ -80,47 +76,28 @@ services:
       HABANA_VISIBLE_DEVICES: all
       OMPI_MCA_btl_vader_single_copy_mechanism: none
       MAX_WARMUP_SEQUENCE_LENGTH: 512
-    command: --model-id ${RERANK_MODEL_ID} --auto-truncate --otlp-endpoint $OTEL_EXPORTER_OTLP_TRACES_ENDPOINT
-  tgi-service:
-    image: ghcr.io/huggingface/tgi-gaudi:2.0.6
-    container_name: tgi-gaudi-server
+    command: --model-id ${RERANK_MODEL_ID} --auto-truncate
+  vllm-service:
+    image: ${REGISTRY:-opea}/vllm-gaudi:${TAG:-latest}
+    container_name: vllm-gaudi-server
     ports:
-      - "8005:80"
+      - "8007:80"
     volumes:
       - "./data:/data"
     environment:
       no_proxy: ${no_proxy}
       http_proxy: ${http_proxy}
       https_proxy: ${https_proxy}
-      HUGGING_FACE_HUB_TOKEN: ${HUGGINGFACEHUB_API_TOKEN}
-      HF_HUB_DISABLE_PROGRESS_BARS: 1
-      HF_HUB_ENABLE_HF_TRANSFER: 0
+      HF_TOKEN: ${HUGGINGFACEHUB_API_TOKEN}
       HABANA_VISIBLE_DEVICES: all
       OMPI_MCA_btl_vader_single_copy_mechanism: none
-      ENABLE_HPU_GRAPH: true
-      LIMIT_HPU_GRAPH: true
-      USE_FLASH_ATTENTION: true
-      FLASH_ATTENTION_RECOMPUTE: true
+      LLM_MODEL_ID: ${LLM_MODEL_ID}
+      VLLM_TORCH_PROFILER_DIR: "/mnt"
     runtime: habana
     cap_add:
       - SYS_NICE
     ipc: host
-    command: --model-id ${LLM_MODEL_ID} --max-input-length 2048 --max-total-tokens 4096 --otlp-endpoint $OTEL_EXPORTER_OTLP_TRACES_ENDPOINT
-  jaeger:
-    image: jaegertracing/all-in-one:latest
-    container_name: jaeger
-    ports:
-      - "16686:16686"
-      - "4317:4317"
-      - "4318:4318"
-      - "9411:9411"
-    ipc: host
-    environment:
-      no_proxy: ${no_proxy}
-      http_proxy: ${http_proxy}
-      https_proxy: ${https_proxy}
-      COLLECTOR_ZIPKIN_HOST_PORT: 9411
-    restart: unless-stopped
+    command: --model $LLM_MODEL_ID --tensor-parallel-size 1 --host 0.0.0.0 --port 80 --block-size 128 --max-num-seqs 256 --max-seq_len-to-capture 2048
   chatqna-gaudi-backend-server:
     image: ${REGISTRY:-opea}/chatqna:${TAG:-latest}
     container_name: chatqna-gaudi-backend-server
@@ -129,7 +106,7 @@ services:
       - tei-embedding-service
       - retriever
       - tei-reranking-service
-      - tgi-service
+      - vllm-service
     ports:
       - "8888:8888"
     environment:
@@ -142,11 +119,10 @@ services:
       - RETRIEVER_SERVICE_HOST_IP=retriever
       - RERANK_SERVER_HOST_IP=tei-reranking-service
       - RERANK_SERVER_PORT=${RERANK_SERVER_PORT:-80}
-      - LLM_SERVER_HOST_IP=tgi-service
+      - LLM_SERVER_HOST_IP=vllm-service
       - LLM_SERVER_PORT=${LLM_SERVER_PORT:-80}
       - LLM_MODEL=${LLM_MODEL_ID}
       - LOGFLAG=${LOGFLAG}
-      - TELEMETRY_ENDPOINT=${TELEMETRY_ENDPOINT}
     ipc: host
     restart: always
   chatqna-gaudi-ui-server:

ChatQnA/docker_compose/intel/hpu/gaudi/compose_guardrails.yaml

@@ -118,9 +118,9 @@ services:
       OMPI_MCA_btl_vader_single_copy_mechanism: none
       MAX_WARMUP_SEQUENCE_LENGTH: 512
     command: --model-id ${RERANK_MODEL_ID} --auto-truncate
-  tgi-service:
-    image: ghcr.io/huggingface/tgi-gaudi:2.0.6
-    container_name: tgi-gaudi-server
+  vllm-service:
+    image: ${REGISTRY:-opea}/vllm-gaudi:${TAG:-latest}
+    container_name: vllm-gaudi-server
     ports:
       - "8008:80"
     volumes:
@@ -129,20 +129,16 @@ services:
       no_proxy: ${no_proxy}
       http_proxy: ${http_proxy}
       https_proxy: ${https_proxy}
-      HUGGING_FACE_HUB_TOKEN: ${HUGGINGFACEHUB_API_TOKEN}
-      HF_HUB_DISABLE_PROGRESS_BARS: 1
-      HF_HUB_ENABLE_HF_TRANSFER: 0
+      HF_TOKEN: ${HUGGINGFACEHUB_API_TOKEN}
       HABANA_VISIBLE_DEVICES: all
       OMPI_MCA_btl_vader_single_copy_mechanism: none
-      ENABLE_HPU_GRAPH: true
-      LIMIT_HPU_GRAPH: true
-      USE_FLASH_ATTENTION: true
-      FLASH_ATTENTION_RECOMPUTE: true
+      LLM_MODEL_ID: ${LLM_MODEL_ID}
+      VLLM_TORCH_PROFILER_DIR: "/mnt"
     runtime: habana
     cap_add:
       - SYS_NICE
     ipc: host
-    command: --model-id ${LLM_MODEL_ID} --max-input-length 1024 --max-total-tokens 2048
+    command: --model $LLM_MODEL_ID --tensor-parallel-size 1 --host 0.0.0.0 --port 80 --block-size 128 --max-num-seqs 256 --max-seq_len-to-capture 2048
   chatqna-gaudi-backend-server:
     image: ${REGISTRY:-opea}/chatqna-guardrails:${TAG:-latest}
     container_name: chatqna-gaudi-guardrails-server
@@ -153,7 +149,7 @@ services:
       - tei-embedding-service
       - retriever
       - tei-reranking-service
-      - tgi-service
+      - vllm-service
     ports:
       - "8888:8888"
     environment:
@@ -168,7 +164,7 @@ services:
       - RETRIEVER_SERVICE_HOST_IP=retriever
       - RERANK_SERVER_HOST_IP=tei-reranking-service
       - RERANK_SERVER_PORT=${RERANK_SERVER_PORT:-80}
-      - LLM_SERVER_HOST_IP=tgi-service
+      - LLM_SERVER_HOST_IP=vllm-service
       - LLM_SERVER_PORT=${LLM_SERVER_PORT:-80}
       - LLM_MODEL=${LLM_MODEL_ID}
       - LOGFLAG=${LOGFLAG}