hoptorch

hoptorch is a small compatibility package for PyTorch higher-order operators. Its first helper is a safe wrapper around torch._higher_order_ops.scan that checks whether scan backward is healthy on the requested device and lazily installs version-specific compatibility patches for affected PyTorch internals. For PyTorch 2.7, where scan backward is not implemented, hoptorch backports a small eager scan implementation with ordinary autograd support.

Install

pip install hoptorch

Runtime dependency:

torch>=2.7
pyvers>=0.2.2

Usage

import torch
from hoptorch import scan
from hoptorch.scan import ensure_scan_backward, scan_unavailable_reason

if ensure_scan_backward("cpu"):
    xs = torch.arange(4.0)

    def step(carry, x):
        next_carry = carry + x
        return next_carry, next_carry.clone()

    carry, ys = scan(step, torch.zeros(()), xs)
else:
    print(scan_unavailable_reason("cpu"))

For compiled code, warm the health check before entering torch.compile:

from hoptorch.scan import ensure_scan_backward

ensure_scan_backward("cpu")
compiled_fn = torch.compile(fn)

If a wrapper call happens during Dynamo tracing before the health result is cached, hoptorch fails closed instead of tracing the probe.

Public API

from hoptorch import scan
from hoptorch.scan import (
    ensure_scan_backward,
    has_scan,
    patch_scan_backward,
    scan_unavailable_reason,
)

• scan(fn, init, xs, *, dim=0, **kwargs): calls PyTorch scan only when scan backward is known to be healthy for the inferred device.
• has_scan(): reports whether torch._higher_order_ops.scan.scan exists.
• ensure_scan_backward(device=None): runs or reads the cached health check, patching lazily if needed, and returns True only after a passing probe.
• scan_unavailable_reason(device=None): returns None when scan backward is usable, otherwise a stable human-readable reason.
• patch_scan_backward(): attempts to install the compatibility patch and is intended mainly for diagnostics.

Tests

python -m unittest discover -s tests