Update: ignore padding support for TransfoXL training when n_clusters==0 (#22457)

* Update: ignore padding support for TransfoXL training when n_clusters==0

* Update: transformer XL always pad

* Update: drop doc

src/transformers/models/transfo_xl/modeling_transfo_xl_utilities.py

@@ -86,7 +86,7 @@ class ProjectedAdaptiveLogSoftmax(nn.Module):
         """
         Params:
             hidden :: [len*bsz x d_proj]
-            labels :: [len*bsz
+            labels :: [len*bsz]
 
         Return:
             if labels is None: out :: [len*bsz x n_tokens] log probabilities of tokens over the vocabulary else: out ::
@@ -109,7 +109,11 @@ class ProjectedAdaptiveLogSoftmax(nn.Module):
         if self.n_clusters == 0:
             logit = self._compute_logit(hidden, self.out_layers[0].weight, self.out_layers[0].bias, self.out_projs[0])
             if labels is not None:
-                out = -nn.functional.log_softmax(logit, dim=-1).gather(1, labels.unsqueeze(1)).squeeze(1)
+                mask = labels != -100
+                out = torch.zeros_like(labels, dtype=hidden.dtype, device=hidden.device)
+                out[mask] = (
+                    -nn.functional.log_softmax(logit, dim=-1)[mask].gather(1, labels[mask].unsqueeze(1)).squeeze(1)
+                )
             else:
                 out = nn.functional.log_softmax(logit, dim=-1)
         else: