perf-overview.md

(perf-overview)=

Important

As of TensorRT-LLM v0.10, these performance benchmarks have changed methodology to utilize in-flight batching and no longer utilize static benchmarking. These numbers are initial measurements and are expected to improve in future releases.

Overview

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.

Known Issues

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

Unexpected extra GPU memory allocation when enabling --multiple_profiles

We observed that enabling multiple profiles can lead to extra unexpected GPU memory usage on some cases starting from v0.11. The issue will be addressed in future releases.

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=enable is enabled). There is also a more efficient implementation that runs single Matmul + SwiGLU fused kernel for FP8 on Hopper (when --use_fused_mlp=enable --gemm_swiglu_plugin fp8 is enabled). The gemm_swiglu_plugin will support more data types and GPU architectures in the future release.

Throughput Measurements

The below table shows performance data where a local inference client is fed requests at an infinite rate (no delay between messages), and shows the throughput client-server scenario under maximum load.

The performance numbers below were collected using the steps described in this document.

All data in the table below was generated using version 0.12.0 and presents token throughput in tokens/second.

		GPU	H200 141GB HBM3	GH200 120GB	H100 80GB HBM3	H100 80GB HBM3	A100-SXM4-80GB	L40S
		Precision	FP8	FP8	FP8	Mixed	Mixed	FP8
Model	Input/Output Lengths	TP
GPTJ 6B	128/128	1	24834.76	22454.79	24429.55	13085.91	5864.81	7647.24
	128/2048	1	8348.93	6656.25	7831.38	3882.21	2194.57	1843.91
	128/4096	1	5062.80	3678.91	3968.98	2046.53	1118.22	980.67
	2048/128	1	2776.53	2491.03	2724.38	1488.56	657.01	741.06
	2048/2048	1	3631.54	2994.81	3004.17	1280.54	854.37	754.16
LLaMA v2 7B	128/128	1	19706.35	17803.58	19068.99	11393.48	5272.39	6345.72
	128/2048	1	7651.12	5472.34	6610.03	2964.65	1785.79	1551.37
	128/4096	1	4424.90	3271.61	3649.38	1596.87	957.12	817.24
	2048/128	1	2385.54	2035.42	2271.63	1189.06	564.77	625.09
	2048/2048	1	3191.34	2726.29	2802.41	1243.96	735.19	641.56
LLaMA v3 8B	128/128	1	28288.75	25420.52	27399.75	15567.44	6586.88	8745.80
	128/2048	1	23230.62	16426.68	19198.73	8817.39	4882.13	5084.49
	128/4096	1	16144.44	9832.66	12084.97	5352.37	3079.90	2755.13
	2048/128	1	3623.79	3290.22	3463.26	1852.48	781.63	980.86
	2048/2048	1	11093.62	7573.35	8894.11	3986.83	2268.13	2051.79
Mistral 7B	128/128	1	30223.01	27696.90	29788.46	16319.25	6807.02	9612.58
	128/2048	1	24989.54	17942.29	20509.72	9982.01	5296.02	5444.89
	128/4096	1	17036.14	10846.03	12807.80	5718.89	3241.33	2931.17
	2048/128	1	3678.80	3294.02	3521.71	1887.75	786.43	1002.49
	2048/2048	1	11510.54	8357.75	9214.61	4284.82	2363.25	2154.26
Mixtral 8x7B	128/128	2	24895.03	8785.80	24394.71	15529.86	5921.41
		4	42014.24	38828.53	40197.42	28132.17	11414.95	6820.26
	128/2048	2	29389.21	5474.69	20873.02	7066.02	4306.98
		4	52348.10	41573.66	40588.05	21285.72	10974.83	7467.15
	128/4096	2	21480.27	2277.66	12838.28	3986.01	2400.11
		4	39182.04	28626.55	28337.31	12447.13	7278.89	5233.43
	2048/128	2	2934.44	1003.51	2898.27	1834.77	693.51
		4	5152.40	4724.01	5028.61	3393.18	1362.93	805.49
	2048/2048	2	14029.17	2671.88	10479.45	3531.31	1945.88
		4	25436.05	20302.56	19971.72	9622.66	5221.74	3616.30
LLaMA v3 70B	128/128	2	5386.88			2959.22	1301.14
		4	8944.26	8587.01	8642.05	5966.47	2413.95
		8	16125.20		15397.47	10406.55	4548.32	1364.08
	128/2048	2	7007.27			720.73	500.83
		4	12906.75	10761.53	8978.95	4736.61	2380.02
		8	19417.37		14822.93	6672.14	3815.08	1809.40
	128/4096	2	6183.85			369.29	251.24
		4	8859.54	7270.77	6073.48	2969.99	1634.82
		8	13969.95		10094.57	4358.77	2847.54	1313.78
	2048/128	2	696.59			301.46	140.88
		4	1044.35	1000.55	1022.06	681.72	278.76
		8	2018.47		1933.15	1279.46	543.73	163.36
	2048/2048	2	3525.18				87.54
		4	6550.76	4859.38	4870.26	2379.66	1209.69
		8	9706.95		7670.04	3692.41	2192.28	895.23
LLaMA v2 70B	128/128	2	6355.16			2927.71	1374.05
		4	10818.97	10819.19	10754.99	6603.10	2765.94
		8	16667.25		16074.84	11369.11	4796.89	1402.92
	128/2048	2	6185.77			668.52	445.04
		4	12884.76	11356.48	8870.71	5067.06	2710.53
		8	19053.13		17534.62	8805.16	5665.93	2203.33
	128/4096	2	4873.24			334.10	215.70
		4	8664.90	6311.85	7564.99	3354.02	1884.46
		8	15110.32		10584.03	5373.10	3672.80	1787.76
	2048/128	2	732.09			302.49	141.70
		4	1272.90	1269.58	1265.80	774.93	320.79
		8	2015.77		1943.96	1355.78	569.48	165.52
	2048/2048	2	3508.50			321.95	212.97
		4	6642.69	5545.83	4889.26	2439.10	1276.58
		8	10178.71		8071.77	4275.74	2589.60	1083.45
Falcon 180B	128/128	4	5129.55
		8	8370.98		8268.72
	128/2048	4	7823.79
		8	13278.59		13107.48
	128/4096	4	6374.10
		8	12660.89		10493.79
	2048/128	4	601.67
		8	1002.57		991.22
	2048/2048	4	3869.76
		8	7134.33		6386.83
TP stands for Tensor Parallelism

Reproducing Benchmarked Results

[!NOTE] The only models supported in this workflow are those listed in the table above.

The following tables are references for commands that are used as part of the benchmarking process. For a more detailed description of this benchmarking workflow, see the Benchmarking Suite README.

Commands

Stage	Description	Command
Dataset	Create a synthetic dataset	python benchmarks/cpp/prepare_dataset.py --tokenizer=$model_name --stdout token-norm-dist --num-requests=$num_requests --input-mean=$isl --output-mean=$osl --input-stdev=0 --output-stdev=0 > $dataset_file
Build	Build a TensorRT-LLM engine	trtllm-bench --model $model_name build --tp_size $tp_size --quantization FP8 --dataset $dataset_file
Run	Run a benchmark with a dataset	trtllm-bench --model $model_name throughput --dataset $dataset_file --engine_dir $engine_dir

Variables

Name	Description
$isl	Benchmark input sequence length.
$osl	Benchmark output sequence length.
$tp_size	Number of GPUs to run the benchmark with
$engine_dir	Location to store built engine file (can be deleted after running benchmarks).
$model_name	HuggingFace model name eg. meta-llama/Llama-2-7b-hf or use the path to a local weights directory
$dataset_file	Location of the dataset file generated by prepare_dataset.py
$num_requests	The number of requests to generate for dataset generation
$seq_len	A sequence length of ISL + OSL

Preparing a Dataset

In order to prepare a dataset, you can use the provided script. To generate a synthetic dataset, run the following command:

python benchmarks/cpp/prepare_dataset.py --output=$dataset_file --tokenizer=$model_name token-norm-dist --num-requests=$num_requests --input-mean=$isl --output-mean=$osl --input-stdev=0 --output-stdev=0 > $dataset_file

The command will generate a text file located at the path specified $dataset_file where all requests are of the same input/output sequence length combinations. The script works by using the tokenizer to retrieve the vocabulary size and randomly sample token IDs from it to create entirely random sequences. In the command above, all requests will be uniform because the standard deviations for both input and output sequences are set to 0.

For each input and output sequence length combination, the table below details the $num_requests that were used. For shorter input and output lengths, a larger number of messages were used to guarantee that the system hit a steady state because requests enter and exit the system at a much faster rate. For longer input/output sequence lengths, requests remain in the system longer and therefore require less requests to achieve steady state.

Input Length	Output Length	$seq_len	$num_requests
128	128	256	30000
128	2048	2176	3000
128	4096	4224	1500
2048	128	2176	3000
2048	2048	4096	1500

Engine Building

All engines are built using the trtllm-bench build sub-command. The basic command for FP8 quantized engines is as follows:

trtllm-bench --model $model_name build --tp_size $tp_size --quantization FP8 --dataset $dataset_file

or if you would like to build for a specific sequence length:

trtllm-bench --model $model_name build --tp_size $tp_size --quantization FP8 --max_seq_length $seq_len

If you would like to build an FP16 engine without any quantization, simply remove the --quantization FP8 option.

[!NOTE] If you specify FP8 quantization, the KV cache will automatically be set to FP8 as well!

The trtllm-bench build sub-command will output the path where the engine is located upon a successful build. For example,

===========================================================
ENGINE SAVED: /tmp/meta-llama/Llama-2-7b-hf/tp_1_pp_1
===========================================================

Running the Benchmark

To run the benchmark with the generated data set, simply use the trtllm-bench throughput sub-command. The benchmarker will run an offline maximum throughput scenario such that all requests are queued in rapid succession. You simply need to provide the patch to the engine from the build phase and a generated dataset.

trtllm-bench --model $model_name throughput --dataset $dataset_file --engine_dir $engine_dir

The results will be printed to the terminal upon benchmark completion. For example,

===========================================================
= ENGINE DETAILS
===========================================================
Model:                  meta-llama/Llama-2-7b-hf
Engine Directory:       /tmp/meta-llama/Llama-2-7b-hf/tp_1_pp_1
TensorRT-LLM Version:   0.12.0
Dtype:                  float16
KV Cache Dtype:         FP8
Quantization:           FP8
Max Input Length:       2048
Max Sequence Length:    4098

===========================================================
= WORLD + RUNTIME INFORMATION
===========================================================
TP Size:                1
PP Size:                1
Max Runtime Batch Size: 4096
Max Runtime Tokens:     8192
Scheduling Policy:      Guaranteed No Evict
KV Memory Percentage:   99.0%
Issue Rate (req/sec):   3.680275266452667e+18
===========================================================
= STATISTICS
===========================================================
Number of requests:             3000
Average Input Length (tokens):  128.0
Average Output Length (tokens): 128.0
Token Throughput (tokens/sec):  23405.927228471104
Request Throughput (req/sec):   182.8588064724305
Total Latency (seconds):        16.406100739
===========================================================

[!WARNING] In some cases, the benchmarker may not print anything at all. This behavior usually means that the benchmark has hit an out of memory issue. Try reducing the KV cache percentage using the --kv_cache_free_gpu_mem_fraction option to lower the percentage of used memory.