FP8-RL: A Practical and Stable Low-Precision Stack for LLM Reinforcement Learning

TL;DR

This paper introduces a practical FP8 low-precision stack for LLM reinforcement learning, significantly improving rollout throughput while maintaining training stability and accuracy.

Contribution

It presents a comprehensive FP8 rollout stack with techniques for quantization, cache management, and mismatch correction, enabling efficient and stable LLM RL training.

Findings

Up to 44% increase in rollout throughput.

Maintains learning behavior comparable to BF16 baselines.

Supports common training and inference backends.

Abstract

Reinforcement learning (RL) for large language models (LLMs) is increasingly bottlenecked by rollout (generation), where long output sequence lengths make attention and KV-cache memory dominate end-to-end step time. FP8 offers an attractive lever for accelerating RL by reducing compute cost and memory traffic during rollout, but applying FP8 in RL introduces unique engineering and algorithmic challenges: policy weights change every step (requiring repeated quantization and weight synchronization into the inference engine) and low-precision rollouts can deviate from the higher-precision policy assumed by the trainer, causing train-inference mismatch and potential instability. This report presents a practical FP8 rollout stack for LLM RL, implemented in the veRL ecosystem with support for common training backends (e.g., FSDP/Megatron-LM) and inference engines (e.g., vLLM/SGLang). We (i) enable FP8 W8A8 linear-layer rollout using blockwise FP8 quantization, (ii) extend FP8 to KV-cache to remove long-context memory bottlenecks via per-step QKV scale recalibration, and (iii) mitigate mismatch using importance-sampling-based rollout correction (token-level TIS/MIS variants). Across dense and MoE models, these techniques deliver up to 44% rollout throughput gains while preserving learning behavior comparable to BF16 baselines.