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.

Versions

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

Check nightly-pipeline.yaml artifact for more details.

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

Input length: randomly sample 1000 prompts from ShareGPT dataset (with fixed random seed).
Output length: the corresponding output length of these 1000 prompts.
Batch size: dynamically determined by vllm and the arrival pattern of the requests.
Average QPS (query per second): 4 for 8B model and 2 for larger models. For each QPS, the arrival time of each query is determined using a random Poisson process (with fixed random seed).
Models: llama-3 8B, llama-3 70B, mixtral 8x7B.
Evaluation metrics: Throughput, TTFT (time to the first token, with mean and std), ITL (inter-token latency, with mean and std).

Check nightly-tests.json artifact for more details.

TGI v2.1 crashes when running mixtral model, see tgi issue #2122
pin the transformers library to 4.41.2 to avoid lmdeploy missing cache_position error, see lmdeploy issue 1885.

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

Test name	GPU	Successful req.	Tput (req/s)	Mean TTFT (ms)	Std TTFT (ms)	Mean ITL (ms)	Std ITL (ms)	Engine
trt_mixtral8x7B_tp2_qps_2	A100-SXM4-80GB	500	1.85765	103.951	23.5344	32.9535	6.03545	trt
lmdeploy_mixtral8x7B_tp2_qps_2	A100-SXM4-80GB	500	1.72606	1462.69	488.098	60.7996	111.024	lmdeploy