DDTree-MLX
A tree-based speculative decoding method for Apple Silicon that extends the DFlash draft model with a draft tree, verifying several candidate continuations in one pass.
DDTree-MLX extends DFlash speculative decoding for Apple's MLX framework by building a draft tree from per-position token probabilities rather than betting on one draft sequence, then verifying every node of the tree against the target model in a single forward pass using tree attention masks and per-token RoPE positions. Because every token is still checked against the target model, the output stays lossless.
The README reports, on a Mac Studio M3 Ultra running a 27B model in 4-bit, roughly 10 to 15 percent faster generation than DFlash alone on code and structured content, where the draft model's predictions are accepted often enough for the extra branches to pay off; on creative prose, where acceptance is low, the tree adds little.
It is implemented in Python with custom Metal kernels and is based on a paper on block-diffusion draft trees.