Paligemma causal attention mask (#30967)

* PaliGemma working causal attention

* Formatting

* Style

* Docstrings + remove commented code

* Update docstring for PaliGemma Config

* PaliGemma - add separator ind to model/labels

* Refactor + docstring paligemma processor method

* Style

* return token type ids when tokenizing labels

* use token type ids when building causal mask

* add token type ids to tester

* remove separator from config

* fix style

* don't ignore separator

* add processor documentation

* simplify tokenization

* fix causal mask

* style

* fix label propagation, revert suffix naming

* fix style

* fix labels tokenization

* [run-slow]paligemma

* add eos if suffixes are present

* [run-slow]paligemma

* [run-slow]paligemma

* add misssing tokens to fast version

* Apply suggestions from code review

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* fix style

* [run-slow]paligemma

---------

Co-authored-by: Peter Robicheaux <peter@roboflow.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

src/transformers/models/paligemma/modeling_paligemma.py

@@ -283,9 +283,14 @@ class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel):
         self.vocab_size = model_embeds.num_embeddings
         return model_embeds
 
-    def _merge_input_ids_with_image_features(self, image_features, inputs_embeds, input_ids, attention_mask, labels):
+    def _merge_input_ids_with_image_features(
+        self, image_features, inputs_embeds, input_ids, attention_mask, labels, token_type_ids, cache_position
+    ):
         _, _, embed_dim = image_features.shape
         batch_size, sequence_length = input_ids.shape
+        dtype, device = inputs_embeds.dtype, inputs_embeds.device
+        min_dtype = torch.finfo(dtype).min
+
         scaled_image_features = image_features / (self.config.hidden_size**0.5)
         final_embedding = torch.zeros(
             batch_size, sequence_length, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
@@ -306,24 +311,43 @@ class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel):
             image_mask.unsqueeze(-1).expand_as(final_embedding), scaled_image_features
         )
         final_embedding = torch.where(pad_mask_expanded, torch.zeros_like(final_embedding), final_embedding)
+        if attention_mask is not None:
+            position_ids = (attention_mask.cumsum(-1)).masked_fill_((attention_mask == 0), 1)
+        else:
+            position_ids = None
 
-        final_attention_mask_4d = attention_mask.unsqueeze(1).unsqueeze(2) * attention_mask.unsqueeze(1).unsqueeze(-1)
-        final_attention_mask_4d = final_attention_mask_4d.float().expand(
-            -1, self.config.text_config.num_key_value_heads, -1, -1
-        )
+        if token_type_ids is not None and labels is not None:
+            # we are training thus we need to create a full mask on the image + prefix but causal on suffix
+            target_length = cache_position[-1] + 1
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(inputs_embeds.shape[0], 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                # unmask the prefill
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    token_type_ids[:, None, None, :] == 0, 0
+                )
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
 
-        # position_ids = torch.arange(0, sequence_length, device=input_ids.device).expand(batch_size, -1)
-        # position_ids = torch.where(input_ids == self.pad_token_id, torch.ones_like(position_ids), position_ids)
-        position_ids = (attention_mask.cumsum(-1)).masked_fill_((attention_mask == 0), 1)
-
-        if labels is not None:
             final_labels = torch.full(
                 (batch_size, sequence_length), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
             )
             final_labels = torch.where(input_ids != self.pad_token_id, labels, final_labels)
         else:
+            causal_mask = attention_mask.unsqueeze(1).unsqueeze(2) * attention_mask.unsqueeze(1).unsqueeze(-1)
+            causal_mask = causal_mask.to(dtype).expand(-1, self.config.text_config.num_key_value_heads, -1, -1)
             final_labels = None
-        return final_embedding, final_attention_mask_4d, final_labels, position_ids
+        return final_embedding, causal_mask, final_labels, position_ids
 
     @add_start_docstrings_to_model_forward(PALIGEMMA_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=PaliGemmaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
@@ -334,6 +358,7 @@ class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel):
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
         past_key_values: Optional[Union[List[torch.FloatTensor], Cache]] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
         cache_position: Optional[torch.LongTensor] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
         labels: Optional[torch.LongTensor] = None,
@@ -397,8 +422,10 @@ class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel):
                 selected_image_feature = image_outputs.last_hidden_state
                 image_features = self.multi_modal_projector(selected_image_feature)
 
+                if cache_position is None:
+                    cache_position = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device)
                 inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features(
-                    image_features, inputs_embeds, input_ids, attention_mask, labels
+                    image_features, inputs_embeds, input_ids, attention_mask, labels, token_type_ids, cache_position
                 )
 
             else:
@@ -487,6 +514,7 @@ class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel):
         cache_position=None,
         pixel_values=None,
         attention_mask=None,
+        token_type_ids=None,
         **kwargs,
     ):
         past_length = 0
@@ -545,6 +573,7 @@ class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel):
                 "use_cache": kwargs.get("use_cache"),
                 "attention_mask": attention_mask,
                 "pixel_values": pixel_values,
+                "token_type_ids": token_type_ids,
             }
         )
         return model_inputs

tests/models/paligemma/test_modeling_paligemma.py

@@ -163,6 +163,8 @@ class PaliGemmaVisionText2TextModelTester:
             "pixel_values": pixel_values,
             "input_ids": input_ids,
             "attention_mask": attention_mask,
+            "labels": input_ids,
+            "token_type_ids": torch.zeros_like(input_ids),
         }
         return config, inputs_dict