QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs
This addresses efficiency bottlenecks for researchers and practitioners training RL on large language models, though it appears incremental as it builds on existing quantization and adaptation techniques.
The paper tackles the resource-intensive nature of reinforcement learning (RL) for large language models by proposing QeRL, a framework that combines NVFP4 quantization with Low-Rank Adaptation to accelerate rollout phases and reduce memory overhead, achieving over 1.5 times speedup and enabling RL training of a 32B model on a single GPU while matching full-parameter fine-tuning performance on benchmarks like GSM8K (90.8%) and MATH 500 (77.4%).
We propose QeRL, a Quantization-enhanced Reinforcement Learning framework for large language models (LLMs). While RL is essential for LLMs' reasoning capabilities, it is resource-intensive, requiring substantial GPU memory and long rollout durations. QeRL addresses these issues by combining NVFP4 quantization with Low-Rank Adaptation (LoRA), accelerating rollout phase of RL while reducing memory overhead. Beyond efficiency, our findings show that quantization noise increases policy entropy, enhancing exploration, and enabling the discovery of better strategies during RL. To further optimize exploration, QeRL introduces an Adaptive Quantization Noise (AQN) mechanism, which dynamically adjusts noise during training. Experiments demonstrate that QeRL delivers over 1.5 times speedup in the rollout phase. Moreover, this is the first framework to enable RL training of a 32B LLM on a single H100 80GB GPU, while delivering overall speedups for RL training. It also achieves faster reward growth and higher final accuracy than 16-bit LoRA and QLoRA, while matching the performance of full-parameter fine-tuning on mathematical benchmarks such as GSM8K (90.8%) and MATH 500 (77.4%) in the 7B model. These results establish QeRL as an efficient and effective framework for RL training in LLMs.