Parallel Drafter
A draft generation approach in speculative decoding that outputs candidate tokens for multiple positions simultaneously in a single forward pass; the draft length must be determined statically before inference, and representative methods include Medusa and Lookahead Decoding.
Timeline (last 90 days)
A Parallel Drafter generates an entire block of draft tokens in a single forward pass, making it suitable for high-load LLM serving scenarios.
The draft length of a parallel drafter must be statically determined before inference and cannot be dynamically adjusted based on context like an autoregressive drafter.
All Facts (2)
A Parallel Drafter generates an entire block of draft tokens in a single forward pass, making it suitable for high-load LLM serving scenarios.
50%UnverifiedThe draft length of a parallel drafter must be statically determined before inference and cannot be dynamically adjusted based on context like an autoregressive drafter.
50%