Module llava 

LLaVA Model V2 - Clean implementation using solid abstractions

This implements the LLaVA (Large Language and Vision Assistant) architecture including:

• LLaVA-1.5-7B, LLaVA-1.5-13B, LLaVA-1.6-7B, LLaVA-1.6-13B, LLaVA-1.6-34B

Architecture components:

• Vision Tower: CLIP-style vision encoder with patch embeddings and bidirectional attention
• Multimodal Projector: Projects vision features to language model dimension (Linear or MLP)
• Language Model: LLaMA-style decoder with RoPE, GQA, and SwiGLU MLP

Structs

LLaVAConfig

LLaVALanguageAttention

LLaMA-style Attention with RoPE and GQA

LLaVALanguageLayer

LLaMA-style Transformer Layer

LLaVALanguageMLP

LLaMA-style MLP with SwiGLU

LLaVALanguageModel

Complete Language Model

LLaVAModelV2

Complete LLaVA Model

LLaVAMultiModalProjector

Projects vision features to language model dimension

LLaVAVisionAttention

CLIP Vision Attention (bidirectional, no causal mask)

LLaVAVisionEmbeddings

CLIP-style Vision Embeddings

LLaVAVisionEncoder

CLIP Vision Encoder

LLaVAVisionLayer

CLIP Vision Encoder Layer

LLaVAVisionMLP

CLIP Vision MLP (GELU activation)

LLaVAVisionTower

Complete Vision Tower (CLIP ViT)