[M2M100, XGLM] fix create_position_ids_from_inputs_embeds (#15751)

src/transformers/models/m2m_100/modeling_m2m_100.py

@@ -167,7 +167,7 @@ class M2M100SinusoidalPositionalEmbedding(nn.Module):
             )
         else:
             bsz, seq_len = inputs_embeds.size()[:-1]
-            position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
+            position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds, past_key_values_length)
 
         # expand embeddings if needed
         max_pos = self.padding_idx + 1 + seq_len + past_key_values_length
@@ -176,7 +176,7 @@ class M2M100SinusoidalPositionalEmbedding(nn.Module):
 
         return self.weights.index_select(0, position_ids.view(-1)).view(bsz, seq_len, -1).detach()
 
-    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds, past_key_values_length):
         """
         We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
 
@@ -191,7 +191,7 @@ class M2M100SinusoidalPositionalEmbedding(nn.Module):
         position_ids = torch.arange(
             self.padding_idx + 1, sequence_length + self.padding_idx + 1, dtype=torch.long, device=inputs_embeds.device
         )
-        return position_ids.unsqueeze(0).expand(input_shape).contiguous()
+        return position_ids.unsqueeze(0).expand(input_shape).contiguous() + past_key_values_length
 
 
 # Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->M2M100

src/transformers/models/xglm/modeling_xglm.py

@@ -211,7 +211,7 @@ class XGLMSinusoidalPositionalEmbedding(nn.Module):
             )
         else:
             bsz, seq_len = inputs_embeds.size()[:-1]
-            position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
+            position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds, past_key_values_length)
 
         # expand embeddings if needed
         max_pos = self.padding_idx + 1 + seq_len + past_key_values_length
@@ -220,7 +220,7 @@ class XGLMSinusoidalPositionalEmbedding(nn.Module):
 
         return self.weights.index_select(0, position_ids.view(-1)).view(bsz, seq_len, -1).detach()
 
-    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds, past_key_values_length):
         """
         We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
 
@@ -235,7 +235,7 @@ class XGLMSinusoidalPositionalEmbedding(nn.Module):
         position_ids = torch.arange(
             self.padding_idx + 1, sequence_length + self.padding_idx + 1, dtype=torch.long, device=inputs_embeds.device
         )
-        return position_ids.unsqueeze(0).expand(input_shape).contiguous()
+        return position_ids.unsqueeze(0).expand(input_shape).contiguous() + past_key_values_length
 
 
 # Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->XGLM