INPUTS:
  input_ids: ['text_batch_size', 'sequence_length'] tensor(int64)
  pixel_values: ['image_batch_size', 'num_channels', 'height', 'width'] tensor(float)
  attention_mask: ['text_batch_size', 'sequence_length'] tensor(int64)
OUTPUTS:
  logits_per_image: ['image_batch_size', 'text_batch_size'] tensor(float)
  logits_per_text: ['text_batch_size', 'image_batch_size'] tensor(float)
  text_embeds: ['text_batch_size', 512] tensor(float)
  image_embeds: ['image_batch_size', 512] tensor(float)