Reproduce “Inference economics of language models” 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
Ege Erdil
The authors develop a hardware-aware model estimating inference speed and cost from computation, memory traffic, GPU communication, and fixed collective latency. It searches batch sizes and GPU parallelism to construct speed-cost frontiers. For 16-bit Llama 3 70B with speculative decoding, it predicts speed rises from 69 to 95 tokens per second at nearly equal modeled costs; these are predictions, not live-system measurements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.