TensorRT-LLM Backend

The Triton backend for TensorRT-LLM. You can learn more about Triton backends in the backend repo. The goal of TensorRT-LLM Backend is to let you serve TensorRT-LLM models with Triton Inference Server. The inflight_batcher_llm directory contains the C++ implementation of the backend supporting inflight batching, paged attention and more.

Where can I ask general questions about Triton and Triton backends? Be sure to read all the information below as well as the general Triton documentation available in the main server repo. If you don't find your answer there you can ask questions on the issues page.

Table of Contents

• TensorRT-LLM Backend
  • Table of Contents
  • Getting Started
  • Quick Start
    • Update the TensorRT-LLM submodule
    • Launch Triton TensorRT-LLM container
    • Prepare TensorRT-LLM engines
    • Prepare the Model Repository
    • Modify the Model Configuration
    • Serving with Triton
    • Send an Inference Request
    • Using the generate endpoint
    • Using the client scripts
      • Early stopping
      • Return context logits and/or generation logits
    • Requests with batch size > 1
  • Building from Source
  • Supported Models
  • Model Config
  • Model Deployment
  • TRT-LLM Multi-instance Support
    • Leader Mode
    • Orchestrator Mode
    • Running Multiple Instances of LLaMa Model
  • Multi-node Support
  • Model Parallelism
    • Tensor Parallelism, Pipeline Parallelism and Expert Parallelism
  • MIG Support
  • Scheduling
  • Key-Value Cache
  • Decoding
    • Decoding Modes - Top-k, Top-p, Top-k Top-p, Beam Search, Medusa, ReDrafter, Lookahead and Eagle
    • Speculative Decoding
  • Chunked Context
  • Quantization
  • LoRa
  • Launch Triton server within Slurm based clusters
  • Prepare some scripts
  • Submit a Slurm job
  • Triton Metrics
  • Benchmarking
  • Testing the TensorRT-LLM Backend

Getting Started

Quick Start

Below is an example of how to serve a TensorRT-LLM model with the Triton TensorRT-LLM Backend on a 4-GPU environment. The example uses the GPT model from the TensorRT-LLM repository with the NGC Triton TensorRT-LLM container. Make sure you are cloning the same version of TensorRT-LLM backend as the version of TensorRT-LLM in the container. Please refer to the support matrix to see the aligned versions.

In this example, we will use Triton 24.07 with TensorRT-LLM v0.11.0.

Update the TensorRT-LLM submodule

git clone -b v0.11.0 https://github.com/triton-inference-server/tensorrtllm_backend.git
cd tensorrtllm_backend
git submodule update --init --recursive
git lfs install
git lfs pull

Launch Triton TensorRT-LLM container

Launch Triton docker container nvcr.io/nvidia/tritonserver:<xx.yy>-trtllm-python-py3 with TensorRT-LLM backend.

Make an engines folder outside docker to reuse engines for future runs. Make sure to replace the <xx.yy> with the version of Triton that you want to use.

docker run --rm -it --net host --shm-size=2g \
    --ulimit memlock=-1 --ulimit stack=67108864 --gpus all \
    -v </path/to/tensorrtllm_backend>:/tensorrtllm_backend \
    -v </path/to/engines>:/engines \
    nvcr.io/nvidia/tritonserver:24.07-trtllm-python-py3

Prepare TensorRT-LLM engines

You can skip this step if you already have the engines ready. Follow the guide in TensorRT-LLM repository for more details on how to to prepare the engines for all the supported models. You can also check out the tutorials to see more examples with serving TensorRT-LLM models.

cd /tensorrtllm_backend/tensorrt_llm/examples/gpt

# Download weights from HuggingFace Transformers
rm -rf gpt2 && git clone https://huggingface.co/gpt2-medium gpt2
pushd gpt2 && rm pytorch_model.bin model.safetensors && wget -q https://huggingface.co/gpt2-medium/resolve/main/pytorch_model.bin && popd

# Convert weights from HF Tranformers to TensorRT-LLM checkpoint
python3 convert_checkpoint.py --model_dir gpt2 \
        --dtype float16 \
        --tp_size 4 \
        --output_dir ./c-model/gpt2/fp16/4-gpu

# Build TensorRT engines
trtllm-build --checkpoint_dir ./c-model/gpt2/fp16/4-gpu \
        --gpt_attention_plugin float16 \
        --remove_input_padding enable \
        --kv_cache_type paged \
        --gemm_plugin float16 \
        --output_dir /engines/gpt/fp16/4-gpu

See here for more details on the parameters.

Prepare the Model Repository

Next, create the model repository that will be used by the Triton server. The models can be found in the all_models folder. The folder contains two groups of models:

• gpt: Using TensorRT-LLM pure Python runtime.
• inflight_batcher_llm`: Using the C++ TensorRT-LLM backend with the executor API, which includes the latest features including inflight batching.

There are five models in all_models/inflight_batcher_llm that will be used in this example:

Model	Description
ensemble	This model is used to chain the preprocessing, tensorrt_llm and postprocessing models together.
preprocessing	This model is used for tokenizing, meaning the conversion from prompts(string) to input_ids(list of ints).
tensorrt_llm	This model is a wrapper of your TensorRT-LLM model and is used for inferencing. Input specification can be found here
postprocessing	This model is used for de-tokenizing, meaning the conversion from output_ids(list of ints) to outputs(string).
tensorrt_llm_bls	This model can also be used to chain the preprocessing, tensorrt_llm and postprocessing models together.

To learn more about ensemble and BLS models, please see the Ensemble Models and Business Logic Scripting documentation.

To learn more about the benefits and the limitations of using the BLS model, please see the model config section.

mkdir /triton_model_repo
cp -r /tensorrtllm_backend/all_models/inflight_batcher_llm/* /triton_model_repo/

Modify the Model Configuration

Use the script to fill in the parameters in the model configuration files. For optimal performance or custom parameters, please refer to perf_best_practices. For more details on the model configuration and the parameters that can be modified, please refer to the model config section.

ENGINE_DIR=/engines/gpt/fp16/4-gpu
TOKENIZER_DIR=/tensorrtllm_backend/tensorrt_llm/examples/gpt/gpt2
MODEL_FOLDER=/triton_model_repo
TRITON_MAX_BATCH_SIZE=4
INSTANCE_COUNT=1
MAX_QUEUE_DELAY_MS=0
MAX_QUEUE_SIZE=0
FILL_TEMPLATE_SCRIPT=/tensorrtllm_backend/tools/fill_template.py
DECOUPLED_MODE=false

python3 ${FILL_TEMPLATE_SCRIPT} -i ${MODEL_FOLDER}/ensemble/config.pbtxt triton_max_batch_size:${TRITON_MAX_BATCH_SIZE}
python3 ${FILL_TEMPLATE_SCRIPT} -i ${MODEL_FOLDER}/preprocessing/config.pbtxt tokenizer_dir:${TOKENIZER_DIR},triton_max_batch_size:${TRITON_MAX_BATCH_SIZE},preprocessing_instance_count:${INSTANCE_COUNT}
python3 ${FILL_TEMPLATE_SCRIPT} -i ${MODEL_FOLDER}/tensorrt_llm/config.pbtxt triton_backend:tensorrtllm,triton_max_batch_size:${TRITON_MAX_BATCH_SIZE},decoupled_mode:${DECOUPLED_MODE},engine_dir:${ENGINE_DIR},max_queue_delay_microseconds:${MAX_QUEUE_DELAY_MS},batching_strategy:inflight_fused_batching,max_queue_size:${MAX_QUEUE_SIZE}
python3 ${FILL_TEMPLATE_SCRIPT} -i ${MODEL_FOLDER}/postprocessing/config.pbtxt tokenizer_dir:${TOKENIZER_DIR},triton_max_batch_size:${TRITON_MAX_BATCH_SIZE},postprocessing_instance_count:${INSTANCE_COUNT},max_queue_size:${MAX_QUEUE_SIZE}
python3 ${FILL_TEMPLATE_SCRIPT} -i ${MODEL_FOLDER}/tensorrt_llm_bls/config.pbtxt triton_max_batch_size:${TRITON_MAX_BATCH_SIZE},decoupled_mode:${DECOUPLED_MODE},bls_instance_count:${INSTANCE_COUNT}

NOTE: It is recommended to match the number of pre/post_instance_counts with triton_max_batch_size for better performance.

Serving with Triton

Now, you're ready to launch the Triton server with the TensorRT-LLM model.

Use the launch_triton_server.py script. This launches multiple instances of tritonserver with MPI.

# 'world_size' is the number of GPUs you want to use for serving. This should
# be aligned with the number of GPUs used to build the TensorRT-LLM engine.
python3 /tensorrtllm_backend/scripts/launch_triton_server.py --world_size=4 --model_repo=${MODEL_FOLDER}

You should see the following logs when the server is successfully deployed.

...
I0503 22:01:25.210518 1175 grpc_server.cc:2463] Started GRPCInferenceService at 0.0.0.0:8001
I0503 22:01:25.211612 1175 http_server.cc:4692] Started HTTPService at 0.0.0.0:8000
I0503 22:01:25.254914 1175 http_server.cc:362] Started Metrics Service at 0.0.0.0:8002

To stop Triton Server inside the container, run:

pkill tritonserver

Send an Inference Request

Using the generate endpoint

The general format of the generate endpoint:

curl -X POST localhost:8000/v2/models/${MODEL_NAME}/generate -d '{"{PARAM1_KEY}": "{PARAM1_VALUE}", ... }'

In the case of the models used in this example, you can replace MODEL_NAME with ensemble or tensorrt_llm_bls. Examining the ensemble and tensorrt_llm_bls model's config.pbtxt file, you can see that 4 parameters are required to generate a response for this model:

• text_input: Input text to generate a response from
• max_tokens: The number of requested output tokens
• bad_words: A list of bad words (can be empty)
• stop_words: A list of stop words (can be empty)

Therefore, we can query the server in the following way:

• if using the ensemble model

  curl -X POST localhost:8000/v2/models/ensemble/generate -d '{"text_input": "What is machine learning?", "max_tokens": 20, "bad_words": "", "stop_words": ""}'

• if using the tensorrt_llm_bls model

curl -X POST localhost:8000/v2/models/tensorrt_llm_bls/generate -d '{"text_input": "What is machine learning?", "max_tokens": 20, "bad_words": "", "stop_words": ""}'

Which should return a result similar to (formatted for readability):

{
  "model_name": "ensemble",
  "model_version": "1",
  "sequence_end": false,
  "sequence_id": 0,
  "sequence_start": false,
  "text_output": "What is machine learning?\n\nMachine learning is a method of learning by using machine learning algorithms to solve problems.\n\n"
}

Using the client scripts

You can refer to the client scripts in the inflight_batcher_llm/client to see how to send requests via Python scripts.

Below is an example of using inflight_batcher_llm_client to send requests to the tensorrt_llm model.

pip3 install tritonclient[all]
INFLIGHT_BATCHER_LLM_CLIENT=/tensorrtllm_backend/inflight_batcher_llm/client/inflight_batcher_llm_client.py
python3 ${INFLIGHT_BATCHER_LLM_CLIENT} --request-output-len 200 --tokenizer-dir ${TOKENIZER_DIR}

The result should be similar to the following:

Using pad_id:  50256
Using end_id:  50256
Input sequence:  [28524, 287, 5093, 12, 23316, 4881, 11, 30022, 263, 8776, 355, 257]
Got completed request
Input: Born in north-east France, Soyer trained as a
Output beam 0:  chef before moving to London in the early 1990s. He has since worked in restaurants in London, Paris, Milan and New York.

He is married to the former model and actress, Anna-Marie, and has two children, a daughter, Emma, and a son, Daniel.

Soyer's wife, Anna-Marie, is a former model and actress.

He is survived by his wife, Anna-Marie, and their two children, Daniel and Emma.

Soyer was born in the north-east of France, and moved to London in the early 1990s.

He was a chef at the London restaurant, The Bistro, before moving to New York in the early 2000s.

He was a regular at the restaurant, and was also a regular at the restaurant, The Bistro, before moving to London in the early 2000s.

Soyer was a regular at the restaurant, and was
Output sequence:  [28524, 287, 5093, 12, 23316, 4881, 11, 30022, 263, 8776, 355, 257, 21221, 878, 3867, 284, 3576, 287, 262, 1903, 6303, 82, 13, 679, 468, 1201, 3111, 287, 10808, 287, 3576, 11, 6342, 11, 21574, 290, 968, 1971, 13, 198, 198, 1544, 318, 6405, 284, 262, 1966, 2746, 290, 14549, 11, 11735, 12, 44507, 11, 290, 468, 734, 1751, 11, 257, 4957, 11, 18966, 11, 290, 257, 3367, 11, 7806, 13, 198, 198, 50, 726, 263, 338, 3656, 11, 11735, 12, 44507, 11, 318, 257, 1966, 2746, 290, 14549, 13, 198, 198, 1544, 318, 11803, 416, 465, 3656, 11, 11735, 12, 44507, 11, 290, 511, 734, 1751, 11, 7806, 290, 18966, 13, 198, 198, 50, 726, 263, 373, 4642, 287, 262, 5093, 12, 23316, 286, 4881, 11, 290, 3888, 284, 3576, 287, 262, 1903, 6303, 82, 13, 198, 198, 1544, 373, 257, 21221, 379, 262, 3576, 7072, 11, 383, 347, 396, 305, 11, 878, 3867, 284, 968, 1971, 287, 262, 1903, 4751, 82, 13, 198, 198, 1544, 373, 257, 3218, 379, 262, 7072, 11, 290, 373, 635, 257, 3218, 379, 262, 7072, 11, 383, 347, 396, 305, 11, 878, 3867, 284, 3576, 287, 262, 1903, 4751, 82, 13, 198, 198, 50, 726, 263, 373, 257, 3218, 379, 262, 7072, 11, 290, 373]

Early stopping

You can also stop the generation process early by using the --stop-after-ms option to send a stop request after a few milliseconds:

python3 ${INFLIGHT_BATCHER_LLM_CLIENT} --stop-after-ms 200 --request-output-len 200 --tokenizer-dir ${TOKENIZER_DIR}

You will find that the generation process is stopped early and therefore the number of generated tokens is lower than 200. You can have a look at the client code to see how early stopping is achieved.

Return context logits and/or generation logits

If you want to get context logits and/or generation logits, you need to enable --gather_context_logits and/or --gather_generation_logits when building the engine (or --gather_all_token_logits to enable both at the same time). For more setting details about these two flags, please refer to build.py or gpt_runtime.

After launching the server, you could get the output of logits by passing the corresponding parameters --return-context-logits and/or --return-generation-logits in the client scripts (end_to_end_grpc_client.py and inflight_batcher_llm_client.py).

For example:

python3 ${INFLIGHT_BATCHER_LLM_CLIENT} --request-output-len 20 --tokenizer-dir ${TOKENIZER_DIR} --return-context-logits --return-generation-logits

The result should be similar to the following:

Input sequence:  [28524, 287, 5093, 12, 23316, 4881, 11, 30022, 263, 8776, 355, 257]
Got completed request
Input: Born in north-east France, Soyer trained as a
Output beam 0:  has since worked in restaurants in London,
Output sequence:  [21221, 878, 3867, 284, 3576, 287, 262, 1903, 6303, 82, 13, 679, 468, 1201, 3111, 287, 10808, 287, 3576, 11]
context_logits.shape: (1, 12, 50257)
context_logits: [[[ -65.9822     -62.267445   -70.08991   ...  -76.16964    -78.8893
    -65.90678  ]
  [-103.40278   -102.55243   -106.119026  ... -108.925415  -109.408585
   -101.37687  ]
  [ -63.971176   -64.03466    -67.58809   ...  -72.141235   -71.16892
    -64.23846  ]
  ...
  [ -80.776375   -79.1815     -85.50916   ...  -87.07368    -88.02817
    -79.28435  ]
  [ -10.551408    -7.786484   -14.524468  ...  -13.805856   -15.767286
     -7.9322424]
  [-106.33096   -105.58956   -111.44852   ... -111.04858   -111.994194
   -105.40376  ]]]
generation_logits.shape: (1, 1, 20, 50257)
generation_logits: [[[[-106.33096  -105.58956  -111.44852  ... -111.04858  -111.994194
    -105.40376 ]
   [ -77.867424  -76.96638   -83.119095 ...  -87.82542   -88.53957
     -75.64877 ]
   [-136.92282  -135.02484  -140.96051  ... -141.78284  -141.55045
    -136.01668 ]
   ...
   [-100.03721   -98.98237  -105.25507  ... -108.49254  -109.45882
     -98.95136 ]
   [-136.78777  -136.16165  -139.13437  ... -142.21495  -143.57468
    -134.94667 ]
   [  19.222942   19.127287   14.804495 ...   10.556551    9.685863
      19.625107]]]]

Requests with batch size > 1

The TRT-LLM backend supports requests with batch size greater than one. When sending a request with a batch size greater than one, the TRT-LLM backend will return multiple batch size 1 responses, where each response will be associated with a given batch index. An output tensor named batch_index is associated with each response to indicate which batch index this response corresponds to.

The client script end_to_end_grpc_client.py demonstrates how a client can send requests with batch size > 1 and consume the responses returned from Triton. When passing --batch-inputs to the client script, the client will create a request with multiple prompts, and use the batch_index output tensor to associate the responses to the original prompt. For example one could run:

python3 end_to_end_grpc_client.py -o 5 -p '["This is a test","I want you to","The cat is"]'  --batch-inputs

to send a request with a batch size of 3 to the Triton server.

Building from Source

Please refer to the build.md for more details on how to build the Triton TRT-LLM container from source.

Supported Models

Only a few examples are listed here. For all the supported models, please refer to the support matrix.

• LLaMa

  • End to end workflow to run llama 7b with Triton
  • Build and run a LLaMA model in TensorRT-LLM
  • Llama Multi-instance
  • Deploying Hugging Face Llama2-7b Model in Triton

• Gemma

  • End to end workflow to run sp model with Triton
  • Run Gemma on TensorRT-LLM

• Mistral

  • Build and run a Mixtral model in TensorRT-LLM

• Multi-modal

  • End to end workflow to run multimodal models(e.g. BLIP2-OPT, LLava1.5-7B, VILA) with Triton
  • Deploying Hugging Face Llava1.5-7b Model in Triton

• Encoder-Decoder

  • End to end workflow to run an Encoder-Decoder model with Triton

Model Config

Please refer to the model config for more details on the model configuration.

Model Deployment

TRT-LLM Multi-instance Support

TensorRT-LLM backend relies on MPI to coordinate the execution of a model across multiple GPUs and nodes. Currently, there are two different modes supported to run a model across multiple GPUs, Leader Mode and Orchestrator Mode.

Note: This is different from the model multi-instance support from Triton Server which allows multiple instances of a model to be run on the same or different GPUs. For more information on Triton Server multi-instance support, please refer to the Triton model config documentation.

Leader Mode

In leader mode, TensorRT-LLM backend spawns one Triton Server process for every GPU. The process with rank 0 is the leader process. Other Triton Server processes, do not return from the TRITONBACKEND_ModelInstanceInitialize call to avoid port collision and allowing the other processes to receive requests.

The overview of this mode is described in the diagram below:

This mode is friendly with slurm deployments since it doesn't use MPI_Comm_spawn.

Orchestrator Mode

In orchestrator mode, the TensorRT-LLM backend spawns a single Triton Server process that acts as an orchestrator and spawns one Triton Server process for every GPU that each model requires. This mode is mainly used when serving multiple models with TensorRT-LLM backend. In this mode, the MPI world size must be one as TRT-LLM backend will automatically create new workers as needed. The overview of this mode is described in the diagram below:

Since this mode uses MPI_Comm_spawn, it might not work properly with slurm deployments. Additionally, this currently only works for single node deployments.

Running Multiple Instances of LLaMa Model

Please refer to Running Multiple Instances of the LLaMa Model for more information on running multiple instances of LLaMa model in different configurations.

Multi-node Support

Check out the Multi-Node Generative AI w/ Triton Server and TensorRT-LLM tutorial for Triton Server and TensorRT-LLM multi-node deployment.

Model Parallelism

Tensor Parallelism, Pipeline Parallelism and Expert Parallelism

Tensor Parallelism, Pipeline Parallelism and Expert parallelism are supported in TensorRT-LLM.

See the models in the examples folder for more details on how to build the engines with tensor parallelism, pipeline parallelism and expert parallelism.

Some examples are shown below:

• Build LLaMA v3 70B using 4-way tensor parallelism and 2-way pipeline parallelism.

python convert_checkpoint.py --model_dir ./tmp/llama/70B/hf/ \
                            --output_dir ./tllm_checkpoint_8gpu_tp4_pp2 \
                            --dtype float16 \
                            --tp_size 4 \
                            --pp_size 2

trtllm-build --checkpoint_dir ./tllm_checkpoint_8gpu_tp4_pp2 \
            --output_dir ./tmp/llama/70B/trt_engines/fp16/8-gpu/ \
            --gemm_plugin auto

• Build Mixtral8x22B with tensor parallelism and expert parallelism

python ../llama/convert_checkpoint.py --model_dir ./Mixtral-8x22B-v0.1 \
                             --output_dir ./tllm_checkpoint_mixtral_8gpu \
                             --dtype float16 \
                             --tp_size 8 \
                             --moe_tp_size 2 \
                             --moe_ep_size 4
trtllm-build --checkpoint_dir ./tllm_checkpoint_mixtral_8gpu \
                 --output_dir ./trt_engines/mixtral/tp2ep4 \
                 --gemm_plugin float16

See the doc to learn more about how TensorRT-LLM expert parallelism works in Mixture of Experts (MoE).

MIG Support

See the MIG tutorial for more details on how to run TRT-LLM models and Triton with MIG.

Scheduling

The scheduler policy helps the batch manager adjust how requests are scheduled for execution. There are two scheduler policies supported in TensorRT-LLM, MAX_UTILIZATION and GUARANTEED_NO_EVICT. See the batch manager design to learn more about how scheduler policies work. You can specify the scheduler policy via the batch_scheduler_policy parameter in the model config of tensorrt_llm model.

Key-Value Cache

See the KV Cache section for more details on how TensorRT-LLM supports KV cache. Also, check out the KV Cache Reuse documentation to learn more about how to enable KV cache reuse when building the TRT-LLM engine. Parameters for KV cache can be found in the model config of tensorrt_llm model.

Decoding

Decoding Modes - Top-k, Top-p, Top-k Top-p, Beam Search, Medusa, ReDrafter, Lookahead and Eagle

TensorRT-LLM supports various decoding modes, including top-k, top-p, top-k top-p, beam search Medusa, ReDrafter, Lookahead and Eagle. See the Sampling Parameters section to learn more about top-k, top-p, top-k top-p and beam search decoding. Please refer to the speculative decoding documentation for more details on Medusa, ReDrafter, Lookahead and Eagle.

Parameters for decoding modes can be found in the model config of tensorrt_llm model.

Speculative Decoding

See the Speculative Decoding documentation to learn more about how TensorRT-LLM supports speculative decoding to improve the performance. The parameters for speculative decoding can be found in the model config of tensorrt_llm_bls model.

Chunked Context

For more details on how to use chunked context, please refer to the Chunked Context section. Parameters for chunked context can be found in the model config of tensorrt_llm model.

Quantization

Check out the Quantization Guide to learn more about how to install the quantization toolkit and quantize TensorRT-LLM models. Also, check out the blog post Speed up inference with SOTA quantization techniques in TRT-LLM to learn more about how to speed up inference with quantization.

LoRa

Refer to lora.md for more details on how to use LoRa with TensorRT-LLM and Triton.

Launch Triton server within Slurm based clusters

Prepare some scripts

tensorrt_llm_triton.sub

#!/bin/bash
#SBATCH -o logs/tensorrt_llm.out
#SBATCH -e logs/tensorrt_llm.error
#SBATCH -J <REPLACE WITH YOUR JOB's NAME>
#SBATCH -A <REPLACE WITH YOUR ACCOUNT's NAME>
#SBATCH -p <REPLACE WITH YOUR PARTITION's NAME>
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=8
#SBATCH --time=00:30:00

sudo nvidia-smi -lgc 1410,1410

srun --mpi=pmix \
    --container-image triton_trt_llm \
    --container-mounts /path/to/tensorrtllm_backend:/tensorrtllm_backend \
    --container-workdir /tensorrtllm_backend \
    --output logs/tensorrt_llm_%t.out \
    bash /tensorrtllm_backend/tensorrt_llm_triton.sh

tensorrt_llm_triton.sh

TRITONSERVER="/opt/tritonserver/bin/tritonserver"
MODEL_REPO="/tensorrtllm_backend/triton_model_repo"

${TRITONSERVER} --model-repository=${MODEL_REPO} --disable-auto-complete-config --backend-config=python,shm-region-prefix-name=prefix${SLURM_PROCID}_

If srun initializes the mpi environment, you can use the following command to launch the Triton server:

srun --mpi pmix launch_triton_server.py --oversubscribe

Submit a Slurm job

sbatch tensorrt_llm_triton.sub

You might have to contact your cluster's administrator to help you customize the above script.

Triton Metrics

Starting with the 23.11 release of Triton, users can now obtain TRT LLM Batch Manager statistics by querying the Triton metrics endpoint. This can be accomplished by launching a Triton server in any of the ways described above (ensuring the build code / container is 23.11 or later) and querying the server. Upon receiving a successful response, you can query the metrics endpoint by entering the following:

curl localhost:8002/metrics

Batch manager statistics are reported by the metrics endpoint in fields that are prefixed with nv_trt_llm_. Your output for these fields should look similar to the following (assuming your model is an inflight batcher model):

# HELP nv_trt_llm_request_metrics TRT LLM request metrics
# TYPE nv_trt_llm_request_metrics gauge
nv_trt_llm_request_metrics{model="tensorrt_llm",request_type="context",version="1"} 1
nv_trt_llm_request_metrics{model="tensorrt_llm",request_type="scheduled",version="1"} 1
nv_trt_llm_request_metrics{model="tensorrt_llm",request_type="max",version="1"} 512
nv_trt_llm_request_metrics{model="tensorrt_llm",request_type="active",version="1"} 0
# HELP nv_trt_llm_runtime_memory_metrics TRT LLM runtime memory metrics
# TYPE nv_trt_llm_runtime_memory_metrics gauge
nv_trt_llm_runtime_memory_metrics{memory_type="pinned",model="tensorrt_llm",version="1"} 0
nv_trt_llm_runtime_memory_metrics{memory_type="gpu",model="tensorrt_llm",version="1"} 1610236
nv_trt_llm_runtime_memory_metrics{memory_type="cpu",model="tensorrt_llm",version="1"} 0
# HELP nv_trt_llm_kv_cache_block_metrics TRT LLM KV cache block metrics
# TYPE nv_trt_llm_kv_cache_block_metrics gauge
nv_trt_llm_kv_cache_block_metrics{kv_cache_block_type="tokens_per",model="tensorrt_llm",version="1"} 64
nv_trt_llm_kv_cache_block_metrics{kv_cache_block_type="used",model="tensorrt_llm",version="1"} 1
nv_trt_llm_kv_cache_block_metrics{kv_cache_block_type="free",model="tensorrt_llm",version="1"} 6239
nv_trt_llm_kv_cache_block_metrics{kv_cache_block_type="max",model="tensorrt_llm",version="1"} 6239
# HELP nv_trt_llm_inflight_batcher_metrics TRT LLM inflight_batcher-specific metrics
# TYPE nv_trt_llm_inflight_batcher_metrics gauge
nv_trt_llm_inflight_batcher_metrics{inflight_batcher_specific_metric="micro_batch_id",model="tensorrt_llm",version="1"} 0
nv_trt_llm_inflight_batcher_metrics{inflight_batcher_specific_metric="generation_requests",model="tensorrt_llm",version="1"} 0
nv_trt_llm_inflight_batcher_metrics{inflight_batcher_specific_metric="total_context_tokens",model="tensorrt_llm",version="1"} 0
# HELP nv_trt_llm_general_metrics General TRT LLM metrics
# TYPE nv_trt_llm_general_metrics gauge
nv_trt_llm_general_metrics{general_type="iteration_counter",model="tensorrt_llm",version="1"} 0
nv_trt_llm_general_metrics{general_type="timestamp",model="tensorrt_llm",version="1"} 1700074049
# HELP nv_trt_llm_disaggregated_serving_metrics TRT LLM disaggregated serving metrics
# TYPE nv_trt_llm_disaggregated_serving_metrics counter
nv_trt_llm_disaggregated_serving_metrics{disaggregated_serving_type="kv_cache_transfer_ms",model="tensorrt_llm",version="1"} 0
nv_trt_llm_disaggregated_serving_metrics{disaggregated_serving_type="request_count",model="tensorrt_llm",version="1"} 0

If, instead, you launched a V1 model, your output will look similar to the output above except the inflight batcher related fields will be replaced with something similar to the following:

# HELP nv_trt_llm_v1_metrics TRT LLM v1-specific metrics
# TYPE nv_trt_llm_v1_metrics gauge
nv_trt_llm_v1_metrics{model="tensorrt_llm",v1_specific_metric="total_generation_tokens",version="1"} 20
nv_trt_llm_v1_metrics{model="tensorrt_llm",v1_specific_metric="empty_generation_slots",version="1"} 0
nv_trt_llm_v1_metrics{model="tensorrt_llm",v1_specific_metric="total_context_tokens",version="1"} 5

Please note that versions of Triton prior to the 23.12 release do not support base Triton metrics. As such, the following fields will report 0:

# HELP nv_inference_request_success Number of successful inference requests, all batch sizes
# TYPE nv_inference_request_success counter
nv_inference_request_success{model="tensorrt_llm",version="1"} 0
# HELP nv_inference_request_failure Number of failed inference requests, all batch sizes
# TYPE nv_inference_request_failure counter
nv_inference_request_failure{model="tensorrt_llm",version="1"} 0
# HELP nv_inference_count Number of inferences performed (does not include cached requests)
# TYPE nv_inference_count counter
nv_inference_count{model="tensorrt_llm",version="1"} 0
# HELP nv_inference_exec_count Number of model executions performed (does not include cached requests)
# TYPE nv_inference_exec_count counter
nv_inference_exec_count{model="tensorrt_llm",version="1"} 0
# HELP nv_inference_request_duration_us Cumulative inference request duration in microseconds (includes cached requests)
# TYPE nv_inference_request_duration_us counter
nv_inference_request_duration_us{model="tensorrt_llm",version="1"} 0
# HELP nv_inference_queue_duration_us Cumulative inference queuing duration in microseconds (includes cached requests)
# TYPE nv_inference_queue_duration_us counter
nv_inference_queue_duration_us{model="tensorrt_llm",version="1"} 0
# HELP nv_inference_compute_input_duration_us Cumulative compute input duration in microseconds (does not include cached requests)
# TYPE nv_inference_compute_input_duration_us counter
nv_inference_compute_input_duration_us{model="tensorrt_llm",version="1"} 0
# HELP nv_inference_compute_infer_duration_us Cumulative compute inference duration in microseconds (does not include cached requests)
# TYPE nv_inference_compute_infer_duration_us counter
nv_inference_compute_infer_duration_us{model="tensorrt_llm",version="1"} 0
# HELP nv_inference_compute_output_duration_us Cumulative inference compute output duration in microseconds (does not include cached requests)
# TYPE nv_inference_compute_output_duration_us counter
nv_inference_compute_output_duration_us{model="tensorrt_llm",version="1"} 0
# HELP nv_inference_pending_request_count Instantaneous number of pending requests awaiting execution per-model.
# TYPE nv_inference_pending_request_count gauge
nv_inference_pending_request_count{model="tensorrt_llm",version="1"} 0

Benchmarking

Check out GenAI-Perf tool for benchmarking TensorRT-LLM models.

You can also use the benchmark_core_model script to benchmark the core model tensosrrt_llm. The script sends requests directly to deployed tensorrt_llm model. The benchmark core model latency indicates the inference latency of TensorRT-LLM, not including the pre/post-processing latency which is usually handled by a third-party library such as HuggingFace.

benchmark_core_model can generate traffic from 2 sources. 1 - dataset (json file containing prompts and optional responses) 2 - token normal distribution (user specified input, output seqlen)

By default, exponential distrution is used to control arrival rate of requests. It can be changed to constant arrival time.

cd tools/inflight_batcher_llm

Example: Run dataset with 10 req/sec requested rate with provided tokenizer.

python3 benchmark_core_model.py -i grpc --request_rate 10 dataset --dataset <dataset path> --tokenizer_dir <> --num_requests 5000

Example: Generate I/O seqlen tokens with input normal distribution with mean_seqlen=128, stdev=10. Output normal distribution with mean_seqlen=20, stdev=2. Set stdev=0 to get constant seqlens.

python3 benchmark_core_model.py -i grpc --request_rate 10 token_norm_dist --input_mean 128 --input_stdev 5 --output_mean 20 --output_stdev 2 --num_requests 5000

Expected outputs

[INFO] Warm up for benchmarking.
[INFO] Start benchmarking on 5000 prompts.
[INFO] Total Latency: 26585.349 ms
[INFO] Total request latencies: 11569672.000999955 ms
+----------------------------+----------+
|            Stat            |  Value   |
+----------------------------+----------+
|        Requests/Sec        |  188.09  |
|       OP tokens/sec        | 3857.66  |
|     Avg. latency (ms)      | 2313.93  |
|      P99 latency (ms)      | 3624.95  |
|      P90 latency (ms)      | 3127.75  |
| Avg. IP tokens per request |  128.53  |
| Avg. OP tokens per request |  20.51   |
|     Total latency (ms)     | 26582.72 |
|       Total requests       | 5000.00  |
+----------------------------+----------+

Please note that the expected outputs in that document are only for reference, specific performance numbers depend on the GPU you're using.

Testing the TensorRT-LLM Backend

Please follow the guide in ci/README.md to see how to run the testing for TensorRT-LLM backend.