Fix PyTorch pad NumPy mode aliases

src/pyrecest/backend_support/_pytorch_pad_mode_contract.py

@@ -0,0 +1,47 @@
+"""PyTorch ``pad`` NumPy mode-name compatibility hook."""
+
+from __future__ import annotations
+
+_PAD_MODE_ALIASES = {
+    "edge": "replicate",
+    "wrap": "circular",
+}
+
+
+def _torch_pad_mode(mode):
+    """Return the PyTorch padding mode corresponding to a NumPy mode name."""
+
+    return _PAD_MODE_ALIASES.get(mode, mode)
+
+
+def patch_pytorch_pad_mode_contract() -> None:
+    """Patch raw/public PyTorch ``pad`` to accept NumPy mode names."""
+
+    try:
+        import pyrecest._backend.pytorch as pytorch_backend  # pylint: disable=import-outside-toplevel
+        import pyrecest.backend as backend  # pylint: disable=import-outside-toplevel
+    except ModuleNotFoundError:  # pragma: no cover - PyTorch backend may be unavailable
+        return
+
+    original_pad = getattr(pytorch_backend, "pad", None)
+    if original_pad is None:
+        return
+    if getattr(original_pad, "_pyrecest_numpy_pad_mode_contract", False):
+        if getattr(backend, "__backend_name__", None) == "pytorch":
+            backend.pad = original_pad
+        return
+
+    def pad(a, pad_width, mode="constant", constant_values=0.0):
+        return original_pad(
+            a,
+            pad_width,
+            mode=_torch_pad_mode(mode),
+            constant_values=constant_values,
+        )
+
+    pad.__name__ = getattr(original_pad, "__name__", "pad")
+    pad.__doc__ = getattr(original_pad, "__doc__", None)
+    pad._pyrecest_numpy_pad_mode_contract = True
+    pytorch_backend.pad = pad
+    if getattr(backend, "__backend_name__", None) == "pytorch":
+        backend.pad = pad

src/pyrecest/stability.py

@@ -9,8 +9,12 @@
 from pyrecest.backend_support._pytorch_allclose_device_contract import (
     patch_pytorch_allclose_device_contract as _patch_pytorch_allclose_device_contract,
 )
+from pyrecest.backend_support._pytorch_pad_mode_contract import (
+    patch_pytorch_pad_mode_contract as _patch_pytorch_pad_mode_contract,
+)
 
 _patch_pytorch_allclose_device_contract()
+_patch_pytorch_pad_mode_contract()
 
 P = ParamSpec("P")
 R = TypeVar("R")

tests/backend/test_pytorch_pad_mode_contract.py

@@ -0,0 +1,40 @@
+import pytest
+from tests.support.backend_runner import run_backend_code
+
+
+torch = pytest.importorskip("torch")
+
+import pyrecest.backend_tools  # noqa: E402,F401
+import pyrecest._backend.pytorch as pytorch_backend  # noqa: E402
+
+
+def _to_list(value):
+    return value.detach().cpu().tolist()
+
+
+def test_raw_pytorch_pad_accepts_numpy_edge_and_wrap_mode_names():
+    values = torch.tensor([[1.0, 2.0, 3.0]], dtype=torch.float64)
+
+    edge = pytorch_backend.pad(values, ((0, 0), (2, 1)), mode="edge")
+    wrap = pytorch_backend.pad(values, ((0, 0), (1, 2)), mode="wrap")
+
+    assert _to_list(edge) == [[1.0, 1.0, 1.0, 2.0, 3.0, 3.0]]
+    assert _to_list(wrap) == [[3.0, 1.0, 2.0, 3.0, 1.0, 2.0]]
+
+
+def test_public_pytorch_pad_accepts_numpy_edge_and_wrap_mode_names():
+    result = run_backend_code(
+        "pytorch",
+        r'''
+import pyrecest.backend as backend
+
+values = backend.asarray([[1.0, 2.0, 3.0]])
+edge = backend.pad(values, ((0, 0), (2, 1)), mode="edge")
+wrap = backend.pad(values, ((0, 0), (1, 2)), mode="wrap")
+
+assert backend.to_numpy(edge).tolist() == [[1.0, 1.0, 1.0, 2.0, 3.0, 3.0]]
+assert backend.to_numpy(wrap).tolist() == [[3.0, 1.0, 2.0, 3.0, 1.0, 2.0]]
+''',
+    )
+
+    assert result.returncode == 0, result.stderr