Fix copies for DBRX - neuron fix (#30610)

src/transformers/models/dbrx/modeling_dbrx.py

@@ -1256,8 +1256,11 @@ class DbrxModel(DbrxPreTrainedModel):
             causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
             if attention_mask.dim() == 2:
                 mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
-                causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
             elif attention_mask.dim() == 4:
                 # backwards compatibility: we allow passing a 4D attention mask shorter than the input length with
                 # cache. In that case, the 4D attention mask attends to the newest tokens only.