Researchers Release FlashMap for 6.14x Faster LLM Decoding

FlashMap replaces traditional discrete diffusion draft methods with a continuous flow map approach. This allows it to draft a full 16-token block in a single forward pass, ensuring token consistency within the block.

The research team reported significant speedups, achieving up to 6.14x lossless speedup on the GSM8K benchmark using the Qwen3-8B model. This performance makes FlashMap the fastest drafter tested on Qwen3-8B, outperforming a retrained DFlash by approximately 7% on average.

FlashMap operates by mapping Gaussian noise to a token block in one step, with a small drafter proposing K tokens. The target model then checks these K tokens in parallel, retaining the longest matching prefix. This process ensures the output remains identical to what the target model would produce.

The efficiency of FlashMap depends on the acceptance rate of drafted tokens and the cost of drafting. Unlike autoregressive methods like EAGLE-3, where cost increases with block length, FlashMap drafts the entire block in one pass using block diffusion.

The researchers noted that diffusion draft methods can lose acceptance due to predicting each slot independently, a phenomenon they term 'factorized approximation.' FlashMap addresses this by using noise samples that jointly determine the mode for every slot within a block.