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

TL;DR

RAMP is a reinforcement learning framework that adaptively assigns mixed precision bit widths to LLM layers, optimizing accuracy and efficiency for on-device inference across various models and hardware.

Contribution

It introduces a novel off-policy Soft Actor Critic method for layer-wise bit width selection, with a transferable policy and a new Scale Folding technique for stable ultra-low precision quantization.

Findings

Achieves 5.54 perplexity at 3.68GB on Llama 2 7B, outperforming uniform quantization methods.

Zero shot transfer of quantization policy to larger and different models.

Retains 99.5% of FP16 reasoning performance in efficient inference pipeline.

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.