Understanding Medusa: A Framework for LLM Inference Acceleration with Multiple Decoding Heads

Understanding Medusa: A Framework for LLM Inference Acceleration with Multiple Decoding Heads

Oxen via YouTube Direct link

Introducing Daniel Varoli from Zapata.ai

1 of 15

1 of 15

Introducing Daniel Varoli from Zapata.ai

Class Central Classrooms beta

YouTube videos curated by Class Central.

Classroom Contents

Understanding Medusa: A Framework for LLM Inference Acceleration with Multiple Decoding Heads

Automatically move to the next video in the Classroom when playback concludes

  1. 1 Introducing Daniel Varoli from Zapata.ai
  2. 2 The Problem with LLMs Today
  3. 3 How we Can Solve These Problems
  4. 4 Normal vs. Speculative Architecture
  5. 5 Speculative Decoding Example
  6. 6 Introducing Medusa
  7. 7 Medusa’s Decoding Heads
  8. 8 Generating Tokens With Medusa Heads
  9. 9 Verifying Candidates With Medusa
  10. 10 What if we Mess Up?
  11. 11 Rejecting Sampling For Accepting Candidates
  12. 12 Considering Many Completion Candidates at Once
  13. 13 Tree Attention Diagrams
  14. 14 How to Integrate Medusa Into a LLM
  15. 15 Results

Never Stop Learning.

Get personalized course recommendations, track subjects and courses with reminders, and more.

Someone learning on their laptop while sitting on the floor.