EAGLE-3
The third-generation speculative decoding algorithm developed by the SGLang team. It performs draft prediction in the hidden-state feature space rather than the vocabulary space, substantially improving token acceptance rates, with particularly strong results on code and mathematics tasks.
Timeline (last 90 days)
After adding speculative decoding to the MLX inference backend, reply speed on Mac is up to 2.8x faster with identical output, and draft depth is set automatically, supporting multiple draft models such as MTP, DFlash, DSpark and EAGLE-3.
EAGLE-3 is the latest iteration of the EAGLE series and is a general-purpose speculative decoding solution with high community recognition.
Even in EAGLE-3, the total size of the draft tree remains a pre-configured fixed value during inference.
EAGLE-3 is a strong baseline in the field of speculative decoding, employing fixed draft tree depth, width, and node count.
AdaptiveSpec is implemented on the SGLang production-grade serving engine and achieves up to 56% throughput improvement compared to EAGLE-3.
All Facts (5)
EAGLE-3 is the latest iteration of the EAGLE series and is a general-purpose speculative decoding solution with high community recognition.
70%UnverifiedEven in EAGLE-3, the total size of the draft tree remains a pre-configured fixed value during inference.
50%UnverifiedEAGLE-3 is a strong baseline in the field of speculative decoding, employing fixed draft tree depth, width, and node count.
50%UnverifiedAdaptiveSpec is implemented on the SGLang production-grade serving engine and achieves up to 56% throughput improvement compared to EAGLE-3.
50%UnverifiedAfter adding speculative decoding to the MLX inference backend, reply speed on Mac is up to 2.8x faster with identical output, and draft depth is set automatically, supporting multiple draft models such as MTP, DFlash, DSpark and EAGLE-3.
50%