This directory contains ASV-based benchmarks for PhysicsNeMo. Benchmarks are discovered from the benchmarks/ tree and configured via asv.conf.json in the repository root.

Resources:

• ASV documentation

Run all benchmarks from the repo root:

./benchmarks/run_benchmarks.sh

Note: the first run may take a while because ASV builds its virtual environment.

Run a subset by name or regex:

./benchmarks/run_benchmarks.sh -b knn

Publish results to the local HTML report:

asv publish

Preview the report in a local web server:

asv preview

The generated site is written to .asv/html/ (open index.html if you prefer).

1. Add a new file under benchmarks/ following the package structure (for example, benchmarks/physicsnemo/nn/neighbors/my_benchmark.py).
2. Define a benchmark class and at least one time_* method. See documentation for available benchmark types.
3. Use setup() to create inputs and keep benchmarks deterministic. See documentation for available benchmark attributes.

Example:

import torch


class MyOpBenchmark:
    params = [1024, 4096]
    param_names = ["n"]

    def setup(self, n: int) -> None:
        self.x = torch.randn(n, n, device="cuda")

    def time_my_op(self, n: int) -> None:
        _ = self.x @ self.x
        torch.cuda.synchronize()