Training large language models (LLMs) at 4-bit precision offers substantial efficiency gains but remains challenging due to the limited dynamic range and coarse numerical resolution. Existing 4-bit training pipelines typically rely on max-scaling, which is ill-suited for heavy-tailed LLM tensor distributions and leads to severe under-utilization of the FP4 quantization grid in the low-magnitude region. This effect causes pronounced representation collapse and large rounding errors for the values that dominate LLM computation. In this work, we derive the theoretically optimal scaling for FP4 under heavy-tailed inputs, revealing why max-scaling is intrinsically suboptimal. Guided by this analysis, we propose Half-S, a simple and efficient scaling strategy that uses half-scaling as a hardware-friendly default and falls back to an MSE-based clipping threshold when needed, yielding a close approximation to the theoretical optimum under real LLM statistics. Extensive experiments on large-scale pretraining and downstream fine-tuning show that Half-S consistently narrows the gap to BF16 in both convergence and final model quality, while preserving the efficiency benefits of 4-bit computation. Under native FP4 support, Half-S is estimated to provide up to 1.8× end-to-end training speedup. These results indicate that Half-S provides a simple and effective correction to max-scaling, substantially improving the stability and accuracy of 4-bit LLM training.

Diffusion Large Language Models (dLLMs) deliver strong long-context processing capability in a non-autoregressive decoding paradigm. However, the considerable computational cost of bidirectional full attention limits the inference efficiency. Although sparse attention is promising, existing methods remain ineffective. This stems from the need to estimate attention importance for tokens yet to be decoded, while the unmasked token positions are unknown during diffusion. In this paper, we present **Focus-dLLM**, a novel training-free attention sparsification framework tailored for accurate and efficient long-context dLLM inference. Based on the finding that token confidence strongly correlates across adjacent steps, we first design a *past confidence-guided indicator* to predict unmasked regions. Built upon this, we propose a *sink-aware pruning strategy* to accurately estimate and remove redundant attention computation, while preserving highly influential attention sinks. To further reduce overhead, this strategy reuses identified sink locations across layers, leveraging the observed cross-layer consistency. Experimental results show that our method offers more than 29× lossless speedup under 32K context length.

Extensive LLM applications demand efficient structured generations, particularly for LR(1) grammars, to produce outputs in specified formats (e.g., JSON). Existing methods primarily parse LR(1) grammars into a pushdown automaton (PDA), leading to runtime execution overhead for context-dependent token processing, especially inefficient under large inference batches.To address these issues, we propose Pre³ that exploits deterministic pushdown automata (DPDA) to optimize the constrained LLM decoding efficiency.First, by **pre**computing **pre**fix-conditioned edges during the **pre**processing, Pre³ enables ahead-of-time edge analysis and thus makes parallel transition processing possible.Futher, leveraging the prefix-conditioned edges, Pre³ introduces a novel approach that transforms LR(1) transition graphs into DPDA, eliminating the need for runtime path exploration and achieving edge transitions with minimal overhead.Pre³ can be seamlessly integrated into standard LLM inference frameworks, improving time per output token (TPOT) by up to 40% and throughput by up to 36% in our experiments. Our code is available at https://github.com/ModelTC/lightllm.

Recent advancements in large language models (LLMs) are propelling us toward artificial general intelligence with their remarkable emergent abilities and reasoning capabilities. However, the substantial computational and memory requirements limit the widespread adoption. Quantization, a key compression technique, can effectively mitigate these demands by compressing and accelerating LLMs, albeit with potential risks to accuracy. Numerous studies have aimed to minimize the accuracy loss associated with quantization. However, their quantization configurations vary from each other and cannot be fairly compared. In this paper, we present LLMC, a plug-and-play compression toolkit, to fairly and systematically explore the impact of quantization. LLMC integrates dozens of algorithms, models, and hardware, offering high extensibility from integer to floating-point quantization, from LLM to vision-language (VLM) model, from fixed-bit to mixed precision, and from quantization to sparsification. Powered by this versatile toolkit, our benchmark covers three key aspects: calibration data, algorithms (three strategies), and data formats, providing novel insights and detailed analyses for further research and practical guidance for users. Our toolkit is available at https://github.com/ModelTC/llmc.

Post-training quantization (PTQ) of transformer language models faces significant challenges due to the existence of detrimental outliers in activations. We observe that these outliers are concentrated in specific channels and are asymmetric across channels. To address this issue, we propose the Outlier Suppression+ (OS+) framework, which contains the channel-wise shifting for asymmetry and channel-wise scaling for concentration. We show that these operations can be seamlessly migrated into subsequent modules while maintaining equivalence. Second, we propose a fast and stable scheme to calculate effective shifting and scaling values. The channel-wise shifting aligns the center of each channel for removal of outlier asymmetry. The channel-wise scaling quantitatively evaluates changes brought by migration and quantization for better quantization burden balance. We validate our OS+ under both standard and fine-grained quantization settings with models including BERT, OPT, BLOOM, BLOOMZ, and LLaMA. Comprehensive results across various tasks demonstrate the superiority of our approach. Especially, with standard quantization, OS+ can achieve near-floating-point performance on both small models and large language models on 8-bit and 6-bit. Besides, we establish a new state-of-the-art for 4-bit BERT with 15.5% improvement. Our code is available at https://github.com/ModelTC/Outlier_Suppression_Plus.

In this paper, we propose a new knowledge distillation approach called adaptive contrastive knowledge distillation (ACKD) for BERT compression. Different from existing knowledge distillation methods for BERT that implicitly learn discriminative student features by mimicking the teacher features, we first introduce a novel contrastive distillation loss (CDL) based on hidden state features in BERT as the explicit supervision to learn discriminative student features. We further observe sentences with similar features may have completely different meanings, which makes them hard to distinguish. Existing methods do not pay sufficient attention to these hard samples with less discriminative features. Therefore, we propose a new strategy called sample adaptive reweighting (SAR) to adaptively pay more attention to these hard samples and strengthen their discrimination abilities. We incorporate our SAR strategy into our CDL and form the adaptive contrastive distillation loss, based on which we construct our ACKD framework. Comprehensive experiments on multiple natural language processing tasks demonstrate the effectiveness of our ACKD framework.