import Tabs from '@theme/Tabs'; import TabItem from '@theme/TabItem';

Quick start

Run your first measurement

llem run --model gpt2 --engine transformers

from llenergymeasure import run_experiment

result = run_experiment(model="gpt2", engine="transformers")
print(result)

On first run, GPT-2 (around 500 MB) downloads from HuggingFace and the engine Docker image is resolved. Subsequent runs use the local cache and typically complete in under two minutes.

A progress indicator prints to stderr. When the experiment finishes, a short summary prints to stdout and a structured result.json is written under results/. Numeric values vary by hardware; the shape is:

Result: gpt2-transformers-bf16-<timestamp>

Energy
  Total          <joules>
  Baseline       <watts>
  Adjusted       <joules - baseline * duration>

Performance
  Throughput     <tokens/sec>
  FLOPs          <estimate>

Timing
  Duration       <wall-clock>
  Warmup         <n prompts excluded>

---

Read the result

Field	What it means
Total (J)	Raw GPU energy across the prompt set
Baseline (W)	Idle GPU power measured before the run
Adjusted (J)	Total minus Baseline x Duration: the energy attributable to inference
Throughput (tok/s)	Output tokens per second across all prompts
FLOPs	Estimated floating-point operations (validity check, not a headline metric)
Duration	Wall-clock time for the full experiment
Warmup	Prompts excluded for thermal stabilisation

:::tip Use Adjusted for cross-experiment comparisons Adjusted isolates inference energy from idle-GPU draw. Use it whenever you are comparing two configurations - engine choice, dtype, batch size - so that ambient GPU power does not inflate the difference. :::

For the full field list in result.json, see Results schema.

---

Result file

Results are written to results/ in your working directory:

results/
└── gpt2-transformers-bf16-<timestamp>/
    └── result.json

The experiment ID encodes the model, engine, dtype, and a timestamp.

---

What's next

• Deeper tutorial - Your first measurement walks through each step with full explanations.
• Understand the result - How to interpret results explains what numbers are normal and how to compare across runs.
• Find the right workflow - Choose your path routes you to the right guide for your use case.