This is a BentoML example project, showing you how to serve and deploy open-source Large Language Models (LLMs) using TensorRT-LLM, a Python API that optimizes LLM inference on NVIDIA GPUs using TensorRT engine.

See here for a full list of BentoML example projects.

💡 This example is served as a basis for advanced code customization, such as custom model, inference logic or LMDeploy options. For simple LLM hosting with OpenAI compatible endpoint without writing any code, see OpenLLM.

Prerequisites

• You have installed Python 3.10+ and pip. See the Python downloads page to learn more.
• You have a basic understanding of key concepts in BentoML, such as Services. We recommend you read Quickstart first.
• You have installed Docker, which will be used to create a container environment to run TensorRT-LLM.
• If you want to test the Service locally, you need a Nvidia GPU with at least 20G VRAM.
• This example uses Llama 3. Make sure you have gained access to the model.
• (Optional) We recommend you create a virtual environment for dependency isolation for this project. See the Conda documentation or the Python documentation for details.

Set up the environment

Clone the project repo and TensorRT-LLM repo.

bash git clone https://github.com/bentoml/BentoTRTLLM.git cd BentoTRTLLM/llama-3-8b-instruct git clone -b v0.10.0 https://github.com/NVIDIA/TensorRT-LLM.git cd TensorRT-LLM

Note: To deploy Llama 3 70B AWQ, go to the llama-3-70b-instruct directory.

Create the base Docker environment to compile the model.

bash git lfs install git clone https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct docker run --rm --runtime=nvidia --gpus all --volume ${PWD}:/TensorRT-LLM --entrypoint /bin/bash -it --workdir /TensorRT-LLM nvidia/cuda:12.1.0-devel-ubuntu22.04

Install dependencies inside the Docker container. Note that TensorRT-LLM requires Python 3.10.

```bash apt-get update && apt-get -y install python3.10 python3-pip openmpi-bin libopenmpi-dev

Install the stable version (corresponding to the cloned branch) of TensorRT-LLM.

pip3 install tensorrt_llm==0.10.0 -U --extra-index-url https://pypi.nvidia.com pip3 install --force-reinstall -U --extra-index-url https://pypi.nvidia.com tensorrt-cu12==10.0.1

Log in to huggingface-cli

You can get your token from huggingface.co/settings/token

apt-get install -y git huggingface-cli login --token ***** ```

Build the Llama 8B model using a single GPU and BF16.

```bash python3 examples/llama/convertcheckpoint.py --modeldir ./Meta-Llama-3-8B-Instruct \ --outputdir ./tllmcheckpoint1gpubf16 \ --dtype bfloat16

trtllm-build --checkpointdir ./tllmcheckpoint1gpubf16 \ --outputdir ./tmp/llama/8B/trtengines/bf16/1-gpu \ --gptattentionplugin bfloat16 \ --gemmplugin bfloat16 \ --maxbatchsize 2048 \ --maxinputlen 2048 \ --maxnumtokens 2048 \ --multipleprofiles enable \ --pagedkvcache enable \ --usepagedcontext_fmha enable ```

The model should be successfully built now. Exit the Docker image.

bash exit

Clone the tensorrtllm_backend repo.

bash cd .. git clone -b v0.10.0 https://github.com/triton-inference-server/tensorrtllm_backend.git

Now, the BentoTRTLLM/ directory should have one TenosrRT-LLM/ directory and one tensorrtllm_backend/ directory.

Copy the model.

bash cd tensorrtllm_backend cp ../TensorRT-LLM/tmp/llama/8B/trt_engines/bf16/1-gpu/* all_models/inflight_batcher_llm/tensorrt_llm/1/

Set the tokenizer_dir and engine_dir paths.

```bash HFLLAMAMODEL=TensorRT-LLM/Meta-Llama-3-8B-Instruct ENGINEPATH=tensorrtllmbackend/allmodels/inflightbatcherllm/tensorrtllm/1

python3 tools/filltemplate.py -i allmodels/inflightbatcherllm/preprocessing/config.pbtxt tokenizerdir:${HFLLAMAMODEL},tokenizertype:auto,tritonmaxbatchsize:2048,preprocessinginstance_count:1

python3 tools/filltemplate.py -i allmodels/inflightbatcherllm/postprocessing/config.pbtxt tokenizerdir:${HFLLAMAMODEL},tokenizertype:auto,tritonmaxbatchsize:2048,postprocessinginstance_count:8

python3 tools/filltemplate.py -i allmodels/inflightbatcherllm/tensorrtllmbls/config.pbtxt tritonmaxbatchsize:2048,decoupledmode:True,blsinstancecount:1,accumulate_tokens:False

python3 tools/filltemplate.py -i allmodels/inflightbatcherllm/ensemble/config.pbtxt tritonmaxbatch_size:2048

python3 tools/filltemplate.py -i allmodels/inflightbatcherllm/tensorrtllm/config.pbtxt tritonbackend:tensorrtllm,tritonmaxbatchsize:2048,decoupledmode:True,maxbeamwidth:1,enginedir:${ENGINEPATH},maxtokensinpagedkvcache:,maxattentionwindowsize:2560,kvcachefreegpumemfraction:0.9,excludeinputinoutput:True,batchingstrategy:inflightfusedbatching,maxqueuedelaymicroseconds:0,enablechunkedcontext:True ```

Import the model

Install BentoML.

bash pip install bentoml

Make sure you are in the llama-3-8b-instruct directory and import the model to the BentoML Model Store.

bash python pack_model.py

To verify it, run:

```bash $ bentoml models list

Tag Size Creation Time meta-llama--meta-llama-3-8b-instruct-trtllm-rtx4000:7eu4l2reqwohx3lu 45.80 GiB 2024-06-07 04:25:30 ```

Run the BentoML Service

We have defined a BentoML Service in service.py. To serve it locally, first create a Docker container environment for TensorRT-LLM:

bash docker run --runtime=nvidia --gpus all -v ${PWD}:/BentoTRTLLM -v ~/bentoml:/root/bentoml -p 3000:3000 --entrypoint /bin/bash -it --workdir /BentoTRTLLM nvcr.io/nvidia/tritonserver:24.06-trtllm-python-py3

Install the dependencies.

bash pip install -r requirements.txt

Start the Service.

bash $ bentoml serve . 2024-06-07T05:16:38+0000 [INFO] [cli] Starting production HTTP BentoServer from "service:TRTLLM" listening on http://localhost:3000 (Press CTRL+C to quit) I0607 05:16:39.805180 117 pinned_memory_manager.cc:275] Pinned memory pool is created at '0x7f7c64000000' with size 268435456 I0607 05:16:39.805431 117 cuda_memory_manager.cc:107] CUDA memory pool is created on device 0 with size 67108864 I0607 05:16:39.810192 117 model_lifecycle.cc:469] loading: postprocessing:1 I0607 05:16:39.810243 117 model_lifecycle.cc:469] loading: preprocessing:1 I0607 05:16:39.810385 117 model_lifecycle.cc:469] loading: tensorrt_llm:1 I0607 05:16:39.810426 117 model_lifecycle.cc:469] loading: tensorrt_llm_bls:1 I0607 05:16:39.841462 117 python_be.cc:2391] TRITONBACKEND_ModelInstanceInitialize: postprocessing_0_0 (CPU device 0) I0607 05:16:39.841462 117 python_be.cc:2391] TRITONBACKEND_ModelInstanceInitialize: preprocessing_0_0 (CPU device 0) [TensorRT-LLM][WARNING] gpu_device_ids is not specified, will be automatically set [TensorRT-LLM][WARNING] max_tokens_in_paged_kv_cache is not specified, will use default value [TensorRT-LLM][WARNING] batch_scheduler_policy parameter was not found or is invalid (must be max_utilization or guaranteed_no_evict) [TensorRT-LLM][WARNING] enable_chunked_context is not specified, will be set to false. ...

The server is now active at http://localhost:3000. You can interact with it using the Swagger UI or in other different ways.

CURL

```bash curl -X 'POST' \ 'http://localhost:3000/generate' \ -H 'accept: text/event-stream' \ -H 'Content-Type: application/json' \ -d '{ "prompt": "Explain superconductors like I'\''m five years old", "max_tokens": 1024 }' ```

Python client

```python import bentoml with bentoml.SyncHTTPClient("http://localhost:3000") as client: response_generator = client.generate( prompt="Explain superconductors like I'm five years old", max_tokens=1024 ) for response in response_generator: print(response, end='') ```

Deploy to BentoCloud

After the Service is ready, you can deploy the application to BentoCloud for better management and scalability. Sign up if you haven't got a BentoCloud account.

Make sure you have logged in to BentoCloud, then run the following command to deploy it. Note that you need to specify the CUDA version in bentofile.yaml.

bash bentoml deploy .

Once the application is up and running on BentoCloud, you can access it via the exposed URL.

Note: For custom deployment in your own infrastructure, use BentoML to generate an OCI-compliant image.