Command line

All options are in the pytest-torch group of pytest --help. For markers, see Markers.

Options

Option	Purpose
--torch-dtype	Dtype(s) for torch_dtype (repeatable): float32, bfloat16, float16, float64. Default: float32
--torch-device	Device(s) for torch_device (repeatable), e.g. cpu, cuda, xpu, mps, cuda:1. Default: all available. The contract's device_parity then compares only among these devices
--torch-seed	Base seed used by torch_seed (default 0)
--torch-seed-mode={fixed,per-test}	per-test derives a stable seed from the test id
--torch-autoseed	Seed before every test, even those that don't request torch_seed
--torch-deterministic	Use deterministic algorithms inside torch_seed
--torch-update-golden	Rewrite golden files instead of comparing
--torch-update-perf	Record performance baselines instead of comparing
--torch-perf-strict	Fail (instead of skip) when a performance baseline is missing
--torch-perf-file	Baseline file (default <rootdir>/.torch-perf-baselines.json)
--torch-benchmark	Run benchmark suites (deselected otherwise) and print the benchmark report
--torch-benchmark-json=PATH	Also write the benchmark results as JSON
--torch-xdist-gpus	Under xdist: CUDA or XPU GPUs to spread workers over: auto, none, or e.g. 0,1
--torch-num-threads	CPU threads per test process: an integer, none, or auto (default): under pytest-xdist, the cores divided between the workers unless OMP_NUM_THREADS or MKL_NUM_THREADS is set; otherwise torch's default
--torch-gpu-memory-fraction	Per-process CUDA memory cap under xdist: none, auto or a float
--torch-update-metrics	Record metric baselines (metric_baseline) instead of comparing
--torch-metrics-strict	Fail (instead of skip) when a metric baseline is missing
--torch-metrics-file	Metric baseline file (default <rootdir>/.torch-metric-baselines.json)
--torch-model=SPEC	Test a model with no test file (repeatable): hf:NAME[@AutoClass], hf-tiny:NAME[@AutoClass], timm:NAME, torchvision:NAME, package.module:Factory, path/to/file.py:Factory
--torch-model-checks	Checks for --torch-model: contract, invariance, checkpoint, gradients (default), export, compile
--torch-model-inputs	Inputs for --torch-model when they can't be inferred, e.g. 4x3x32x32 or input_ids=long:4x16:1000,attention_mask=mask:4x16
--torch-model-pretrained	Load pretrained weights for --torch-model (default: random init)
--torch-repro-dir=DIR	Write a reproduction bundle for every failing test
--torch-repro=BUNDLE	Replay a bundle: only its test, same seed and inputs
--torch-repro-max-weights-mb	Store weights passed to record_repro up to this size (default 0)
--torch-failure-artifacts=DIR	Save a heatmap PNG for every tensor mismatch (needs matplotlib)
--torch-github-annotations	Print GitHub Actions ::error annotations for failing tests