arXiv 2503.13565: ML-SpecQD: Multi-Level Speculative Decoding with Quantized Drafts
paper
Fields
id
sources/w15-pap-ml-specqdPublisher
arXiv
Kind
paper
Published
2025-03-17
Accessed
2026-10-11
notes
Paper by Evangelos Georganas, Dhiraj Kalamkar, Alexander Kozlov, Alexander Heinecke (Intel Labs). Proposes multi-level speculative decoding using plug-and-play direct-cast 4-bit (MXFP4) quantized draft models to accelerate memory-bound CPU/edge decoding without fine-tuning separate draft models.
Notes
Paper by Evangelos Georganas, Dhiraj Kalamkar, Alexander Kozlov, Alexander Heinecke (Intel Labs). Proposes multi-level speculative decoding using plug-and-play direct-cast 4-bit (MXFP4) quantized draft models to accelerate memory-bound CPU/edge decoding without fine-tuning separate draft models.
Cross-references
This record references no other records.