ALiBi: Attention with Linear Biases
Learn ALiBi, the position encoding method that adds linear biases to attention scores for exceptional length extrapolation in transformers.
19 min readConcept
Explore machine learning concepts related to position-encoding. Clear explanations and practical insights.
Learn ALiBi, the position encoding method that adds linear biases to attention scores for exceptional length extrapolation in transformers.
Learn Rotary Position Embeddings (RoPE), the elegant position encoding using rotation matrices, powering LLaMA, Mistral, and modern LLMs.