Self-host LLMs with TensorRT-LLM and BentoML

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.

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.

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.

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.

python3 examples/llama/convert_checkpoint.py --model_dir ./Meta-Llama-3-8B-Instruct \
            --output_dir ./tllm_checkpoint_1gpu_bf16 \
            --dtype bfloat16

trtllm-build --checkpoint_dir ./tllm_checkpoint_1gpu_bf16 \
            --output_dir ./tmp/llama/8B/trt_engines/bf16/1-gpu \
            --gpt_attention_plugin bfloat16 \
            --gemm_plugin bfloat16 \
            --max_batch_size 2048 \
            --max_input_len 2048 \
            --max_num_tokens 2048 \
            --multiple_profiles enable \
            --paged_kv_cache enable \
            --use_paged_context_fmha enable

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

exit

Clone the tensorrtllm_backend repo.

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.

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.

HF_LLAMA_MODEL=TensorRT-LLM/Meta-Llama-3-8B-Instruct
ENGINE_PATH=tensorrtllm_backend/all_models/inflight_batcher_llm/tensorrt_llm/1

python3 tools/fill_template.py -i all_models/inflight_batcher_llm/preprocessing/config.pbtxt tokenizer_dir:${HF_LLAMA_MODEL},tokenizer_type:auto,triton_max_batch_size:2048,preprocessing_instance_count:1

python3 tools/fill_template.py -i all_models/inflight_batcher_llm/postprocessing/config.pbtxt tokenizer_dir:${HF_LLAMA_MODEL},tokenizer_type:auto,triton_max_batch_size:2048,postprocessing_instance_count:8

python3 tools/fill_template.py -i all_models/inflight_batcher_llm/tensorrt_llm_bls/config.pbtxt triton_max_batch_size:2048,decoupled_mode:True,bls_instance_count:1,accumulate_tokens:False

python3 tools/fill_template.py -i all_models/inflight_batcher_llm/ensemble/config.pbtxt triton_max_batch_size:2048

python3 tools/fill_template.py -i all_models/inflight_batcher_llm/tensorrt_llm/config.pbtxt triton_backend:tensorrtllm,triton_max_batch_size:2048,decoupled_mode:True,max_beam_width:1,engine_dir:${ENGINE_PATH},max_tokens_in_paged_kv_cache:,max_attention_window_size:2560,kv_cache_free_gpu_mem_fraction:0.9,exclude_input_in_output:True,batching_strategy:inflight_fused_batching,max_queue_delay_microseconds:0,enable_chunked_context:True

Import the model

Install BentoML.

pip install bentoml

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

python pack_model.py

To verify it, run:

$ 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:

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.

pip install -r requirements.txt

Start the Service.

$ 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

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

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.

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.