mixed-precision

The paper introduces LEAN, a loss-engineered regularization method that reshapes a trained model's weight-sensitivity distribution, shrinking the sensitive tail 5-13x and enabling models to survive hundreds of bit-flip attacks while storing ImageNet ResNet-50 at a 3.29-bit payload (2.4x smaller than INT8).

提出PN-Mix，一种面向RoPE解码器LLM的混合精度键缓存量化方法，通过查询-键对范数识别显著通道对，在约2.26位平均精度下达到或超越现有方法在2.32–2.39位下的精度，且可优雅降级至1.5位。

We trained a model on agentic coding tasks using fully asynchronous RL with mixed-precision, mixed-accelerator samplers, achieving training stability on par with homogeneous clusters at lower cost by applying off-policy correction.

论文提出 RAMP（Recurrent Adaptive Mixed Precision），通过分析量化误差在循环深度上的传播规律，将低精度计算置于早期循环步骤、终端步骤恢复全精度，从而在保持循环深度语言模型长上下文预填充质量的同时，在 H100 上实现最高 1.16× 的端到端首填充加速（13.9% 延迟降低）。

Q-PACE 提出一种基于二阶敏感度模型的动态精度分配方法，用于量化感知训练（QAT），在预训练和监督微调实验中（模型规模达4B参数）相比现有混合精度训练方案取得了更优效果，并在更低内存预算下达到可比拟的损失。研究还发现量化敏感度可通过深度和层类型高度预测，从而支持低成本的低频重校准。

This paper presents accurate software models of AMD GPU matrix cores for CDNA 1/2/3 architectures, validated for bit-level reproducibility against hardware, and demonstrates their use in numerical applications to compare accuracy with NVIDIA tensor cores.

TileMix introduces a tile-centric mixed-precision attention mechanism to accelerate long-context prefill in large language models, balancing accuracy and efficiency by routing score-tile groups through FP16 or INT8 paths.

MiCoPro presents an end-to-end hardware/software co-design framework for mixed-precision quantization, using a hardware-aware proxy model to search optimal layer-wise bitwidths under latency constraints and deploy directly to edge hardware, achieving up to 40% latency reduction with less than 3% accuracy drop.

This paper presents NANQ, a noise-floor-aware mixed-precision non-uniform quantization framework for analog compute-in-memory (CIM) systems, which adapts quantization precision to hardware noise characteristics and improves model accuracy under low-bit constraints.

EdgeRazor is a lightweight framework for compressing large language models using entropy-guided mixed-precision quantization-aware distillation, achieving 1.88 bits per parameter while preserving teacher model competence and requiring no changes to inference implementations like llama.cpp. The method is demonstrated on small models such as MobileLLM and Qwen variants.

MixQuant proposes an adaptive mixed-precision quantization framework for LLMs that handles variable memory budgets by marginalizing layer distortion over random upstream configurations, outperforming existing methods across multiple models and budgets.

This paper proposes an operator-aware calibration method for absolute tolerances in tensor kernel correctness tests, using error distribution data to set tighter thresholds. The method achieves a 9.3% absolute recall improvement in bug detection with minimal false positives, demonstrated on the gpuemu corpus.

LaCache introduces a training-free acceleration framework for diffusion-based LLMs that uses lossless caching and precision-adaptive inference to eliminate redundant computation during denoising steps, achieving up to 40.2× end-to-end speedup while maintaining task accuracy.

This paper analyzes the structure of quantization loss in mixed-precision neural networks, showing that saturation makes the loss additive per layer and proposes a coverage model that predicts configuration loss with few parameters, validated on large-scale models.

KronQ is a post-training quantization framework that incorporates gradient covariance using a Kronecker-factored Hessian approximation, enabling bidirectional incoherence processing and improved sensitivity metrics for mixed-precision allocation. It achieves low perplexity even at 2-bit weight-only quantization on large models like LLaMA-3-70B.

A community researcher enabled running GLM-5.2 (753B parameters, all 256 experts) in vLLM without pruning via a hybrid quantization (NVFP4, NF3, MXFP8), fitting on 4×96GB GPUs with ~307k KV cache and near-FP8 accuracy.

This paper introduces MODE, a modality-decomposed expert-level mixed-precision quantization framework for MoE multimodal LLMs that addresses biases in expert importance estimation by decomposing selection frequency by modality and filtering redundant vision tokens, achieving minimal performance loss under aggressive quantization.

A mixed-precision quantization of Google's Gemma-4-12B-it model using NVFP4 for MLP weights and FP8 for attention layers, achieving 25% smaller footprint and faster throughput while maintaining quality.

dMX is a differentiable mixed-precision quantization framework that learns optimal floating-point bit-width assignments per layer for LLMs, targeting the MXFP family of formats defined by the OCP standard. It uses continuous optimization with temperature-based annealing and a budget-aware regularization term, consistently outperforming KL-divergence heuristics on Llama, Qwen3, and SmolLM2 models.

Researchers from UiT and University of Oslo propose a differentiable NAS framework that jointly optimizes architectural configurations and mixed-precision quantization for LLM compression, achieving up to 1.4× faster inference or 6% higher accuracy across seven reasoning tasks compared to sequential NAS-then-quantization baselines.