Fast LLM Inference - Optimized Task Plan

I hope to implement some acceleration technologies for Large Language Models (LLMs) because I enjoy doing this myself and love the challenge of bringing research papers into real-world applications.

If there are any technologies you'd like to develop or discuss, feel free to reach out. Thanks!

I'm excited to dive deeper into AI research!

---

Updates Log

2024

• 2024/12/16: Add the Medusa-1 Training Script v2
• 2024/12/15: Add the Medusa-1 Training Script
• 2024/12/12: Update the KV Cache support for Speculative Decoding
• 2024/12/04: Add the Kangaroo Training Script v2
• 2024/11/26: Add the Kangaroo Training Script
• 2024/11/22: Update the Target Model Keep Generation Mechanism experiment
• 2024/11/18: Update the Self-Speculative Decoding experiment results of google--gemma-2-9b-it.
• 2024/11/12: Reviewing implementation challenges for Self-Speculative Decoding and evaluating model compatibility for improved efficiency.
• 2024/11/10: Initial setup for Self-Speculative Decoding completed; data pipeline in place for testing draft-and-verify.
• 2024/11/08: Speculative Decoding successfully implemented. Verified improved inference time with no noticeable accuracy degradation.

Pending Decisions

• Batched Speculative Decoding:
• Prompt lookup decoding: Determine timeline after reviewing initial implementations.
• UAG Integration: Assess when to integrate after Medusa and Kangaroo are in place.

---

TODO List

November 2024

• 2024/11/08 | Complete Speculative Decoding following the paper Fast Inference from Transformers via Speculative Decoding
• 2024/11/15 | Implement Self-Speculative Decoding as per Draft & Verify - Lossless Large Language Model Acceleration via Self-Speculative Decoding
  • LayerSkip model architecture
  • Bayesian Optimization for Layer Skip Selection (AR)
  • Adaption Draft-Exiting Mechanism
  • Optimization
  • Bayesian Optimization for Layer Skip Selection (Speed)
  • gemma-2-9b-it experiment
• 2024/11/22 | Develop Kangaroo following Kangaroo - Lossless Self-Speculative Decoding via Double Early Exiting
  • Kangaroo model
  • Training Script
  • Implement double early exits to improve speed.
• 2024/11/29 | Implement Medusa from Medusa - Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
  • Medusa model
  • Training Script (Medusa-1)
  • Testing
• 2025/03 | Implement Hydra from Hydra: Sequentially-Dependent Draft Heads for Medusa Decoding
• 2025/03 | Implement Lookahead Decoding from Break the Sequential Dependency of LLM Inference Using Lookahead Decoding
• 2025/04 | Implement Eagle from EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty
• 2025/04 | Implement Eagle-2 from EAGLE-2: Faster Inference of Language Models with Dynamic Draft Trees
• 2025/04 | Implement Eagle-3 from EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test

Additional Enhancements

• TBD | Implement Batched Speculative Decoding from The Synergy of Speculative Decoding and Batching in Serving Large Language Models
• TBD | Implement prompt lookup decoding from prompt-lookup-decoding GitHub
• TBD | Implement UAG (Universal Assisted Generation) from Universal Assisted Generation Blog