ConMeZO: Adaptive Descent-Direction Sampling for Gradient-Free Finetuning of Large Language Models
This addresses the memory bottleneck in finetuning large language models for AI researchers, though it is incremental as it builds on existing zeroth-order methods.
The paper tackles the slow convergence of zeroth-order optimization for finetuning large language models by proposing ConMeZO, which uses adaptive directional sampling to accelerate convergence, achieving up to 2X faster performance than MeZO while maintaining low memory usage.
Zeroth-order or derivative-free optimization (MeZO) is an attractive strategy for finetuning large language models (LLMs) because it eliminates the memory overhead of backpropagation. However, it converges slowly due to the inherent curse of dimensionality when searching for descent directions in the high-dimensional parameter space of billion-scale LLMs. We propose ConMeZO, a novel zeroth-order optimizer that accelerates convergence by adaptive directional sampling. Instead of drawing the direction uniformly at random, ConMeZO restricts the sampling to a cone centered around a momentum estimate. This concentrates the search in directions where the true gradient is more likely to lie and thus reduces the effect of high dimensions. We prove that ConMeZO achieves the same worst-case convergence rate as MeZO. Empirically, when finetuning LLMs on natural language tasks, ConMeZO is up to 2X faster than MeZO while retaining the low-memory footprint of zeroth-order methods.