Skip to content

Diffusion Transformer for Cosmos

The RBLNCosmosTransformer3DModel is the RBLN-optimized version of the core transformer block used in the Cosmos World Foundation Models.

Supported NPU

Diffusers pipelines run on ATOM™ only. Rebel100™ does not support them yet.

This model is the diffusion transformer (DiT) at the core of every Cosmos pipeline. At each denoising step, it processes the latent video representation together with the text embeddings from the pipeline's text encoder and the timestep. When conditioning frames are given (Image2World and Video2World), a condition mask marks them in the input.

Usage within Pipelines

Typically, you do not interact with RBLNCosmosTransformer3DModel directly. Instead, it is automatically loaded and managed as part of an RBLN Cosmos pipeline, such as RBLNCosmosTextToWorldPipeline, RBLNCosmosVideoToWorldPipeline, and RBLNCosmos2_5_PredictBasePipeline.

When configuring an RBLN Cosmos pipeline, you can pass specific settings for the transformer via the transformer argument in the pipeline's configuration object:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
from optimum.rbln import (
    RBLNCosmosTextToWorldPipeline,
    RBLNCosmosTextToWorldPipelineConfig,
)

# Example: Configure pipeline with specific batch size for the transformer
config = RBLNCosmosTextToWorldPipelineConfig(
    batch_size=1,
    height=704,
    width=1280,
    transformer={
        "num_devices": 4,
        "device": [0, 1, 2, 3],  # Allocate appropriate number of NPUs
    },
)

# ... load pipeline using this config ...

API Reference

CosmosPerFrameAdaLNMixin

CosmosTransformerBlock.forward with the adaLN modulation kept per-frame.

Shared by the transformer and the ControlNet: a per-frame timestep gives only T distinct modulation values, so computing them on [B, T, C] and broadcasting over the [B, T, HW, C] view of the hidden states keeps the H*W-fold expansion out of the graph.

CosmosTransferTransformerWrapper

Bases: CosmosTransformer3DModelWrapper

Transfer2.5 graph: the base blocks plus ControlNet residual injection and image context.

A separate wrapper (instead of optional inputs on the base one) keeps the Predict graph and its input order byte-identical, and removes any ambiguity in positional input binding.

RBLNCosmosTransformer3DModel

Bases: RBLNModel

RBLN implementation of CosmosTransformer3DModel for diffusion models like Cosmos.

The CosmosTransformer3DModel takes text and/or image embeddings from encoders (like CLIP) and maps them to a shared latent space that guides the diffusion process to generate the desired image.

This class inherits from [RBLNModel]. Check the superclass documentation for the generic methods the library implements for all its models.

forward(hidden_states, timestep, encoder_hidden_states, attention_mask=None, fps=None, condition_mask=None, padding_mask=None, block_controlnet_hidden_states=None, return_dict=True)

Forward pass for the RBLN-optimized CosmosTransformer3DModel.

Parameters:

Name Type Description Default
hidden_states Tensor

The currently predicted image embeddings.

 required
timestep Tensor

Current denoising step.

 required
encoder_hidden_states Tensor

Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.

 required
fps int | None

(int | None): Frames per second for the video being generated.

 None
condition_mask Tensor | None

Tensor of condition mask.

 None
padding_mask Tensor | None

Tensor of padding mask.

 None
return_dict bool

Whether or not to return a [~diffusers.models.modeling_output.Transformer2DModelOutput] instead of a plain tuple.

 True

Returns:

Type Description
Transformer2DModelOutput | tuple

(~diffusers.models.modeling_output.Transformer2DModelOutput | tuple)

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod

The from_pretrained() function is utilized in its standard form as in the HuggingFace transformers library. User can use this function to load a pre-trained model from the HuggingFace library and convert it to a RBLN model to be run on RBLN NPUs.

Parameters:

Name Type Description Default
model_id str | Path

The model id of the pre-trained model to be loaded. It can be downloaded from the HuggingFace model hub or a local path, or a model id of a compiled model using the RBLN Compiler.

 required
export bool | None

A boolean flag to indicate whether the model should be compiled. If None, it will be determined based on the existence of the compiled model files in the model_id.

 None
rbln_config dict | RBLNModelConfig | None

Configuration for RBLN model compilation and runtime. This can be provided as a dictionary or an instance of the model's configuration class (e.g., RBLNLlamaForCausalLMConfig for Llama models). For detailed configuration options, see the specific model's configuration class documentation.

 None
kwargs Any

Additional keyword arguments. Arguments with the prefix rbln_ are passed to rbln_config, while the remaining arguments are passed to the HuggingFace library.

 {}

Returns:

Type Description
RBLNModel

A RBLN model instance ready for inference on RBLN NPU devices.

get_hf_class() classmethod

Lazily load and cache the HuggingFace class this model wraps.

The class is looked up in hf_library_name under the name without the RBLN prefix (e.g. RBLNLlamaForCausalLM -> LlamaForCausalLM).

save_pretrained(save_directory, push_to_hub=False, **kwargs)

Saves a model and its configuration file to a directory, so that it can be re-loaded using the [~optimum.rbln.modeling_base.RBLNBaseModel.from_pretrained] class method.

Parameters:

Name Type Description Default
save_directory str | Path

Directory where to save the model file.

 required
push_to_hub bool

Whether or not to push your model to the HuggingFace model hub after saving it.

 False

from_model(model, config=None, rbln_config=None, model_save_dir=None, subfolder='', **kwargs) classmethod

Converts and compiles a pre-trained HuggingFace library model into a RBLN model. This method performs the actual model conversion and compilation process.

Parameters:

Name Type Description Default
model PreTrainedModel | ModelMixin

The PyTorch model to be compiled. The object must be a HuggingFace transformers PreTrainedModel or a Diffusers component based on diffusers.ModelMixin, such as a VAE, UNet or ControlNet.

 required
config PretrainedConfig | None

The configuration object associated with the model.

 None
rbln_config RBLNModelConfig | dict | None

Configuration for RBLN model compilation and runtime. This can be provided as a dictionary or an instance of the model's configuration class (e.g., RBLNLlamaForCausalLMConfig for Llama models). For detailed configuration options, see the specific model's configuration class documentation.

 None
kwargs Any

Additional keyword arguments. Arguments with the prefix rbln_ are passed to rbln_config, while the remaining arguments are passed to the HuggingFace library.

 {}

The method performs the following steps:

  1. Compiles the PyTorch model into an optimized RBLN graph
  2. Configures the model for the specified NPU device
  3. Creates the necessary runtime objects if requested
  4. Saves the compiled model and configurations

Returns:

Type Description
RBLNModel

A RBLN model instance ready for inference on RBLN NPU devices.

RBLNCosmosTransformer3DModelConfig

Bases: RBLNModelConfig

Configuration class for RBLN Cosmos Transformer models.

This class inherits from RBLNModelConfig and provides specific configuration options for Transformer models used in diffusion models like Cosmos.

__init__(batch_size=None, num_frames=None, height=None, width=None, max_seq_len=None, embedding_dim=None, num_channels_latents=None, num_latent_frames=None, latent_height=None, latent_width=None, uses_per_frame_timestep=None, **kwargs)

Parameters:

Name Type Description Default
batch_size int | None

The batch size for inference. Defaults to 1.

 None
num_frames int | None

The number of frames in the generated video. Defaults to 121.

 None
height int | None

The height in pixels of the generated video. Defaults to 704.

 None
width int | None

The width in pixels of the generated video. Defaults to 1280.

 None
max_seq_len int | None

Maximum sequence length of prompt embeds.

 None
embedding_dim int | None

Embedding vector dimension of prompt embeds.

 None
num_channels_latents int | None

The number of channels in latent space.

 None
latent_height int | None

The height in pixels in latent space.

 None
latent_width int | None

The width in pixels in latent space.

 None
uses_per_frame_timestep bool | None

Whether the pipeline feeds a per-frame timestep tensor ([B, 1, T, 1, 1]; conditioning frames at t=0) instead of a per-batch scalar.

 None
kwargs Any

Additional arguments passed to the parent RBLNModelConfig.

 {}

Raises:

Type Description
ValueError

If batch_size is not a positive integer.

from_pretrained(path, rbln_config=None, return_unused_kwargs=False, **kwargs) classmethod

Load a RBLNModelConfig from a path.

Parameters:

Name Type Description Default
path str

Path to the RBLNModelConfig file or directory containing the config file.

 required
rbln_config dict[str, Any] | None

Additional configuration to override.

 None
return_unused_kwargs bool

Whether to return unused kwargs.

 False
kwargs Any

Additional keyword arguments to override configuration values. Keys starting with 'rbln_' will have the prefix removed and be used to update the configuration.

 {}

Returns:

Name Type Description
RBLNModelConfig Self | tuple[Self, dict[str, Any]]

The loaded configuration instance.
Note

This method loads the configuration from the specified path and applies any provided overrides. If the loaded configuration class doesn't match the expected class, a warning will be logged.

Examples:

1
config = RBLNResNetForImageClassificationConfig.from_pretrained("/path/to/model")

initialize_submodule_config(submodule_config=None, force_kwargs=False, **kwargs)

Build the config of a submodule from the parent's runtime options and kwargs.

Returns the submodule config instance when its class is known (an instance was passed, or cls_name was given), otherwise the merged kwargs dict for the model class to resolve later.

get_load_overrides()

Extract the options that must cross the load boundary, for passing to from_pretrained.

Returns the explicitly-set runtime options (device, device_map, create_runtimes, activate_profiler, timeout) and the subclass-specific non-save attributes (load-behavior flags such as _load_visual_runtime, which are never serialized so they can only travel with the caller), recursively including submodules. Compile-time attributes are not included: when loading a compiled model they come from the artifact's rbln_config.json, and this dict is merged on top of it.

Returns:

Type Description
dict[str, Any]

Dictionary of load overrides, keyed like the rbln_config dict accepted by from_pretrained.