Nightly benchmark

The main goal of this benchmarking is two-fold:

Performance clarity: Provide clarity on which one (vllm, tensorrt-llm, lmdeploy and tgi) leads in performance in what workload.
Reproducible: one can run the exact same set of benchmarking commands inside the exact same docker by following reproducing instructions in reproduce.md.

Docker images

We benchmark vllm, tensorrt-llm, lmdeploy and tgi using the following docker images:

vllm/vllm-openai:v0.5.0.post1
nvcr.io/nvidia/tritonserver:24.04-trtllm-python-py3
openmmlab/lmdeploy:v0.5.0
ghcr.io/huggingface/text-generation-inference:2.1

Hardware

One AWS node with 8x NVIDIA A100 GPUs.

Workload description

We benchmark vllm, tensorrt-llm, lmdeploy and tgi using the following workload:

Input length: randomly sample 500 prompts from ShareGPT dataset (with fixed random seed).
Output length: the corresponding output length of these 500 prompts.
Models: llama-3 8B, llama-3 70B, mixtral 8x7B.
Average QPS (query per second): 4 for the small model (llama-3 8B) and 2 for other two models. For each QPS, the arrival time of each query is determined using a random Poisson process (with fixed random seed).
Evaluation metrics: Throughput (higher the better), TTFT (time to the first token, lower the better), ITL (inter-token latency, lower the better).

Plots

In the following plots, the dot shows the mean and the error bar shows the standard error of the mean. Value 0 means that the corresponding benchmark crashed.

Results

Test name	GPU	Successful req.	Tput (req/s)	Mean TTFT (ms)	Std TTFT (ms)	Mean ITL (ms)	Std ITL (ms)	Input Tput (tok/s)	Output Tput (tok/s)	Engine
tgi_llama8B_tp1_qps_4	A100-SXM4-80GB	500	3.74425	108.705	100.847	16.6054	8.11108	755.553	487.988	tgi
tgi_llama70B_tp4_qps_2	A100-SXM4-80GB	500	1.88792	471.147	299.21	45.5061	26.9567	380.963	335.166	tgi
tgi_mixtral8x7B_tp2_qps_2	A100-SXM4-80GB	500	1.83876	442.111	431.699	38.7051	52.0669	417.204	396.172	tgi
trt_llama70B_tp4_qps_2	A100-SXM4-80GB	500	1.88808	135.072	52.7508	34.9815	11.7329	380.996	284.877	trt
trt_llama8B_tp1_qps_4	A100-SXM4-80GB	500	3.77042	51.2959	16.967	13.6497	4.22742	760.832	575.75	trt
trt_mixtral8x7B_tp2_qps_2	A100-SXM4-80GB	500	1.8578	103.746	23.6912	32.9532	6.10077	421.524	482.858	trt
vllm_llama8B_tp1_qps_4	A100-SXM4-80GB	500	3.74554	20.585	18.9562	16.2997	8.53774	755.813	499.251	vllm
vllm_llama70B_tp4_qps_2	A100-SXM4-80GB	500	1.89946	67.8226	54.8692	43.266	27.2318	383.293	333.633	vllm
vllm_mixtral8x7B_tp2_qps_2	A100-SXM4-80GB	500	1.83413	114.966	268.457	39.1624	55.9786	416.154	395.997	vllm
lmdeploy_llama8B_tp1_qps_4	A100-SXM4-80GB	500	3.79197	20.9606	20.25	12.5817	4.86262	765.181	494.943	lmdeploy
lmdeploy_llama70B_tp4_qps_2	A100-SXM4-80GB	500	1.91245	64.8334	53.9591	34.2332	14.7409	385.913	341.49	lmdeploy
lmdeploy_mixtral8x7B_tp2_qps_2	A100-SXM4-80GB	500	1.74858	1457.06	477.169	60.1654	110.52	396.743	389.27	lmdeploy

Performance Benchmark

make yapf and ruff happy

Buildkite had a problem