Skip to content

Command-line tools

Command-line operations and the reproducible benchmark entry points.

TorchWave registers five standalone console executables:

torchwave-bench

Benchmark model throughput, batch sizes, autograd backward passes, and cross-framework GPU speedups: 

# Benchmark standard models across batch sizes
torchwave-bench --models popular --batch-sizes 1,10,100,1000

# Benchmark across backends (Torch eager, Torch compiled, JAX JIT, or all)
torchwave-bench --models IMRPhenomD --backend all --batch-sizes 1,10,100

# Benchmark compiled PyTorch with dynamic shapes (can reduce recompilation across batch sizes)
torchwave-bench --models TaylorF2,IMRPhenomD --compile --dynamic --batch-sizes 1,10,50,100

# Measure forward + backward autograd gradient time
torchwave-bench --models IMRPhenomD --backend torch --batch-sizes 100 --backward

# Compare against LALSimulation baseline and export Markdown table
torchwave-bench --models TaylorF2,IMRPhenomD,IMRPhenomXAS --compare-lal --markdown-out benchmark.md

torchwave-profile

Profile operator kernels, execution bottlenecks, and memory allocation: 

# Run comprehensive profiling (cProfile, torch profiler, memory)
torchwave-profile --model IMRPhenomD --mode all --top-k 15

# Export Chrome trace JSON for visualization in chrome://tracing
torchwave-profile --model TaylorF2 --trace-out trace_f2.json

torchwave-data

Manage optional external ROM datasets outside git repositories and wheels: 

# List available external datasets
torchwave-data list

# Inspect dataset metadata
torchwave-data info seobnrv4-rom-v3

# Fetch dataset to user cache or custom directory
torchwave-data fetch seobnrv4-rom-v3

# Verify pinned SHA-256 integrity
torchwave-data verify seobnrv4-rom-v3

torchwave-generate-qnm

Generate Kerr quasinormal mode tables directly from the Leaver continued-fraction eigenproblem: 

torchwave-generate-qnm --output-dir ./qnm_data --modes 2,2,0

torchwave-generate-london-qnm

Recompute the London \(N=61\) quasinormal mode polynomial fits: 

torchwave-generate-london-qnm --output-dir ./london_qnm

Controlled Benchmarking Protocol

To reproduce published results, use scripts/compare_packages.py with the dated campaign's commands and environment. Regenerate its figures with the maintained plot command. For a new experiment, follow the benchmark methodology.

Running a Multi-Backend Benchmark Campaign

The runner guide distinguishes the maintained package comparison, local CLI experiments and legacy campaigns. The older run_benchmark_campaign.py can use JAXWave or self-consistency as its reference; its gate_passed flag alone does not establish LAL agreement.

Generating Figures

Use the published campaign's plot instructions for maintained CPU/GPU figures. plot_benchmark_results.py reads the separate legacy receipt format; see the legacy qualification rules.