144 articles tagged with #quantization. AI-curated summaries with sentiment analysis and key takeaways from 50+ sources.
Researchers demonstrate that low-bit quantization of reasoning models introduces a hidden cost: quantized models generate significantly longer chains of thought to maintain accuracy, offsetting per-token speedup gains. The study introduces metrics to measure this token inflation and finds quantization-aware training as the most effective mitigation strategy.
Researchers introduce LUQ, the first ultra-low-bit quantization method for multimodal large language models that achieves 40% memory reduction compared to 4-bit models by analyzing layer-wise entropy and selectively applying extreme compression to simpler layers. The breakthrough addresses a critical deployment bottleneck for vision-language AI systems by recognizing that multimodal tokens require different precision handling than text tokens.
HyperQuant is a new post-training quantization pipeline that compresses large language and diffusion models to 3-5 bits per weight while maintaining near-lossless quality, outperforming existing methods like HIGGS and TurboQuant. The technique combines Hadamard transforms, optimal lattice quantization, and entropy coding to achieve 3.9x compression on model weights and 3.79x on KV cache, enabling more efficient deployment of large AI models.
TileFuse is a new kernel library that enables efficient quantized large language model inference on AMD's XDNA2 NPUs by supporting industry-standard quantization formats like AWQ directly, rather than requiring model reshaping. The technology delivers up to 2x improvements in latency and energy efficiency on edge devices, making practical LLM deployment on consumer hardware substantially more viable.
Researchers introduce LC-QAT, a novel 2-bit quantization method for large language models that combines vector quantization with learnable affine mappings to achieve superior compression with minimal training data. The approach outperforms existing quantization-aware training methods while requiring only 0.1-10% of typical training data, advancing the practical deployment of extremely low-bit LLMs.
Researchers introduce PiSO (Piecewise Scale Optimization), an algorithm that optimizes quantization scaling factors for compressing large language models more effectively than existing heuristic methods. By using calibration data to compute optimal channel-wise scales, PiSO demonstrates consistent improvements in model perplexity and downstream accuracy across Llama and Qwen models, with gains becoming more pronounced at lower bit-widths.
ScaleSweep introduces an optimized block scale initialization method for NVFP4 quantization of large language models, improving upon traditional AbsMax approaches. The technique theoretically bounds the search space and empirically achieves 93% performance retention under aggressive 4-bit quantization, advancing hardware-efficient AI inference.
Researchers introduce APEX4, a pure INT4 inference system that addresses the long-standing challenge of W4A4 quantization in large language models by adapting compute strategies based on GPU architecture. The system achieves up to 2.09× speedup on consumer GPUs while maintaining quality within 0.63 perplexity points of FP16 baselines, making efficient LLM inference more practical across diverse hardware platforms.
Researchers introduce STAR-KV, an adaptive compression framework that reduces KV cache memory requirements in large language models by up to 75% through low-rank projections and intelligent rank selection. The technique achieves up to 20x compression when combined with quantization and delivers significant speedups in attention computation, addressing a critical bottleneck in LLM inference efficiency.
Researchers introduce I-Segmenter, the first fully integer-only Vision Transformer framework for semantic segmentation that eliminates floating-point operations to enable efficient deployment on resource-constrained devices. The model achieves only 5.1% accuracy loss compared to standard floating-point versions while reducing model size by 3.8x and improving inference speed by 1.2x, with a novel activation function addressing quantization challenges.
Researchers introduce an end-to-end framework for compressing Large Language Models through joint structural pruning and mixed-precision quantization that optimizes global error propagation rather than layer-wise errors. The approach demonstrates significant performance improvements at ultra-low bit precisions (1-3 bits), reducing perplexity by up to 21% compared to existing methods.
ActQuant introduces a novel post-training quantization framework that compresses Vision-Language-Action models to sub-4-bit weights while maintaining 94-95% performance, enabling practical deployment on edge devices. The method combines action-guided bit allocation with curvature-aware optimization, achieving 5.3× compression on major VLA models and validated performance on physical robotic hardware.
SAGE-PTQ introduces a novel ultra-low-bit quantization framework for large language models that dramatically reduces scaling overhead while maintaining accuracy. The method achieves 1.03 weight bits per parameter with minimal scaling costs, outperforming existing approaches like BiLLM by orders of magnitude in perplexity metrics while requiring significantly less GPU memory.
Researchers introduce YAQA, a new quantization algorithm that improves model compression by directly optimizing end-to-end error rather than layer-by-layer error. The method achieves 30% error reduction compared to existing approaches like GPTQ and even outperforms quantization-aware training, with theoretical guarantees backing its performance.
Researchers present Recover-LoRA, a technique that recovers accuracy in large language models aggressively quantized to 2-bit precision by applying low-rank adapters trained on synthetic data. The method achieves 7.5-23.3% throughput improvements while recovering 80-95% of lost accuracy on most benchmarks, enabling practical deployment of compressed models on edge devices.
QuBLAST is a new post-training quantization method that compresses large language models by 40-45% while maintaining performance, using block-level mixed-precision quantization and activation scaling to address computational and memory constraints in LLM deployment.
Researchers demonstrate that 2-bit quantization of large reasoning models causes instability leading to longer inference traces rather than speedup, but introduce lightweight recovery techniques (FP16 planning and loop rescue) that restore accuracy from 17-65% to 74-87% while maintaining computational efficiency.
Researchers introduce BitTP, a quantization technique that compresses LLM-based trajectory prediction models to 1.58-bit weights while maintaining full-precision activations, enabling deployment on resource-constrained edge devices. The approach not only reduces memory and latency but actually improves prediction accuracy by 14-21% compared to full-precision baselines, demonstrating that strategic quantization can serve as an effective regularizer.
Researchers introduce Logit-aware Final-block Quantization (LFQ), a technique that improves low-bit quantization of large language models by optimizing the final transformer block to preserve token probability distributions. This advancement addresses quality degradation in generative tasks while maintaining efficiency gains critical for deploying scaled LLMs.
Researchers present PeRQ, a post-training quantization method that uses permutations to optimize block rotations for neural network compression. The approach recovers up to 90% of full-vector rotation performance when quantizing large language models to INT4, significantly outperforming existing block rotation methods.
Researchers propose Hurwitz Quaternion Multiplicative Quantization (HQMQ), a calibration-free method for compressing KV caches in large language models using quaternion mathematics. The technique achieves 5x compression with minimal perplexity loss, matching full-precision performance at ~5 bits while outperforming existing quantization methods across five major model architectures.
Researchers present MobileMoE, a family of sub-billion parameter Mixture-of-Experts language models optimized for on-device deployment that achieve 2-4x efficiency gains over dense models while matching or exceeding performance. The work establishes new on-device scaling laws and delivers the first practical MoE inference implementation on smartphones, with 1.8-3.8x faster performance than existing mobile baselines.
Researchers have developed a bias correction technique for quantizing KV-cache memory in video diffusion models, addressing a fundamental problem where quantization noise causes inflated attention to cached data. The method recovers near-full quality video generation while using 50% less memory than standard approaches, enabling longer video synthesis without sacrificing output quality.
Researchers address a critical failure mode in quantized Vision-Language Models by proposing LRA-EE, a technique that uses early exit strategies to bypass noise-saturated layers in INT8 CLIP. The method improves zero-shot classification accuracy by 2.44 percentage points while reducing computational load by 13.4%, demonstrating that selective layer utilization can recover performance lost to quantization-induced representation collapse.
Researchers introduce Yeti, a compact protein structure tokenizer that converts protein structures into discrete tokens for multimodal AI models. The approach achieves superior codebook utilization and token diversity while maintaining competitive reconstruction accuracy with 10x fewer parameters than existing solutions, enabling efficient joint generation of protein sequences and structures.
