Skip to content

Estimate Configuration

The Estimate Configuration section controls how WarpRec samples batches during the estimate pipeline.

This block does not change the model architecture or evaluation protocol. Instead, it defines how much lightweight profiling WarpRec performs before extrapolating time and memory usage.

Available Keywords

  • train_batches: Number of measured training batches used to estimate iterative-model training cost. Defaults to 100.
  • eval_batches: Number of measured evaluation batches used to estimate inference and evaluation cost. Defaults to 100.
  • warmup_batches: Number of initial batches ignored before measurements start. Defaults to 10.

Important

train_batches and eval_batches must be strictly positive. warmup_batches must be greater than or equal to zero.

Example Estimate Configuration

estimate:
    warmup_batches: 10
    train_batches: 100
    eval_batches: 100

Practical Guidance

  • Increase train_batches or eval_batches when you want more stable estimates on highly variable workloads.
  • Reduce them when you want faster, cheaper pre-flight checks.
  • Use a small positive warmup_batches value to avoid measuring one-time startup effects such as kernel warmup or initial memory allocation.