DeepSeekVL2<T>

Models & Types · AiDotNet.VisionLanguage.InstructionTuned

DeepSeek-VL2: MoE vision-language model with dynamic tiling and multi-head latent attention.

Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → DeepSeekVL2<T>

Implements: IInstructionTunedVLM<T>

For Beginners

DeepSeek-VL2 is a major upgrade that combines Mixture-of-Experts with dynamic image tiling. It has 64 total experts but only activates 6 per token, making it efficient despite its large capacity. Dynamic tiling lets it handle images at variable resolutions by splitting them into appropriately-sized tiles. It also uses multi-head latent attention (MLA) to compress the key-value cache, reducing memory usage during inference. Available in 1.8B, 4.5B, and 27B sizes for different compute budgets. Default values follow the original paper settings.

How It Works

DeepSeek-VL2 (Wu et al., 2024) advances multimodal understanding with Mixture-of-Experts (MoE) architecture, dynamic image tiling for variable resolution input, and multi-head latent attention for efficient KV cache compression. It uses 64 experts with 6 active per token.

References:

  • Paper: “DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding” (Wu et al., 2024)

Example

// Create a DeepSeek-VL2 model for advanced multimodal understanding
// using MoE architecture with dynamic tiling and MLA attention
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.TwoDimensional,
    taskType: NeuralNetworkTaskType.ImageClassification,
    inputHeight: 224, inputWidth: 224, inputDepth: 3, outputSize: 512);
            
// ONNX inference mode with pre-trained model
var model = new DeepSeekVL2<double>(architecture, "deepseekvl2.onnx");
            
// Training mode with native layers
var trainModel = new DeepSeekVL2<double>(architecture, new DeepSeekVL2Options());

Constructors

Constructor Summary
DeepSeekVL2(NeuralNetworkArchitecture<T> architecture, DeepSeekVL2Options options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
DeepSeekVL2(NeuralNetworkArchitecture<T> architecture, string modelPath, DeepSeekVL2Options options)

Properties

Property Summary
int DecoderEmbeddingDim { get; }
int EmbeddingDimension { get; }
int MaxGenerationLength { get; }
string LanguageModelName { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Chat(Tensor<T> image, IEnumerable<ValueTuple<string, string>> conversationHistory, string userMessage)
Tensor<T> EncodeImage(Tensor<T> image)
Tensor<T> GenerateFromImage(Tensor<T> image, string prompt) Generates text using DeepSeek-VL2’s MoE + dynamic tiling architecture.