arXiv:2601. 07475v2 Announce Type: replace-cross Abstract: The emergence of fine-grained numerical formats like NVFP4 presents new opportunities for efficient Large Language Model (LLM) inference.
DualQuant introduces a dual‑scale, calibration‑free post‑training quantization framework that jointly optimizes weight‑reconstruction MSE and derives per‑channel activation rescalings, enabling effective 4‑bit weight and activation quantization without calibration data. The method employs a non‑differentiable optimization solved via variable splitting and block coordinate descent, supporting a wide range of formats including MX and NVFP4. Experiments on Llama and Qwen models show DualQuant matches or surpasses calibration‑based SmoothQuant and significantly outperforms other calibration‑free baselines, achieving up to 2.61× memory reduction and 2.46× throughput speedup on an RTX 4090.
arXiv:2607. 26515v1 Announce Type: new Abstract: We present, to our knowledge, the first end-to-end FP4 RL post-training, in which both the rollout and training policies, including their forward and backward passes, operate at 4-bit precision.
arXiv:2609.26333v1 Announce Type: new Abstract: Prefill and decode reward different approaches to quantization: low-precision arithmetic accelerates prompt processing, while compact weights reduce me...
The paper presents a comprehensive evaluation of post‑training quantization (PTQ) techniques for large language models (LLMs), addressing gaps in prior work by conducting fair, large‑scale experiments on text‑only LLMs. It decouples quantization into pre‑quantization transformation and error mitigation, systematically studies granularity, symmetry, and data formats (MXFP4, NVFP4), and finds that rotation‑based transformations and low‑rank compensation with GPTQ can improve performance, while finer granularity and asymmetric quantization benefit activations more than weights. The study highlights the importance of understanding component interactions and the impact of scaling factor formats on low‑bit quantization performance.
arXiv:2607. 04422v1 Announce Type: cross Abstract: Recent NVFP4 pretraining methods mainly target transformer linear layers, leaving optimizer states, optimizer arithmetic and attention underexplored in 4-bit pipelines.
arXiv:2606. 26587v1 Announce Type: cross Abstract: Low-bit floating-point formats and semi-structured sparsity are increasingly supported by modern accelerators, yet combining them for LLM activation compression remains challenging: activations contain input-dependent outliers that dominate block scales in FP4 quantization, and directly applying N:M sparsity masks discards moderate values, coupling sparsification loss with quantization error.
The paper investigates low‑bit quantization for Multimodal Large Language Models (MLLMs), showing that MXFP8 retains near‑lossless performance while 4‑bit formats like MXFP4 and HiF4 cause significant degradation. It identifies activation quantization as the main source of this loss and introduces Residual Fallback Quantization (RFQ), a lightweight framework that adds a quantized residual pathway to improve activation fidelity without architectural changes. Experiments on Wan2.2 and Qwen3‑VL demonstrate that RFQ recovers much of the performance gap to BF16 baselines across generation and reasoning tasks.
The paper introduces ALTO, an open‑source training recipe that enables a 20‑billion‑parameter mixture‑of‑experts LLM (GPT‑OSS‑20B) to train on the MLPerf Small MoE benchmark using the MXFP4 sub‑8‑bit format. By combining hybrid 1D/2D block quantization, randomized Hadamard transforms, stochastic rounding, and a simplified weight de‑oscillation scheme, the authors achieve a validation loss of 3.3350 after 16,128 steps—only 0.007 above the BF16 baseline—while reducing the FP4 convergence overhead from +20.0% to +5.0% relative to BF16. The study also reports negative results for differential gradient estimation, outlier clipping, and macro‑block scaling, despite operator‑level SNR improvements, and notes that the experiments use simulated MXFP4 kernels on AMD MI300 hardware without wall‑clock optimization.
ResCom introduces a low‑rank correction that leverages activation rounding residuals to compensate for layer‑output errors in low‑bit weight–activation quantization of large language models. The method selects residual directions based on their impact on output reconstruction and computes the correction in closed form, avoiding gradient training or weight updates. Experiments on Llama‑3‑8B with 4‑bit weights and activations show significant gains: perplexity drops from 7.708 to 7.147, MMLU improves by 2.89 percentage points, and five‑shot GSM8K accuracy rises from 23.96 % to 36.16 %.
arXiv:2606. 21257v4 Announce Type: replace-cross Abstract: openPangu models are attractive targets for private and domestic large-language-model deployment, yet their robustness under aggressive post-training quantization on Ascend NPUs has not been systematically characterized.
SpecQuant is a training‑free framework that merges speculative decoding with multi‑parent quantization to enable adaptive, efficient inference of large language models. It generates several quantized variants (INT4, FP8, FP16) from a single base model and routes queries to the appropriate variant based on predicted complexity, using lightweight models for simple tasks and full‑precision models for complex reasoning. Evaluations on Qwen2.5 models across MMLU, AlpacaEval, and GSM8K show 35‑43% speedups with less than 2% accuracy loss, facilitating practical on‑device LLM deployment without specialized infrastructure.