RAMP: Reinforcement Adaptive Mixed Precision Quantization for Efficient On Device LLM Inference

arXiv cs.LG / 3/19/2026

📰 NewsDeveloper Stack & InfrastructureTools & Practical UsageModels & Research

共有:

Key Points

RAMP is a reinforcement learning-based method that performs per-layer mixed-precision quantization to minimize perplexity under a global bit budget for efficient on-device LLM inference.
The policy conditions on an 11-dimensional embedding of activation statistics, weight properties, and structural descriptors to enable zero-shot transfer across model families and scales.
Scale Folding is a preconditioning technique that migrates activation outliers into weights via per-channel scaling and normalization layer compensation to enable stable sub-4-bit quantization.
On Llama 2 7B, RAMP achieves 5.54 perplexity at 3.68GB (3.65 effective bits), outperforming uniform 4-bit AWQ and GPTQ, and the policy generalizes zero-shot to Llama 2 13B and Mistral 7B, with the HALO pipeline exporting allocations to GGUF for kernel-free inference on CPUs, GPUs, and edge devices while retaining 99.5% of FP16 performance.

Abstract

Post training quantization is essential for deploying large language models (LLMs) on resource constrained hardware, yet state of the art methods enforce uniform bit widths across layers, yielding suboptimal accuracy efficiency trade offs. We present RAMP (Reinforcement Adaptive Mixed Precision), an off policy Soft Actor Critic framework that learns per layer bit width assignments to minimize perplexity under a global bit budget. The policy conditions on an 11 dimensional embedding of activation statistics, weight properties, and structural descriptors, enabling zero shot transfer across model families and scales. To enable stable sub 4 bit quantization, we introduce Scale Folding, a preconditioning technique that migrates activation outliers into weights via per channel scaling and normalization layer compensation. A quality prioritized reward with asymmetric penalties and budget cliffs drives rapid convergence. On Llama 2 7B, RAMP achieves 5.54 perplexity at 3.68GB (3.65 effective bits), outperforming uniform 4 bit AWQ (5.60 at 3.90 GB) and GPTQ by 6% in size and 1% to3% in quality. Critically, a policy trained only on Llama 2 7B generalizes zero shot to Llama 2 13B and Mistral 7B, often surpassing target specific training, supporting the hypothesis that quantization sensitivity is primarily architectural. The HALO pipeline exports allocations to GGUF format for kernel free inference on CPUs, GPUs, and edge devices, retaining 99.5% of FP16 commonsense reasoning performance.