- Fix mask format conversion (float to boolean) for scaled_dot_product_attention
- Fix mask dimensions for proper broadcasting [batch, 1, seq_len, seq_len]
- Resolve conflict between is_causal and custom mask parameters
- Enable training with optimized attention and KV caching
- Complete transformer implementation from scratch
- Training pipeline with gradient accumulation and mixed precision
- Optimized inference with KV caching
- Multi-format data processing (PDFs, images, code, text)
- Comprehensive documentation
- Apache 2.0 license
- Example training plots included in docs/images/
