Skip to content

Latest commit

 

History

History
8 lines (7 loc) · 388 Bytes

File metadata and controls

8 lines (7 loc) · 388 Bytes

Improving Sub-billion Scale LLM Design Experiments

Some of the techniques used in the LLM pretraining design include:

  • Embedding Sharing
  • Grouped Query Design
  • SwiGLU Activations for the Multi Perceptron Layer
  • Intermidate blockwise weight sharing
chrome_7KTClIuUnl