performance.md

Performance of TensorRT-LLM

This document summarizes performance measurements of TensorRT-LLM on H100 (Hopper), L40S (Ada) and A100 (Ampere) GPUs for a few key models.

The data in the following tables is provided as a reference point to help users validate observed performance. It should not be considered as the peak performance that can be delivered by TensorRT-LLM.

Methodology

The different performance numbers below were collected using the methodology described in the benchmarks folder.

Peak Throughput

The below tables provide reference data at large batch sizes, representing high throughput offline tasks.

This data has been updated for v0.6.1, unless specified.

H100 GPUs (FP8)

Model	Batch Size	TP (1)	Input Length	Output Length	Throughput (out tok/s/GPU)
GPT-J 6B	1024	1	128	128	26,150
GPT-J 6B	120	1	128	2048	8,011
GPT-J 6B	64	1	2048	128	2,551
GPT-J 6B	64	1	2048	2048	3,327
LLaMA 7B	768	1	128	128	19,694
LLaMA 7B	112	1	128	2048	6,818
LLaMA 7B	80	1	2048	128	2,244
LLaMA 7B	48	1	2048	2048	2,740
LLaMA 70B	1024	2	128	128	2,657
LLaMA 70B	480	4	128	2048	1,486
LLaMA 70B	96	2	2048	128	306
LLaMA 70B	64	2	2048	2048	547
Falcon 180B	1024	4	128	128	987
Falcon 180B	1024	8	128	2048	724
Falcon 180B	64	4	2048	128	112
Falcon 180B	64	4	2048	2048	264

L40S GPUs (FP8)^*

^{* The following data is from TensorRT-LLM v0.5.}

Model	Batch Size	TP (1)	Input Length	Output Length	Throughput (out tok/s/GPU)
GPT-J 6B	64	1	128	128	3,630
GPT-J 6B	64	1	128	2048	1,859
GPT-J 6B	32	1	2048	128	616
GPT-J 6B	32	1	2048	2048	757
LLaMA 7B	64	1	128	128	3,240
LLaMA 7B	64	1	128	2048	1,622
LLaMA 7B	32	1	2048	128	581
LLaMA 7B	16	1	2048	2048	531

A100 GPUs (FP16)

Model	Batch Size	TP (1)	Input Length	Output Length	Throughput (out tok/s/GPU)
GPT-J 6B	512	1	128	128	6,374
GPT-J 6B	120	2	128	2048	2,192
GPT-J 6B	60	1	2048	128	670
GPT-J 6B	64	2	2048	2048	903
LLaMA 7B	384	1	128	128	5,586
LLaMA 7B	60	1	128	2048	1,928
LLaMA 7B	52	1	2048	128	591
LLaMA 7B	64	2	2048	2048	782
LLaMA 70B	1280	4	128	128	670
LLaMA 70B	240	4	128	2048	525
LLaMA 70B	120	4	2048	128	79
Falcon 180B	1024	8	128	128	232
Falcon 180B	128	8	128	2048	180

(1) TP stands for Tensor Parallelism.

Low Latency^**

^{** The following data is from TensorRT-LLM v0.5. Low latency numbers will soon be updated to reflect real time latency with infight-batching.}

The below tables provide reference data at batch size 1 for first token latency, representing end-user's perceived latency for online streaming tasks.

H100 GPUs (FP8)

Model	Batch Size	TP (1)	Input Length	1st Token Latency (ms)
GPT-J 6B	1	1	128	7
GPT-J 6B	1	1	2048	29
LLaMA 7B	1	1	128	7
LLaMA 7B	1	1	2048	36
LLaMA 70B	1	4	128	26
LLaMA 70B	1	4	2048	109
Falcon 180B	1	8	128	27
Falcon 180B	1	8	2048	205

L40S GPUs (FP8)

Model	Batch Size	TP (1)	Input Length	1st Token Latency (ms)
GPT-J 6B	1	1	128	12
GPT-J 6B	1	1	2048	71
LLaMA 7B	1	1	128	14
LLaMA 7B	1	1	2048	73

A100 GPUs (FP16)

Model	Batch Size	TP (1)	Input Length	1st Token Latency (ms)
GPT-J 6B	1	1	128	12
GPT-J 6B	1	1	2048	129
LLaMA 7B	1	1	128	16
LLaMA 7B	1	1	2048	133
LLaMA 70B	1	4	128	47
LLaMA 70B	1	4	2048	377
Falcon 180B	1	8	128	61
Falcon 180B	1	8	2048	509

(1) TP stands for Tensor Parallelism.

Known Issues

The following issues are being addressed to improve the efficiency of TensorRT-LLM.

Fused Matmul + Gated-SiLU (LLaMA)

The current implementation combines two Matmul operations into one Matmul followed by a separate SwiGLU kernel (when --use_fused_mlp is enabled). The future release will include a more efficient implementation that runs single Matmul + SwiGLU fused kernel.

Reproducing Benchmarked Results

Building the TensorRT-LLM Container

---

In order to benchmark TensorRT-LLM, you will need to follow the Quick Start build process to create a baseline container for building a wheel. Additionally, the development container needs a copy of the source code to build the wheel and the benchmarking script. Create the right build environment, use the following :

git clone https://github.com/NVIDIA/TensorRT-LLM.git
cd TensorRT-LLM
git submodule update --init --recursive
git lfs install
git lfs pull
make -C docker build
make -C docker run LOCAL_USER=1

Warning

If you have elevated privileges on your system, then skip the make -C docker run LOCAL_USER=1 command above as it may make it so that you cannot access some required system libraries within the container because the build forces your UID and GID to match those that are set for your non-elevated user. There are cases where the container will be booted as root (i.e. on some SLURM systems with the pyxis plugin) which will cause libraries to be missing.

If you are benchmarking in a shared environment, you need to specify the GPU indices that you would like the container to use, otherwise the Makefile defaults to loading the container with all GPUs on the system. For example, if you only have the 4 higher indices of GPUs on your system you can configure it using the following example:

NV_GPU=0,1,2,3
make -C docker run LOCAL_USER=1 GPU_OPTS='--gpus \"device=${NV_GPU}\"'

Additionally, if you'd like to mount external storage to access persistent storage, or previously built engines, you can mount directories as follows (simply replace source and destination with the appropriate paths):

make -C docker run LOCAL_USER=1 DOCKER_RUN_ARGS="-v /source:/destination"

Once the container starts, you'll need to build the wheel and the benchmarking scripts. From the code root (the default directory when the container is loaded), the following commands will build the TensorRT-LLM wheel, install dependencies, and build the benchmark scripts:

python3 ./scripts/build_wheel.py --benchmarks --trt_root /usr/local/tensorrt
pip install ./build/tensorrt_llm*.whl

Methodology

Engine Building Setups

Each engine needs to be built before they can be benchmarked, and requires the source code for each of their respective build scripts. For smaller models, it is fine to build the engine on the fly in container; however, for larger engines it is recommended to pre-build and mount a directory with the engine because engine files are quite large and take time to repeatedly build. Additionally, built engines can be used for input lengths, output lengths, and batch sizes up to their build options meaning you can use an engine to benchmark multiple input configurations.

In order to benchmark the various networks, our engine building scheme is as follows:

• For the GPT-J, Llama2-7b, and Llama2-70b benchmarks were ran using a single-setting engine build for each network configured for our maximum expected throughput.
• For Falcon-180B, where memory limits and model size have a higher impact for running the model, our benchmarks transition to a per-configuration engine build.

Below we document how to benchmark each model on an H100-HBM3-80GB system and reproduce the throughput numbers we document on our [Performance section](#performance of-tensorrt-llm).

Running on A100

To run the benchmarks below on A100, you will need to remove the --enable_fp8 --fp8_kv_cache options from each engine build command because FP8 computation is a feature in H100 and newer GPUs.

Reproducing First Token Latency

In order to test the latency to the first token, you can build the engines as specified below (or with the tweaks specified above on A100) -- once built as described in the build steps above, you can then benchmark with a single output token in order to find the time to first token latency. We provide the appropriate command lines below for each of the benchmarked models, but you can use this same method to benchmark other models available in TensorRT-LLM.

Benchmarking per Model

Warning

In some cases, using Group Query Attention (GQA) can improve performance of some networks. These kernels are currently experimental and not enabled by default. In order to enable them, simply run export TRTLLM_ENABLE_XQA=1 in your shell. The kernels are an inference runtime optimization, so previously built engines should still function. For the benchmarks below, we have enabled GQA where our tests displayed performance benefits. If your network is not listed below, be sure to try both GQA-enabled and GQA-disabled configurations to find the configuration that works best. For more details see our documentation about GPT Attention.

GPT-J 6B

---

python examples/gptj/build.py \
	--enable_context_fmha \
	--parallel_build \
	--output_dir /tmp/engines/gptj \
	--dtype float16 \
	--use_gpt_attention_plugin float16 \
	--world_size 1 \
	--max_batch_size 64 \
	--max_input_len 2048 \
	--max_output_len 2048 \
	--hidden_act gelu \
	--enable_fp8 \
	--fp8_kv_cache \
	--strongly_typed \
	--n_layer 28 \
	--n_head 16 \
	--n_embd 4096 \
	--n_positions 2048 \
	--enable_two_optimization_profiles

Throughput Benchmark

in_out_sizes=("64:128,128" "64:128,2048" "64:2048,128" "64:2048,2048")
for in_out in ${in_out_sizes[@]}
do
	batch_size=$(echo $in_out | awk -F':' '{ print $1 }')
	in_out_dims=$(echo $in_out | awk -F':' '{ print $2 }')
	echo "BS: $batch_size, ISL/OSL: $in_out_dims"

	./cpp/build/benchmarks/gptSessionBenchmark --model gptj --engine_dir /tmp/engines/gptj/ --warm_up 1 --batch_size $batch_size --duration 0 --num_runs 5 --input_output_len $in_out_dims
done

First Token Latency Benchmark

in_out_sizes=("64:128,1" "64:2048,1")
for in_out in ${in_out_sizes[@]}
do
	batch_size=$(echo $in_out | awk -F':' '{ print $1 }')
	in_out_dims=$(echo $in_out | awk -F':' '{ print $2 }')
	echo "BS: $batch_size, ISL/OSL: $in_out_dims"

	./cpp/build/benchmarks/gptSessionBenchmark --model gptj --engine_dir /tmp/engines/gptj/ --warm_up 1 --batch_size $batch_size --duration 0 --num_runs 5 --input_output_len $in_out_dims
done

Llama2-7b

---

pip install -r examples/llama/requirements.txt
python examples/llama/build.py \
	--remove_input_padding \
	--enable_context_fmha \
	--parallel_build \
	--output_dir /tmp/engines/llama/7b \
	--dtype float16 \
	--use_gpt_attention_plugin float16 \
	--world_size 1 \
	--tp_size 1 \
	--pp_size 1 \
	--max_batch_size 64 \
	--max_input_len 2048 \
	--max_output_len 2048 \
	--enable_fp8 \
	--fp8_kv_cache \
	--strongly_typed \
	--n_layer 32 \
	--n_head 32 \
	--n_embd 4096 \
	--inter_size 11008 \
	--vocab_size 32000 \
	--n_positions 4096 \
	--hidden_act silu

Throughput Benchmark

in_out_sizes=("64:128,128" "64:128,2048" "64:2048,128" "32:2048,2048")
for in_out in ${in_out_sizes[@]}
do
	batch_size=$(echo $in_out | awk -F':' '{ print $1 }')
	in_out_dims=$(echo $in_out | awk -F':' '{ print $2 }')
	echo "BS: $batch_size, ISL/OSL: $in_out_dims"

	./cpp/build/benchmarks/gptSessionBenchmark --model llama --engine_dir /tmp/engines/llama/7b --warm_up 1 --batch_size $batch_size --duration 0 --num_runs 5 --input_output_len $in_out_dims
done

First Token Latency Benchmark

in_out_sizes=("64:128,1" "32:2048,1")
for in_out in ${in_out_sizes[@]}
do
	batch_size=$(echo $in_out | awk -F':' '{ print $1 }')
	in_out_dims=$(echo $in_out | awk -F':' '{ print $2 }')
	echo "BS: $batch_size, ISL/OSL: $in_out_dims"

	./cpp/build/benchmarks/gptSessionBenchmark --model llama --engine_dir /tmp/engines/llama/7b --warm_up 1 --batch_size $batch_size --duration 0 --num_runs 5 --input_output_len $in_out_dims
done

Llama2-70b

pip install -r examples/llama/requirements.txt
python examples/llama/build.py \
	--remove_input_padding \
	--enable_context_fmha \
	--parallel_build \
	--output_dir /tmp/engines/llama/70b \
	--dtype float16 \
	--use_gpt_attention_plugin float16 \
	--world_size 4 \
	--tp_size 4 \
	--pp_size 1 \
	--max_batch_size 64 \
	--max_input_len 2048 \
	--max_output_len 2048 \
	--enable_fp8 \
	--fp8_kv_cache \
	--strongly_typed \
	--n_layer 80 \
	--n_head 64 \
	--n_kv_head 8 \
	--n_embd 8192 \
	--inter_size 28672 \
	--vocab_size 32000 \
	--n_positions 4096 \
	--hidden_act silu \
	--ffn_dim_multiplier 1.3 \
	--multiple_of 4096

Throughput Benchmark

export TRTLLM_ENABLE_XQA=1
in_out_sizes=("64:128,128" "64:128,2048" "64:2048,128" "64:2048,2048")
for in_out in ${in_out_sizes[@]}
do
	batch_size=$(echo $in_out | awk -F':' '{ print $1 }')
	in_out_dims=$(echo $in_out | awk -F':' '{ print $2 }')
	echo "BS: $batch_size, ISL/OSL: $in_out_dims"

	mpirun -n 4 --allow-run-as-root --oversubscribe ./cpp/build/benchmarks/gptSessionBenchmark --model llama --engine_dir /tmp/engines/llama/70b --warm_up 1 --batch_size $batch_size --duration 0 --num_runs 5 --input_output_len $in_out_dims
done

First Token Latency Benchmark

export TRTLLM_ENABLE_XQA=1
in_out_sizes=("64:128,1" "64:128,1")
for in_out in ${in_out_sizes[@]}
do
	batch_size=$(echo $in_out | awk -F':' '{ print $1 }')
	in_out_dims=$(echo $in_out | awk -F':' '{ print $2 }')
	echo "BS: $batch_size, ISL/OSL: $in_out_dims"

	mpirun -n 4 --allow-run-as-root --oversubscribe ./cpp/build/benchmarks/gptSessionBenchmark --model llama --engine_dir /tmp/engines/llama/70b --warm_up 1 --batch_size $batch_size --duration 0 --num_runs 5 --input_output_len $in_out_dims
done

Falcon-180B

---

Benchmarking Falcon-180B requires a custom engine per batch size, input/output sequence length due to the large footprint of the model and the large input size of 2048. You can build and benchmark each engine one at a time with the following loop.

export TRTLLM_ENABLE_XQA=1
# Benchmark specific batch size:isl:osl combinations.
in_out_sizes=("96:128,128" "96:128,2048" "64:2048,128")
for in_out in ${in_out_sizes[@]}
do
	batch_size=$(echo $in_out | awk -F':' '{ print $1 }')
	in_out_dims=$(echo $in_out | awk -F':' '{ print $2 }')
	isl=$(echo $in_out_dims | awk -F',' '{ print $1 }')
	osl=$(echo $in_out_dims | awk -F',' '{ print $2 }')
	engine_path="/tmp/engines/falcon/180b/${batch_size}_${isl}_${osl}"
	echo "BS: $batch_size, ISL/OSL: ${isl},${osl}"

	# Build the specific engine for the BS,ISL,OSL combination
	python examples/falcon/build.py \
		--use_inflight_batching \
		--paged_kv_cache \
		--remove_input_padding \
		--enable_context_fmha \
		--parallel_build \
		--output_dir $engine_path \
		--dtype float16 \
		--use_gemm_plugin float16 \
		--use_gpt_attention_plugin float16 \
		--world_size 8 \
		--tp 8 \
		--max_batch_size $batch_size \
		--max_input_len $isl \
		--max_output_len $osl \
		--enable_fp8 \
		--fp8_kv_cache \
		--n_layer 80 \
		--n_head 232 \
		--n_kv_head 8 \
		--n_embd 14848 \
		--vocab_size 65024 \
		--new_decoder_architecture
	# Throughput benchmark
	mpirun -n 8 --allow-run-as-root --oversubscribe ./cpp/build/benchmarks/gptSessionBenchmark --model falcon --engine_dir $engine_path --warm_up 1 --batch_size $batch_size --duration 0 --num_runs 5 --input_output_len "${isl},${osl}"
	# Time to first token benchmark
	mpirun -n 8 --allow-run-as-root --oversubscribe ./cpp/build/benchmarks/gptSessionBenchmark --model falcon --engine_dir $engine_path --warm_up 1 --batch_size $batch_size --duration 0 --num_runs 5 --input_output_len "${isl},1"

	# The Falcon-180b engine is quite large, remove after the benchmark to free up space
	# Remove this line if you'd like to save the engines.
	rm -r $engine_path
done