fix attn mask shape of gpt (#472)

model_zoo/gpt/gpt.py

@@ -292,7 +292,7 @@ class GPT(nn.Module):
         # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
         # Adapted from huggingface
         if attention_mask is not None:
-            batch_size = x.shape[0]
+            batch_size = input_ids.shape[0]
             attention_mask = attention_mask.view(batch_size, -1)
             attention_mask = col_nn.partition_batch(attention_mask)
             attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)