Accelerating Production LLMs with Combined Token/Embedding Speculators
Abstrak
This technical report describes the design and training of novel speculative decoding draft models, for accelerating the inference speeds of large language models in a production environment. By conditioning draft predictions on both context vectors and sampled tokens, we can train our speculators to efficiently predict high-quality n-grams, which the base model then accepts or rejects. This allows us to effectively predict multiple tokens per inference forward pass, accelerating wall-clock inference speeds of highly optimized base model implementations by a factor of 2-3x. We explore these initial results and describe next steps for further improvements.
Topik & Kata Kunci
Penulis (7)
Davis Wertheimer
Joshua Rosenkranz
Thomas Parnell
Sahil Suneja
Pavithra Ranganathan
Raghu Ganti
Mudhakar Srivatsa
Akses Cepat
- Tahun Terbit
- 2024
- Bahasa
- en
- Sumber Database
- arXiv
- Akses
- Open Access ✓