Fix fp16 masking in PoolerEndLogits

Necessary to run xlnet (at least in squad) with `--fp16 --fp16_opt_level="O2"`, otherwise loss is immediately `NaN` and fine-tuning cannot proceed.

pytorch_transformers/modeling_utils.py

@@ -478,6 +478,9 @@ class PoolerEndLogits(nn.Module):
         x = self.dense_1(x).squeeze(-1)
 
         if p_mask is not None:
+            if next(self.parameters()).dtype == torch.float16:
+                x = x * (1 - p_mask) - 65500 * p_mask
+            else:
                 x = x * (1 - p_mask) - 1e30 * p_mask
 
         return x