Reproduce “Compiler-Assisted Speculative Sampling for Accelerated LLM Inference on Heterogeneous Edge Devices” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alejandro Ruiz y Mesa, Guilherme Korol, Moritz Riesterer, João Paulo Cardoso de Lima, Jeronimo Castrillon
The authors present a compiler-guided method for deciding when speculative decoding can speed edge translation. A small drafter proposes tokens for a larger target to verify together, while profiling and a cost model guide CPU/GPU placement and draft length. On one edge SoC, the method estimates up to 1.68× speedup under favorable acceptance; at median acceptance 0.17, it predicts none.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.