Adding Multimodal Capabilities

This guide will walk you through the steps needed to extend the capabilities of an Aphrodite model so that it accepts multimodal inputs.

TIP

See also: Adding a New Model.

Step 1: Update the base Aphrodite model

We assume that you have already created a new model by following the steps in the Adding a New Model guide. If not, please do so before proceeding.

1. Implement the aphrodite.modeling.models.interfaces.SupportsMultiModal interface:

from aphrodite.modeling.models.interfaces import SupportsMultiModal  

class YourModelForImage2Seq(nn.Module):  
class YourModelForImage2Seq(nn.Module, SupportsMultiModal):  

INFO

The model class does not have to be named *ForCausalLM. Check out the HuggingFace Transformers documentation for some examples.

2. If you haven't done so already, reserve a keyword parameter in forward() for each input tensor that corresponds to a multi-modal input, as shown in the following example:

  def forward(
      self,
      input_ids: torch.Tensor,
      positions: torch.Tensor,
      kv_caches: List[torch.Tensor],
      attn_metadata: AttentionMetadata,
      pixel_values: torch.Tensor, 
  ) -> SamplerOutput:

Step 2: Register input mappers

For each modality type that the model accepts as input, decorate the model class with aphrodite.multimodal.MultiModalRegistry.register_input_mapper. This decorator accepts a function that maps multi-modal inputs to the keyword arguments you have previously defined in forward().

from aphrodite.modeling.models.interfaces import SupportsMultiModal
from aphrodite.multimodal import MULTIMODAL_REGISTRY

@MULTIMODAL_REGISTRY.register_image_input_mapper()  
class YourModelForImage2Seq(nn.Module, SupportsMultiModal):

A default mapper is available for each modality in the core Aphrodite library. This input mapper will be used if you do not provide your own function.

TIP

See also: Input Processing Pipeline.

Step 3: Register maximum number of multi-modal tokens

For each modality type that the model accepts as input, calculate the maximum possible number of tokens and register it via aphrodite.inputs.InputRegistry.register_max_multimodal_tokens.

from aphrodite.inputs import INPUT_REGISTRY
from aphrodite.modeling.models.interfaces import SupportsMultiModal
from aphrodite.multimodal import MULTIMODAL_REGISTRY

@MULTIMODAL_REGISTRY.register_image_input_mapper()
@MULTIMODAL_REGISTRY.register_max_image_tokens(<your_calculation>)
@INPUT_REGISTRY.register_dummy_data(<your_dummy_data_factory>)  
class YourModelForImage2Seq(nn.Module, SupportsMultiModal):

Here are some examples:

• Static feature size: LLaVA-1.5 Model
• Dynamic feature size: LLaVA-NeXT Model

TIP

See also: Input Processing Pipeline.

Step 4: (Optional) Register dummy data

During startup, dummy data is passed to the Aphrodite model to allocate memory. This only consists of text input by default, which may not be applicable to multi-modal models. In such cases, you can define your own dummy data by registering a factory method via aphrodite.inputs.InputRegistry.register_dummy_data.

from aphrodite.inputs import INPUT_REGISTRY
from aphrodite.modeling.models.interfaces import SupportsMultiModal
from aphrodite.multimodal import MULTIMODAL_REGISTRY

@MULTIMODAL_REGISTRY.register_image_input_mapper()
@MULTIMODAL_REGISTRY.register_max_image_tokens(<your_calculation>)
@INPUT_REGISTRY.register_dummy_data(<your_dummy_data_factory>)  
class YourModelForImage2Seq(nn.Module, SupportsMultiModal):

INFO

The dummy data should have the maximum possible number of multi-modal tokens, as described in the previous step.

Here are some examples:

• Static feature size: LLaVA-1.5 Model
• Dynamic feature size: LLaVA-NeXT Model

TIP

See also: Input Processing Pipeline.

Step 5: (Optional) Register input processor

Sometimes, there's a need to process inputs at the aphrodite.AphroditeEngine level before they are passed to the model executor. This is often due to the fact that unlike implementations in HuggingFace Transformers, the reshaping and/or expansion of multi-modal embeddings needs to take place outside model's forward() method. You can register input processors via aphrodite.inputs.InputRegister.register_input_processor.

from aphrodite.inputs import INPUT_REGISTRY
from aphrodite.modeling.models.interfaces import SupportsMultiModal
from aphrodite.multimodal import MULTIMODAL_REGISTRY

@MULTIMODAL_REGISTRY.register_image_input_mapper()
@MULTIMODAL_REGISTRY.register_max_image_tokens(<your_calculation>)
@INPUT_REGISTRY.register_dummy_data(<your_dummy_data_factory>)
@INPUT_REGISTRY.register_input_processor(<your_input_processor>)  
class YourModelForImage2Seq(nn.Module, SupportsMultiModal):

A common use case of input processors is inserting placeholder tokens to leverage the Aphrodite framework for attention mask generation. Here are some examples:

• Insert static number of image tokens: LLaVA-1.5 Model
• Insert dynamic number of image tokens: LLaVA-NeXT Model

TIP

See also: Input Processing Pipeline.