Attention Quantization

FlashAttention improves efficiency through tiling, but its online softmax still relies on floating-point arithmetic for numerical stability, making full quantization difficult. We identify three main obstacles to integer-only FlashAttention: (1) scale explosion during tile-wise accumulation, (2) inefficient shift-based exponential operations on GPUs, and (3) quantization granularity constraints requiring uniform scales for integer comparison. To address these challenges, we propose QFlash, an end-to-end integer FlashAttention design that performs softmax entirely in the integer domain and runs as a single Triton kernel. On seven attention workloads from ViT, DeiT, and Swin models, QFlash achieves up to 6.73$\times$ speedup over I-ViT and up to 8.69$\times$ speedup on Swin, while reducing energy consumption by 18.8% compared to FP16 FlashAttention, without sacrificing Top-1 accuracy on ViT/DeiT and remaining competitive on Swin under per-tensor quantization. Our code is publicly available at https://github.com/EfficientCompLab/qflash.

We present Open-TQ-Metal, the first implementation of fused compressed-domain attention on Apple Silicon, enabling 128K-context inference for Llama 3.1 70B on a single 64GB consumer Mac -- a configuration impossible with all existing inference frameworks. Open-TQ-Metal quantizes the KV cache to int4 on the fly and computes attention directly on the compressed representation via custom Metal compute shaders, eliminating all intermediate dequantization matrices. Across 330 experiments spanning two model families (Gemma 4 31B and Llama 3.1 70B), the fused sdpa_int4 kernel achieves 48x attention speedup at 128K context over the dequantize-then-attend baseline, reduces KV cache memory from 40 GB to 12.5 GB (3.2x compression), and maintains identical top-1 token predictions to FP16 inference. We further provide the first cross-architecture analysis of KV cache quantization methods, revealing that the attention scale factor -- not model size -- determines whether angular quantization schemes like PolarQuant succeed or fail, with Gemma 4's attn_scale=1.0 amplifying directional error 25-100x more than Llama's standard 1/sqrt(d) scaling.

Transformer-based large language models (LLMs) have demonstrated remarkable performance across a wide range of real-world tasks, but their inference cost remains prohibitively high due to the quadratic complexity of attention and the memory bandwidth limitations of high-precision operations. In this work, we present a low-bit mixed-precision attention kernel using the microscaling floating-point (MXFP) data format, utilizing the computing capability on next-generation GPU architectures. Our Diagonal-Tiled Mixed-Precision Attention (DMA) incorporates two kinds of low-bit computation at the tiling-level, and is a delicate fused kernel implemented using Triton, exploiting hardware-level parallelism and memory efficiency to enable fast and efficient inference without compromising model performance. Extensive empirical evaluations on NVIDIA B200 GPUs show that our kernel maintains generation quality with negligible degradation, and meanwhile achieves significant speedup by kernel fusion. We release our code at https://github.com/yifu-ding/MP-Sparse-Attn.

We propose AXELRAM, a smart SRAM macro architecture that computes attention scores directly from quantized KV cache indices without dequantization. The key enabler is a design-time fixed codebook: orthogonal-transform-based quantization concentrates each coordinate's distribution to N(0,1/d), so the optimal quantizer depends only on dimension d and bit-width b, not on input data. The asymmetric path design -- transform on write, table-lookup on read with no inverse transform -- reduces per-query multiplications by 102.4x (a mathematical identity). Through multi-seed evaluation (10 seeds x 3 models), we discover that sign pattern sensitivity causes catastrophic PPL spikes (Delta > 50) on certain models (Qwen2.5-3B), while others (LLaMA-3.1-8B) are fully stable. This phenomenon extends SpinQuant's observation of rotation variance in weight quantization to the KV cache domain, where the effect is qualitatively more severe. We trace the root cause to layer-wise norm heterogeneity and propose a gradient-free sign pattern selection (200 candidates, 8 calibration samples, one-time) that eliminates catastrophic spikes with zero additional hardware cost. All source code is available at https://github.com/Axelidea/AXELRAM.

Transformers have achieved widespread and remarkable success, while the computational complexity of their attention modules remains a major bottleneck for vision tasks. Existing methods mainly employ 8-bit or 4-bit quantization to balance efficiency and accuracy. In this paper, with theoretical justification, we indicate that binarization of attention preserves the essential similarity relationships, and propose BinaryAttention, an effective method for fast and accurate 1-bit qk-attention. Specifically, we retain only the sign of queries and keys in computing the attention, and replace the floating dot products with bit-wise operations, significantly reducing the computational cost. We mitigate the inherent information loss under 1-bit quantization by incorporating a learnable bias, and enable end-to-end acceleration. To maintain the accuracy of attention, we adopt quantization-aware training and self-distillation techniques, mitigating quantization errors while ensuring sign-aligned similarity. BinaryAttention is more than 2x faster than FlashAttention2 on A100 GPUs. Extensive experiments on vision transformer and diffusion transformer benchmarks demonstrate that BinaryAttention matches or even exceeds full-precision attention, validating its effectiveness. Our work provides a highly efficient and effective alternative to full-precision attention, pushing the frontier of low-bit vision and diffusion transformers. The codes and models can be found at https://github.com/EdwardChasel/BinaryAttention.

Quantization addresses the high resource demand for large language models (LLMs) by alleviating memory pressure and bandwidth congestion and providing significantly scaled compute power with a tolerable impact on accuracy. Four-bit floating point (FP4), the lowest-precision format that preserves essential numerical properties such as exponent and sign, has begun to be adopted in cutting-edge architectures, including Blackwell and AMD CDNA, to support LLM quantization and reduce deployment costs. Although aggressive quantization can yield efficiency gains, the quantization sensitivity of within-transformer layers and whether these sensitivities generalize across existing FP4 formats and model scales remain underexplored. To elucidate quantization sensitivity, this study conducts a systematic analysis of two FP4 formats, MXFP4 and NVFP4, across three Qwen2.5 model scales (0.5B, 7B, and 14B), using controlled component-wise and block-wise isolation methodologies. We observe that MLP up- and down-projection layers consistently dominate in terms of sensitivity, while gate and attention projections are moderately and substantially less sensitive to FP4 quantization, respectively. We further find that sensitivity does not universally localize to the final blocks, but early blocks can be highly sensitive, particularly under MXFP4. Our results provide a diagnostic characterization of the inference behavior of FP4 across components, depths, and FP4 formats.

Low-bit attention, such as SageAttention, has emerged as an effective approach for accelerating model inference, but its applicability to training remains poorly understood. In prior work, we introduced SageBwd, a trainable INT8 attention that quantizes six of seven attention matrix multiplications while preserving fine-tuning performance. However, SageBwd exhibited a persistent performance gap to full-precision attention (FPA) during pre-training. In this work, we investigate why this gap occurs and demonstrate that SageBwd matches full-precision attention during pretraining. Through experiments and theoretical analysis, we reach a few important insights and conclusions: (i) QK-norm is necessary for stable training at large tokens per step, (ii) quantization errors primarily arise from the backward-pass score gradient dS, (iii) reducing tokens per step enables SageBwd to match FPA performance in pre-training, and (iv) K-smoothing remains essential for training stability, while Q-smoothing provides limited benefit during pre-training.

Attention scores in transformers are bilinear forms $S_{ij} = x_i^\top M x_j / \sqrt{d_h}$ whose maximum magnitude governs overflow risk in low-precision training. We derive a rank-aware concentration inequality: when the interaction matrix $M = W^Q W^{K\top}$ has rank $r \ll d$, tail probabilities for $\max_{i,j}|S_{ij}|$ decay as $\exp(-d^{2}α^{2}/(γr))$ rather than $\exp(-dα^{2})$, where $γ> 1$ is a typicality parameter. For transformer attention where $r = d_h$, this yields $8$--$28\times$ tighter concentration than rank-agnostic bounds in modern architectures. We apply this result to FP8 training, deriving geometry-aware scale factors that provide principled overflow guarantees without observing activations. The method computes per-layer scales from the spectral norm $\|W^Q W^{K\top}\|_2$ via implicit power iteration, includes a grouped query attention formulation that avoids key expansion, and remains compatible with fused attention kernels. Across GPT-2 XL to Llama-2-70B, geometry-aware scaling eliminates overflows in transient scenarios where delayed scaling fails, while achieving comparable downstream MMLU accuracy.