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.

Accessing the TensorRT-LLM Backend

There are several ways to access the TensorRT-LLM Backend.

Run the Pre-built Docker Container

Starting with Triton 23.10 release, Triton includes a container with the TensorRT-LLM Backend and Python Backend. This container should have everything to run a TensorRT-LLM model. You can find this container on the Triton NGC page.

Build the Docker Container

Option 1. Build via the build.py Script in Server Repo

Starting with Triton 23.10 release, you can follow steps described in the Building With Docker guide and use the build.py script to build the TRT-LLM backend.

The below commands will build the same Triton TRT-LLM container as the one on the NGC.

# Prepare the TRT-LLM base image using the dockerfile from tensorrtllm_backend.
cd tensorrtllm_backend
git lfs install
git submodule update --init --recursive

# Specify the build args for the dockerfile.
BASE_IMAGE=nvcr.io/nvidia/pytorch:24.04-py3
TRT_VERSION=10.0.1.6
TRT_URL_x86=https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.0.1/tars/TensorRT-10.0.1.6.Linux.x86_64-gnu.cuda-12.4.tar.gz
TRT_URL_ARM=https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.0.1/tars/TensorRT-10.0.1.6.ubuntu-22.04.aarch64-gnu.cuda-12.4.tar.gz

docker build -t trtllm_base \
             --build-arg BASE_IMAGE="${BASE_IMAGE}" \
             --build-arg TRT_VER="${TRT_VERSION}" \
             --build-arg RELEASE_URL_TRT_x86="${TRT_URL_x86}" \
             --build-arg RELEASE_URL_TRT_ARM="${TRT_URL_ARM}" \
             -f dockerfile/Dockerfile.triton.trt_llm_backend .

# Run the build script from Triton Server repo. The flags for some features or
# endpoints can be removed if not needed. Please refer to the support matrix to
# see the aligned versions: https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html
TRTLLM_BASE_IMAGE=trtllm_base
TENSORRTLLM_BACKEND_REPO_TAG=rel
PYTHON_BACKEND_REPO_TAG=r24.04

cd server
./build.py -v --no-container-interactive --enable-logging --enable-stats --enable-tracing \
              --enable-metrics --enable-gpu-metrics --enable-cpu-metrics \
              --filesystem=gcs --filesystem=s3 --filesystem=azure_storage \
              --endpoint=http --endpoint=grpc --endpoint=sagemaker --endpoint=vertex-ai \
              --backend=ensemble --enable-gpu --endpoint=http --endpoint=grpc \
              --no-container-pull \
              --image=base,${TRTLLM_BASE_IMAGE} \
              --backend=tensorrtllm:${TENSORRTLLM_BACKEND_REPO_TAG} \
              --backend=python:${PYTHON_BACKEND_REPO_TAG}

The TRTLLM_BASE_IMAGE is the base image that will be used to build the container. The TENSORRTLLM_BACKEND_REPO_TAG and PYTHON_BACKEND_REPO_TAG are the tags of the TensorRT-LLM backend and Python backend repositories that will be used to build the container. You can also remove the features or endpoints that you don't need by removing the corresponding flags.

Option 2. Build via Docker

The version of Triton Server used in this build option can be found in the Dockerfile.

# Update the submodules
cd tensorrtllm_backend
git lfs install
git submodule update --init --recursive

# Use the Dockerfile to build the backend in a container
# For x86_64
DOCKER_BUILDKIT=1 docker build -t triton_trt_llm -f dockerfile/Dockerfile.trt_llm_backend .
# For aarch64
DOCKER_BUILDKIT=1 docker build -t triton_trt_llm --build-arg TORCH_INSTALL_TYPE="src_non_cxx11_abi" -f dockerfile/Dockerfile.trt_llm_backend .

Using the TensorRT-LLM Backend

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.

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 deployment.

# Update the submodule TensorRT-LLM repository
git submodule update --init --recursive
git lfs install
git lfs pull

# TensorRT-LLM is required for generating engines. You can skip this step if
# you already have the package installed. If you are generating engines within
# the Triton container, you have to install the TRT-LLM package.
(cd tensorrt_llm &&
    bash docker/common/install_cmake.sh &&
    export PATH=/usr/local/cmake/bin:$PATH &&
    python3 ./scripts/build_wheel.py --trt_root="/usr/local/tensorrt" &&
    pip3 install ./build/tensorrt_llm*.whl)

# Go to the tensorrt_llm/examples/gpt directory
cd 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 \
        --paged_kv_cache enable \
        --gemm_plugin float16 \
        --output_dir engines/fp16/4-gpu

Create the model repository

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

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).

ensemble

This model can be used to chain the preprocessing, tensorrt_llm and postprocessing models together.

tensorrt_llm_bls

This model can also be used to chain the preprocessing, tensorrt_llm and postprocessing models together.

When using the BLS model instead of the ensemble, you should set the number of model instances to the maximum batch size supported by the TRT engine to allow concurrent request execution. This can be done by modifying the count value in the instance_group section of the BLS model config.pbtxt.

The BLS model has an optional parameter accumulate_tokens which can be used in streaming mode to call the postprocessing model with all accumulated tokens, instead of only one token. This might be necessary for certain tokenizers.

The BLS model supports speculative decoding. Target and draft triton models are set with the parameters tensorrt_llm_model_name tensorrt_llm_draft_model_name. Speculative decoding is performed by setting num_draft_tokens in the request. use_draft_logits may be set to use logits comparison speculative decoding. Note that return_generation_logits and return_context_logits are not supported when using speculative decoding.

BLS Inputs

Name	Shape	Type	Description
text_input	[ -1 ]	string	Prompt text
max_tokens	[ -1 ]	int32	number of tokens to generate
bad_words	[2, num_bad_words]	int32	Bad words list
stop_words	[2, num_stop_words]	int32	Stop words list
end_id	[1]	int32	End token Id. If not specified, defaults to -1
pad_id	[1]	int32	Pad token Id
temperature	[1]	float32	Sampling Config param: temperature
top_k	[1]	int32	Sampling Config param: topK
top_p	[1]	float32	Sampling Config param: topP
len_penalty	[1]	float32	Sampling Config param: lengthPenalty
repetition_penalty	[1]	float	Sampling Config param: repetitionPenalty
min_length	[1]	int32_t	Sampling Config param: minLength
presence_penalty	[1]	float	Sampling Config param: presencePenalty
frequency_penalty	[1]	float	Sampling Config param: frequencyPenalty
random_seed	[1]	uint64_t	Sampling Config param: randomSeed
return_log_probs	[1]	bool	When true, include log probs in the output
return_context_logits	[1]	bool	When true, include context logits in the output
return_generation_logits	[1]	bool	When true, include generation logits in the output
beam_width	[1]	int32_t	(Default=1) Beam width for this request; set to 1 for greedy sampling
stream	[1]	bool	(Default=false). When true, stream out tokens as they are generated. When false return only when the full generation has completed.
prompt_embedding_table	[1]	float16 (model data type)	P-tuning prompt embedding table
prompt_vocab_size	[1]	int32	P-tuning prompt vocab size
lora_task_id	[1]	uint64	Task ID for the given lora_weights. This ID is expected to be globally unique. To perform inference with a specific LoRA for the first time lora_task_id lora_weights and lora_config must all be given. The LoRA will be cached, so that subsequent requests for the same task only require lora_task_id. If the cache is full the oldest LoRA will be evicted to make space for new ones. An error is returned if lora_task_id is not cached
lora_weights	[ num_lora_modules_layers, D x Hi + Ho x D ]	float (model data type)	weights for a lora adapter. see lora docs for more details.
lora_config	[ num_lora_modules_layers, 3]	int32t	lora configuration tensor. [ module_id, layer_idx, adapter_size (D aka R value) ] see lora docs for more details.
embedding_bias_words	[-1]	string	Embedding bias words
embedding_bias_weights	[-1]	float32	Embedding bias weights
num_draft_tokens	[1]	int32	number of tokens to get from draft model during speculative decoding
use_draft_logits	[1]	bool	use logit comparison during speculative decoding

BLS Outputs

Name	Shape	Type	Description
text_output	[-1]	string	text output
cum_log_probs	[-1]	float	cumulative probabilities for each output
output_log_probs	[beam_width, -1]	float	log probabilities for each output
context_logits	[-1, vocab_size]	float	context logits for input
generation_logtis	[beam_width, seq_len, vocab_size]	float	generatiion logits for each output

To learn more about ensemble and BLS models, please see the Ensemble Models and Business Logic Scripting sections of the Triton Inference Server documentation.

# Create the model repository that will be used by the Triton server
cd tensorrtllm_backend
mkdir triton_model_repo

# Copy the example models to the model repository
cp -r all_models/inflight_batcher_llm/* triton_model_repo/

# Copy the TRT engine to triton_model_repo/tensorrt_llm/1/
cp tensorrt_llm/examples/gpt/engines/fp16/4-gpu/* triton_model_repo/tensorrt_llm/1

Modify the model configuration

The following table shows the fields that may to be modified before deployment:

triton_model_repo/preprocessing/config.pbtxt

Name	Description
tokenizer_dir	The path to the tokenizer for the model. In this example, the path should be set to /tensorrtllm_backend/tensorrt_llm/examples/gpt/gpt2 as the tensorrtllm_backend directory will be mounted to /tensorrtllm_backend within the container

triton_model_repo/tensorrt_llm/config.pbtxt

Name	Description
gpt_model_type	Mandatory. Set to inflight_fused_batching when enabling in-flight batching support. To disable in-flight batching, set to V1
gpt_model_path	Mandatory. Path to the TensorRT-LLM engines for deployment. In this example, the path should be set to /tensorrtllm_backend/triton_model_repo/tensorrt_llm/1 as the tensorrtllm_backend directory will be mounted to /tensorrtllm_backend within the container
batch_scheduler_policy	Mandatory. Set to max_utilization to greedily pack as many requests as possible in each current in-flight batching iteration. This maximizes the throughput but may result in overheads due to request pause/resume if KV cache limits are reached during execution. Set to guaranteed_no_evict to guarantee that a started request is never paused.
decoupled	Optional (default=false). Controls streaming. Decoupled mode must be set to true if using the streaming option from the client.
max_beam_width	Optional (default=1). The maximum beam width that any request may ask for when using beam search.
max_tokens_in_paged_kv_cache	Optional (default=unspecified). The maximum size of the KV cache in number of tokens. If unspecified, value is interpreted as 'infinite'. KV cache allocation is the min of max_tokens_in_paged_kv_cache and value derived from kv_cache_free_gpu_mem_fraction below.
max_attention_window_size	Optional (default=max_sequence_length). When using techniques like sliding window attention, the maximum number of tokens that are attended to generate one token. Defaults attends to all tokens in sequence.
kv_cache_free_gpu_mem_fraction	Optional (default=0.9). Set to a number between 0 and 1 to indicate the maximum fraction of GPU memory (after loading the model) that may be used for KV cache.
exclude_input_in_output	Optional (default=false). Set to true to only return completion tokens in a response. Set to false to return the prompt tokens concatenated with the generated tokens
cancellation_check_period_ms	Optional (default=100). The time for cancellation check thread to sleep before doing the next check. It checks if any of the current active requests are cancelled through triton and prevent further execution of them.
stats_check_period_ms	Optional (default=100). The time for the statistics reporting thread to sleep before doing the next check.
iter_stats_max_iterations	Optional (default=executor::kDefaultIterStatsMaxIterations). The numbers of iteration stats to be kept.
request_stats_max_iterations	Optional (default=executor::kDefaultRequestStatsMaxIterations). The numbers of request stats to be kept.
normalize_log_probs	Optional (default=true). Set to false to skip normalization of output_log_probs
enable_chunked_context	Optional (default=false). Set to true to enable context chunking.
gpu_device_ids	Optional (default=unspecified). Comma-separated list of GPU IDs to use for this model. If not provided, the model will use all visible GPUs.
decoding_mode	Optional. Set to one of the following: {top_k, top_p, top_k_top_p, beam_search, medusa} to select the decoding mode. The top_k mode exclusively uses Top-K algorithm for sampling, The top_p mode uses exclusively Top-P algorithm for sampling. The top_k_top_p mode employs both Top-K and Top-P algorithms, depending on the runtime sampling params of the request. Note that the top_k_top_p option requires more memory and has a longer runtime than using top_k or top_p individually; therefore, it should be used only when necessary. beam_search uses beam search algorithm. If not specified, the default is to use top_k_top_p if max_beam_width == 1; otherwise, beam_search is used. When Medusa model is used, medusa decoding mode should be set. However, TensorRT-LLM detects loaded Medusa model and overwrites decoding mode to medusa with warning.
medusa_choices	Optional. To specify Medusa choices tree in the format of e.g. "{0, 0, 0}, {0, 1}". By default, mc_sim_7b_63 choices are used.
lora_cache_optimal_adapter_size	Optional (default=8) Optimal adapter size used to size cache pages. Typically optimally sized adapters will fix exactly into 1 cache page.
lora_cache_max_adapter_size	Optional (default=64) Used to set the minimum size of a cache page. Pages must be at least large enough to fit a single module, single later adapter_size maxAdapterSize row of weights.
lora_cache_gpu_memory_fraction	Optional (default=0.05) Fraction of GPU memory used for LoRA cache. Computed as a fraction of left over memory after engine load, and after KV cache is loaded
lora_cache_host_memory_bytes	Optional (default=1G) Size of host LoRA cache in bytes
gpu_weights_percent	Optional (default=1.0). Set to a number between 0.0 and 1.0 to specify the percentage of weights that reside on GPU instead of CPU and streaming load during runtime. Values less than 1.0 are only supported for an engine built with weight_streaming on.

triton_model_repo/postprocessing/config.pbtxt

Name	Description
tokenizer_dir	The path to the tokenizer for the model. In this example, the path should be set to /tensorrtllm_backend/tensorrt_llm/examples/gpt/gpt2 as the tensorrtllm_backend directory will be mounted to /tensorrtllm_backend within the container
tokenizer_type	The type of the tokenizer for the model, t5, auto and llama are supported. In this example, the type should be set to auto

Launch Triton server

Please follow the option corresponding to the way you build the TensorRT-LLM backend.

Option 1. Launch Triton server within Triton NGC container

docker run --rm -it --net host --shm-size=2g --ulimit memlock=-1 --ulimit stack=67108864 --gpus all -v /path/to/tensorrtllm_backend:/tensorrtllm_backend nvcr.io/nvidia/tritonserver:23.10-trtllm-python-py3 bash

Option 2. Launch Triton server within the Triton container built via build.py script

docker run --rm -it --net host --shm-size=2g --ulimit memlock=-1 --ulimit stack=67108864 --gpus all -v /path/to/tensorrtllm_backend:/tensorrtllm_backend tritonserver bash

Option 3. Launch Triton server within the Triton container built via Docker

docker run --rm -it --net host --shm-size=2g --ulimit memlock=-1 --ulimit stack=67108864 --gpus all -v /path/to/tensorrtllm_backend:/tensorrtllm_backend triton_trt_llm bash

Once inside the container, you can launch the Triton server with the following command:

cd /tensorrtllm_backend
# --world_size is the number of GPUs you want to use for serving
python3 scripts/launch_triton_server.py --world_size=4 --model_repo=/tensorrtllm_backend/triton_model_repo

In order to use multiple TensorRT-LLM models, use the --multi-model option. The --world_size must be 1 as the TensorRT-LLM backend will dynamically launch TensorRT-LLM workers as needed.

cd /tensorrtllm_backend
python3 scripts/launch_triton_server.py --model_repo=/tensorrtllm_backend/triton_model_repo --multi-model

When using the --multi-model option, the Triton model repository can contain multiple TensorRT-LLM models. When running multiple TensorRT-LLM models, the gpu_device_ids parameter should be specified in the models config.pbtxt configuration files. It is up to you to ensure there is no overlap between allocated GPU IDs.

When successfully deployed, the server produces logs similar to the following ones.

I0919 14:52:10.475738 293 grpc_server.cc:2451] Started GRPCInferenceService at 0.0.0.0:8001
I0919 14:52:10.475968 293 http_server.cc:3558] Started HTTPService at 0.0.0.0:8000
I0919 14:52:10.517138 293 http_server.cc:187] Started Metrics Service at 0.0.0.0:8002

Query the server with the Triton generate endpoint

Starting with Triton 23.10 release, you can query the server using Triton's generate endpoint with a curl command based on the following general format within your client environment/container:

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:

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 ensemble model or

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": ""}'

if using the tensorrt_llm_bls model.

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"
}

Utilize the provided client script to send a request

You can send requests to the "tensorrt_llm" model with the provided python client script as following:

python3 inflight_batcher_llm/client/inflight_batcher_llm_client.py --request-output-len 200 --tokenizer-dir /workspace/tensorrtllm_backend/tensorrt_llm/examples/gpt/gpt2

The result should be similar to the following:

Got completed request
output_ids =  [[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, 6342, 11, 4881, 11, 284, 257, 4141, 2988, 290, 257, 2679, 2802, 13, 198, 198, 1544, 373, 15657, 379, 262, 23566, 38719, 293, 748, 1355, 14644, 12, 3163, 912, 287, 6342, 290, 262, 15423, 4189, 710, 287, 6342, 13, 198, 198, 1544, 373, 257, 2888, 286, 262, 4141, 8581, 286, 13473, 290, 262, 4141, 8581, 286, 11536, 13, 198, 198, 1544, 373, 257, 2888, 286, 262, 4141, 8581, 286, 13473, 290, 262, 4141, 8581, 286, 11536, 13, 198, 198, 50, 726, 263, 373, 257, 2888, 286, 262, 4141, 8581, 286, 13473, 290]]
Input: Born in north-east France, Soyer trained as a
Output:  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 Paris, France, to a French father and a German mother.

He was educated at the prestigious Ecole des Beaux-Arts in Paris and the Sorbonne in Paris.

He was a member of the French Academy of Sciences and the French Academy of Arts.

He was a member of the French Academy of Sciences and the French Academy of Arts.

Soyer was a member of the French Academy of Sciences and

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:

python inflight_batcher_llm/client/inflight_batcher_llm_client.py --stop-after-ms 200 --request-output-len 200 --tokenizer-dir /workspace/tensorrtllm_backend/tensorrt_llm/examples/gpt/gpt2

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/inflight_batcher_llm_client.py --request-output-len 20 --tokenizer-dir /path/to/tokenizer/ \
--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]]]]

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}_

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.

Kill the Triton server

pkill tritonserver

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

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

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:

1. Leader mode
2. Orchestrator mode

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.

Testing the TensorRT-LLM Backend

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