Quickstart¶
This guide walks through running your first profiling session and viewing the results.
Single-Turn Profiling¶
Single-turn profiling sends prompts to an inference server one at a time, capturing energy telemetry for each query.
1. Start your inference server¶
2. Run a profiling session¶
ipw profile \
--client ollama \
--model llama3.2:1b \
--client-base-url http://localhost:11434 \
--max-queries 50
This will:
- Launch the energy monitor subprocess
- Connect to the Ollama server
- Send 50 prompts from the default IPW dataset
- Capture per-query telemetry (power, energy, memory, temperature, latency)
- Score responses using an LLM judge
- Compute IPJ and IPW metrics
- Save results to
./runs/profile_<hardware>_<model>/
3. Analyze the results¶
# Compute accuracy and efficiency metrics
ipw analyze ./runs/profile_*
# Fit regression curves for energy vs. input/output length
ipw analyze ./runs/profile_* --analysis regression
4. Generate plots¶
Plots are saved to ./runs/profile_*/plots/.
Agentic Profiling¶
Agentic profiling runs multi-turn agent workloads where the model can use tools, reason over multiple steps, and interact with external systems.
1. Install an agent extra¶
2. Set your API key¶
The ReAct agent calls a cloud LLM API:
3. Run an agentic profiling session¶
This will:
- Start the energy monitor
- Initialize the ReAct agent with the specified model
- Iterate over GAIA benchmark questions
- For each question: run the agent, record per-turn traces, capture telemetry
- Export results as JSONL traces and HuggingFace Arrow datasets
- Run accuracy analysis
4. View results¶
Listing Available Components¶
Use ipw list to discover what is available:
# List everything
ipw list all
# List specific categories
ipw list clients # ollama, vllm, openai
ipw list datasets # ipw, mmlu-pro, supergpqa, gaia, frames, ...
ipw list analyses # accuracy, regression
ipw list visualizations # regression, output-kde
Choosing a Dataset¶
Different datasets test different capabilities:
# General knowledge (built-in, no download needed)
ipw profile --dataset ipw ...
# Multiple-choice academic knowledge
ipw profile --dataset mmlu-pro ...
# Multi-hop reasoning (agentic)
ipw run --dataset frames ...
# Terminal/CLI tasks (requires Terminus agent)
ipw run --agent terminus --dataset terminalbench ...
# Software engineering (requires agent)
ipw run --agent openhands --dataset swebench ...
Next Steps¶
- Read the Profiling Guide for all
ipw profileoptions - Explore the Datasets Overview for benchmark details
- Set up Cost Tracking for cloud API models
- Check Platform Support for your hardware