Speculative Speculative Decoding

Autoregressive decoding is bottlenecked by its sequential nature. Speculative decoding has become a standard way to accelerate inference by using a fast draft model to predict upcoming tokens from a slower target model, and then verifying them in parallel with a single target model forward pass. However, speculative decoding itself relies on a sequential dependence between speculation and verification. We introduce speculative speculative decoding (SSD) to parallelize these operations. While a verification is ongoing, the draft model predicts likely verification outcomes and prepares speculations pre-emptively for them. If the actual verification outcome is then in the predicted set, a speculation can be returned immediately, eliminating drafting overhead entirely. We identify three key challenges presented by speculative speculative decoding, and suggest principled methods to solve each. The result is Saguaro, an optimized SSD algorithm. Our implementation is up to 2x faster than optimized speculative decoding baselines and up to 5x faster than autoregressive decoding with open source inference engines.

Hide a message in an emoji

Leap Fuels Hands-On Learning, Exploration

Lifelong learners and tech enthusiasts don’t view openSUSE Leap as just a stable operating system, but a launchpad for discovery. Malcolm, who shared with th...

Hype for the Future 19B: New England Decoded

Maine Located at the northeastern extreme of the New England, the Northeast, and the contiguous United States (CONUS) as a whole, the State of Maine is the most sparsely populated state of the New …

Reverse Engineering a 27MHz RC Toy communication using RTL SDR

My kids have this RC fire engine that works in the 27MHz band. I got curious how the communication is, with objective to control the toy from laptop. I had an RTL SDR in my toolbox. I have used it …

Under the hood: Vec<T>

Let's look at Vec<T> to get a better understanding of its inner structure.