SGD.zero_grad overwrites gradients for non‑trainable parameters

[github-actions]

File: leanpass/optim.py

The zero_grad method unconditionally creates a NumPy array for param.grad even when the parameter does not require gradients (requires_grad=False). This breaks the convention that non‑trainable tensors have grad=None and can cause downstream code to treat them as trainable. The method should only zero gradients for parameters that have requires_grad=True.

def zero_grad(self):
    for param in self.params:
-        param.grad = np.zeros_like(param.data)
+        if param.requires_grad:
+            param.grad = np.zeros_like(param.data)

Filed automatically by ai-issue-scan.