Expand description
Quantized BLIP text module implementation.
Provides the text decoder portion of the BLIP model with 8-bit quantization. Uses a BERT-style transformer architecture for text processing.
Key components:
- Text embeddings layer with position embeddings
- Multi-head self attention layers
- Cross-attention for vision-text fusion
- Layer normalization and feed-forward layers
- Quantized linear transformations
References:
Re-exports§
pub use crate::quantized_var_builder::VarBuilder;