pytorch - 💡(How to fix) Fix ConstantLR accepts factor=0 and crashes with ZeroDivisionError after total_iters steps [1 pull requests]

Bug Description

ConstantLR accepts factor=0 without raising an error during initialization. However, after total_iters steps, the get_lr() method computes 1.0 / self.factor, which causes a ZeroDivisionError.

The validation at line 813 of torch/optim/lr_scheduler.py uses factor < 0 (strict inequality), but should use factor <= 0 to also reject zero.

Reproducer

import torch

optimizer = torch.optim.SGD([torch.randn(1)], lr=0.1)
scheduler = torch.optim.lr_scheduler.ConstantLR(optimizer, factor=0.0, total_iters=2)

scheduler.step()  # OK
scheduler.step()  # ZeroDivisionError: float division by zero

Root Cause

In torch/optim/lr_scheduler.py:

• Line 813: if factor > 1.0 or factor < 0: should be if factor > 1.0 or factor <= 0:
• Line 854: get_lr() computes group["lr"] * (1.0 / self.factor) which divides by zero when factor=0

Inconsistency

LinearLR in the same file correctly validates with start_factor <= 0 (line 921), rejecting zero. ConstantLR should follow the same pattern.

Expected Behavior

ConstantLR(optimizer, factor=0.0) should raise a ValueError at construction time, not silently accept the value and crash later.

Environment

• PyTorch version: main branch
• Python version: 3.x
• OS: any

cc @vincentqb @jbschlosser @albanD @janeyx99 @crcrpar @malfet @pytorch/optim