Benchmarks

Automated performance benchmarks for every TorchEBM module. Results are generated by the benchmark suite and rendered here automatically at documentation build time.

Quick Start

First Run (Baseline)After OptimizationQuick TestCI Mode

bash benchmarks/run.sh

Captures baseline performance numbers.

bash benchmarks/run.sh

Automatically compares against baseline and generates report.

bash benchmarks/run.sh --quick

Small-scale only (~1 min).

bash benchmarks/run.sh --ci

Exits non-zero on regression > 5%.

Coverage

• Losses

  Exact SM, Approx SM, Sliced SM, Contrastive Divergence, Equilibrium Matching

• Samplers

  Langevin Dynamics, HMC, Flow ODE

• Integrators

  Leapfrog, Euler-Maruyama, RK4, DOPRI5 (adaptive)

• Models

  Transformer forward, Transformer forward+backward

• Interpolants

  Linear, Cosine, Variance-Preserving

Scale Configurations

Scale	Batch Size	Dimensions	Steps
small	64	8	50
medium	256	32	100
large	1024	128	200

Measurement Methodology

Aspect	GPU	CPU
Timer	torch.cuda.Event	time.perf_counter
Warm-up	10 iterations	3 iterations
Measured	50 iterations	20 iterations
Memory	cuda.max_memory_allocated	N/A
Statistics	median, mean, std, IQR, min, max	same
Comparison	geometric mean speedup	same