Status: Supported optional numerical backend Scope: Backend compatibility; no dedicated CUDA engine or speedup guarantee
Install the optional Torch dependency with:
pip install -e ".[compute-torch]"Request the backend through the public mathematics backend interface and inspect the actual selection:
from tnfr.mathematics.backend import get_backend
backend = get_backend("torch")
print(backend.name, backend.get_device_name())If Torch is unavailable, get_backend("torch") can return the NumPy fallback.
Check backend.name == "torch" when Torch execution is required. A Torch backend
can itself use CPU or CUDA according to the installation, device availability
and TNFR_CUDA_ENABLED; requesting Torch does not guarantee GPU execution.
backend.get_backend_info() exposes additional execution metadata.
Backend agreement is covered by
tests/mathematics/test_backends.py.
Those backend-specific checks skip when the requested dependency is unavailable;
a passing NumPy-only run does not establish Torch agreement.
TNFR does not currently ship the historical tnfr.engines.computation.gpu_engine.TNFRGPUEngine class. The pytorch_cuda_demo.py filename is retained as a compatibility and provenance check: it verifies Torch operations against NumPy and reports the selected device. The canonical graph-pressure adapter currently reports a CPU realization. The example does not assert a CUDA speedup.
A future GPU acceleration claim must include:
- an implementation reachable through a supported public API;
- numerical agreement with the canonical CPU computation;
- reproducible inputs, seeds, hardware and software versions;
- warm-up, transfer-time and memory accounting;
- recorded benchmark results with uncertainty and crossover sizes.
Until those conditions are met, compute-torch means optional Torch numerical
compatibility, not guaranteed CUDA acceleration.