Forward differentiation nan gradient but reverse mode is fine

[unalmis]

Under what circumstances is it possible to have a forward mode nan gradient but not a reverse mode nan gradient? The context for this issue is here.

[bmaxdk]

If the very first gradient evaluation in forward-mode hits a near-zero denominator, you get a NaN right away. Can you use a small safe divisor offset (ϵ) or clamping? Since extremely small value can cause this issue