RLeaky and RSynaptic fail under float64 with dtype mismatch during forward pass

[GraceDyer]

Summary

I encountered a reproducible issue when running recurrent snnTorch neuron modules in float64. Specifically, RLeaky and RSynaptic fail during forward execution with a dtype mismatch error.

This appears to affect high-precision workflows where users may reasonably expect model.double() or float64 inputs/states to work consistently.

Minimal reproduction

import torch
import snntorch as snn

x = torch.randn(2, 4, dtype=torch.float64)

rleaky = snn.RLeaky(beta=0.9, linear_features=4).double()
rleaky(x)

A similar failure is also reproducible with RSynaptic under the same float64 setup.

Expected behavior

Recurrent neuron modules should either:

• run correctly in float64 when the module and inputs are converted to float64, or
• fail with a clearly documented limitation if float64 is not intended to be supported.

Actual behavior

The forward pass fails with:

RuntimeError: mat1 and mat2 must have the same dtype, but got Float and Double

In my reproduction/validation runs, this behavior was stable across repeated reruns and affected at least:

• RLeaky
• RSynaptic

Environment

• torch: 2.8.0
• snntorch: 0.9.4
• python: 3.10

Additional notes

This issue matters for high-precision or numerically sensitive workflows, where converting the model and inputs to float64 is a reasonable expectation.

If float64 support is intentionally limited for these recurrent modules, an explicit documentation note or earlier validation error would still be very helpful.