[debug] DebugUnderflowOverflow doesn't work with DP (#12816)

2021-07-21 09:36:02 -07:00
parent ac3cb660ca
commit cf0755aa6e
3 changed files with 15 additions and 4 deletions
--- a/docs/source/debugging.rst
+++ b/docs/source/debugging.rst
@@ -24,7 +24,11 @@ Underflow and Overflow Detection

 .. note::

-   This feature can be used with any ``nn.Module``-based model
+   For multi-GPU training it requires DDP (``torch.distributed.launch``).
+
+.. note::
+
+   This feature can be used with any ``nn.Module``-based model.

 If you start getting ``loss=NaN`` or the model inhibits some other abnormal behavior due to ``inf`` or ``nan`` in
 activations or weights one needs to discover where the first underflow or overflow happens and what led to it. Luckily