Qwen2.5-VL

The Qwen2.5-VL model is a vision-language model designed for tasks like Visual Question Answering (VQA), image captioning, and video understanding. It can process both image and video inputs alongside text, making it highly versatile for multimodal applications. RBLN NPUs can accelerate Qwen2.5-VL model inference using Optimum RBLN.

Classes

RBLNQwen2_5_VLForConditionalGeneration

Bases: RBLNDecoderOnlyModelForCausalLM

RBLNQwen2_5_VLForConditionalGeneration is a multi-modal model that integrates vision and language processing capabilities, optimized for RBLN NPUs. It is designed for conditional generation tasks that involve both image and text inputs.

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

Important Note

This model includes a Large Language Model (LLM). For optimal performance, it is highly recommended to use tensor parallelism for the language model. This can be achieved by using the rbln_config parameter in the from_pretrained method. Refer to the from_pretrained documentation and the RBLNQwen2_5_VLForConditionalGenerationConfig class for details.

Examples:

from optimum.rbln import RBLNQwen2_5_VLForConditionalGeneration

model = RBLNQwen2_5_VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-VL-7B-Instruct",
    export=True,
    rbln_config={
        "visual": {
            "max_seq_lens": 6400,
            "device": 0,
        },
        "tensor_parallel_size": 8,
        "kvcache_partition_len": 16_384,
        "max_seq_len": 114_688,
        "device": [0, 1, 2, 3, 4, 5, 6, 7],
    },
)

model.save_pretrained("compiled-qwen2.5-vl-7b-instruct")

Functions

from_pretrained(model_id, export=False, 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	Union[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	A boolean flag to indicate whether the model should be compiled.	False
rbln_config	Optional[Union[Dict, RBLNModelConfig]]	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	Dict[str, 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
Self	A RBLN model instance ready for inference on RBLN NPU devices.

from_model(model, *, rbln_config=None, **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	The PyTorch model to be compiled. The object must be an instance of the HuggingFace transformers PreTrainedModel class.	required
rbln_config	Optional[Union[Dict, RBLNModelConfig]]	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	Dict[str, 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
Self	A RBLN model instance ready for inference on RBLN NPU devices.

save_pretrained(save_directory)

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

Parameters:

Name	Type	Description	Default
save_directory	Union[str, PathLike]	The directory to save the model and its configuration files. Will be created if it doesn't exist.	required

generate(input_ids, attention_mask=None, max_length=None)

The generate function is utilized in its standard form as in the HuggingFace transformers library. User can use this function to generate text from the model.

Parameters:

Name	Type	Description	Default
input_ids	LongTensor	The sequence used as a prompt for the generation	required
attention_mask	Optional[Tensor]	The attention mask to apply on the sequence	None
max_length	Optional[int]	The maximum length of the sequence to be generated	None

Classes

RBLNQwen2_5_VisionTransformerPretrainedModelConfig

Bases: RBLNModelConfig

Functions

__init__(max_seq_lens=None, **kwargs)

Parameters:

Name	Type	Description	Default
max_seq_lens	int	The maximum sequence length for the Vision Transformer, representing the number of patches in a sequence derived from an image or a video frame. See the "Max Seq Lens" section below for details.	None
**kwargs	Dict[str, Any]	Additional keyword arguments passed to the parent RBLNModelConfig. These may include general model settings like model architecture or device-specific configurations.	{}

Raises:

Type	Description
ValueError	If max_seq_lens is None or not provided.
ValueError	If max_seq_lens (or any value in the list) is not a positive integer.
ValueError	If max_seq_lens is not a multiple of (window_size / patch_size)^2 for window-based attention, or is insufficient for the expected image/video resolution.
ValueError	If batch_size (inherited from RBLNModelConfig) is not a positive integer.

Max Seq Lens

Since Qwen2_5_VLForConditionalGeneration performs inference on a per-image or per-frame basis, max_seq_lens should be set based on the maximum expected resolution of the input images or video frames, according to the following guidelines:

1. Minimum Value: max_seq_lens must be greater than or equal to the number of patches generated from the input image. For example, a 224x224 image with a patch size of 14 results in (224 / 14) * (224 / 14) = 256 patches. Therefore, max_seq_lens must be at least 256.
2. Alignment Requirement: max_seq_lens must be a multiple of (window_size / patch_size)^2 due to the requirements of the window-based attention mechanism. For instance, if window_size is 112 and patch_size is 14, then (112 / 14)^2 = 64, meaning valid values for max_seq_lens include 64, 128, 192, 256, etc.

RBLNQwen2_5_VLForConditionalGenerationConfig

Bases: RBLNDecoderOnlyModelForCausalLMConfig

Functions

__init__(visual=None, **kwargs)

Parameters:

Name	Type	Description	Default
visual	Optional[RBLNModelConfig]	Configuration for the vision encoder component.	None
**kwargs	Dict[str, Any]	Additional keyword arguments passed to the parent class RBLNDecoderOnlyModelForCausalLMConfig. See the parent class documentation for details.	{}

Raises:

Type	Description
ValueError	If the visual configuration is provided but contains invalid settings, such as an invalid max_seq_lens (e.g., not a positive integer, not a multiple of the window-based attention unit, or insufficient for the expected resolution).
ValueError	If visual is None and no default vision configuration can be inferred for the model architecture.
ValueError	If any inherited parameters violate constraints defined in the parent class, such as batch_size not being a positive integer, prefill_chunk_size not being divisible by 64, or max_seq_len not meeting requirements for Flash Attention.