BLIP: this is correct now (#35081)

this is correct now

src/transformers/models/blip_2/modeling_blip_2.py

@@ -2311,7 +2311,7 @@ class Blip2ForConditionalGeneration(Blip2PreTrainedModel, GenerationMixin):
         if input_ids is None:
             start_tokens = [self.config.text_config.bos_token_id]
             if getattr(self.config, "image_token_index", None) is not None:
-                start_tokens += [self.config.image_token_index] * self.config.num_query_tokens
+                start_tokens = [self.config.image_token_index] * self.config.num_query_tokens + start_tokens
             input_ids = torch.tensor([start_tokens], dtype=torch.long, device=image_embeds.device)
             input_ids = input_ids.repeat(batch_size, 1)
 

src/transformers/models/instructblip/modeling_instructblip.py

@@ -1593,7 +1593,7 @@ class InstructBlipForConditionalGeneration(InstructBlipPreTrainedModel, Generati
         if input_ids is None:
             start_tokens = [self.config.text_config.bos_token_id]
             if getattr(self.config, "image_token_index", None) is not None:
-                start_tokens += [self.config.image_token_index] * self.config.num_query_tokens
+                start_tokens = [self.config.image_token_index] * self.config.num_query_tokens + start_tokens
             input_ids = torch.tensor([start_tokens], dtype=torch.long, device=image_embeds.device)
             input_ids = input_ids.repeat(batch_size, 1)
 

src/transformers/models/instructblipvideo/modeling_instructblipvideo.py

@@ -1628,7 +1628,7 @@ class InstructBlipVideoForConditionalGeneration(InstructBlipVideoPreTrainedModel
         if input_ids is None:
             start_tokens = [self.config.text_config.bos_token_id]
             if getattr(self.config, "video_token_index", None) is not None:
-                start_tokens += [self.config.video_token_index] * self.config.num_query_tokens * 4
+                start_tokens = [self.config.video_token_index] * self.config.num_query_tokens * 4 + start_tokens
             input_ids = torch.tensor([start_tokens], dtype=torch.long, device=image_embeds.device)
             input_ids = input_ids.repeat(batch_size, 1)
 

src/transformers/models/instructblipvideo/modular_instructblipvideo.py

@@ -441,7 +441,7 @@ class InstructBlipVideoForConditionalGeneration(InstructBlipForConditionalGenera
         if input_ids is None:
             start_tokens = [self.config.text_config.bos_token_id]
             if getattr(self.config, "video_token_index", None) is not None:
-                start_tokens += [self.config.video_token_index] * self.config.num_query_tokens * 4
+                start_tokens = [self.config.video_token_index] * self.config.num_query_tokens * 4 + start_tokens
             input_ids = torch.tensor([start_tokens], dtype=torch.long, device=image_embeds.device)
             input_ids = input_ids.repeat(batch_size, 1)