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:

• 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

Check nightly-pipeline.yaml artifact for more details.

Workload description

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.

Known issues

• 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.

Plots

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

Results