Veeksha

Veeksha is a high-fidelity benchmarking framework for LLM inference systems. Whether you're optimizing a production deployment, comparing serving backends, or running capacity planning experiments, Veeksha lets you measure what matters to you: realistic multi-turn conversations, agentic workflows, high-frequency stress tests, or targeted microbenchmarks. One tool, any workload.

From isolated requests to complex agentic sessions, Veeksha captures the full complexity of modern LLM workloads.

👉 Why Veeksha? — Learn what sets Veeksha apart
📚 Documentation — Full guides and API reference

Quick start

No install needed; run directly with uvx:

uvx -p 3.14t veeksha benchmark \
    --client.type openai_chat_completions \
    --client.api_base http://localhost:8000/v1 \
    --client.model meta-llama/Llama-3.2-1B-Instruct \
    --traffic_scheduler.type rate \
    --traffic_scheduler.interval_generator.type poisson \
    --traffic_scheduler.interval_generator.arrival_rate 5.0 \
    --runtime.benchmark_timeout 60

Or use a YAML configuration file:

uvx -p 3.14t veeksha benchmark --config my_benchmark.veeksha.yml

Or install with uv pip install veeksha / pip install veeksha and use veeksha directly.

We require free-threaded Python for worker parallelism.

Installation from source

git clone https://github.com/project-vajra/veeksha.git
cd veeksha

# Install uv if needed
curl -LsSf https://astral.sh/uv/install.sh | sh

# Create an environment
uv venv --python 3.14t
source .venv/bin/activate
uv pip install -e .