7.4.2
Compute Cost and Choosing Group Size
Understand the compute requirements and recommended use cases for different GRPO group sizes.
What is it?
The appropriate group size depends on both task difficulty and compute availability. The source recommends different configurations for easy tasks, general training and very difficult tasks.
Key points
G = 2: Low compute, but never recommended because the signal is too noisy.
G = 4: Moderate compute; quick experiments and easy tasks with pass rate above 50%.
G = 8: High compute; default balance for most tasks.
G = 16: Very high compute; hard tasks with pass rate below 20%.
G = 32: Extreme compute; only for very hard tasks with massive compute resources.
Selecting a Group Size for a Task
For a quick experiment on an easy task, G = 4 may be sufficient. For most tasks, G = 8 is the recommended default. For a difficult task with a pass rate below 20%, G = 16 can provide more attempts to obtain positive examples.
Real-world application
Use the task pass rate and available compute to select an appropriate group size. Avoid selecting a large group without considering its additional generation cost.
Key takeaway
Group size should balance the quality of the learning signal against compute requirements.