TheRock/rocm-systems/projects/rocprofiler-sdk/benchmark/README.md

Benchmark Suite

Generate Data

From the current directory:

cmake -B build-benchmark .
cd build-benchmark
export PATH=${PWD}/bin:${PATH}
rocprofv3-benchmark -i ./example.yml -n 2

sqlite3 benchmark.db

SELECT * FROM benchmark_metrics;
SELECT * FROM benchmark_statistics;

Running vLLM Workload

A vLLM benchmark configuration is provided in vllm.yaml to measure profiler overhead on LLM inference workloads.

Note: The model path is currently fixed to /model/Qwen/Qwen3-30B-A3B in the YAML file. To use a different model, edit the --model parameter in vllm.yaml before running.

Prerequisites

• vLLM installed and in PATH
• Model weights available at the specified path (or update path in vllm.yaml)
• Sufficient GPU memory for the model with tensor parallelism