Low-bit inference on accelerators with native 4-bit support requires quantizing both weights and activations to realize the full throughput gains these formats provide. While post-training weight quantization has been extensively studied, activation quantization remains considerably more challenging and typically requires calibration data. In this work, we introduce DUALQUANT, a dual-scale, calibration-free post-training quantization framework that directly minimizes the weight-reconstruction Mean Squared Error (MSE) and, as an immediate by-product, yields per-channel activation rescalings prior to activation quantization. Our key observation is that the scaling factors obtained by minimizing weight-reconstruction error provide effective activation rescalings, enabling competitive activation quantization without calibration data. We formulate DUALQUANT as a non-differentiable weight-reconstruction problem and develop an efficient solver based on variable splitting and block coordinate descent, where each subproblem admits a closed-form update. DUALQUANT provides a unified algorithm supporting a wide range of formats, including the full MX family and NVFP4. To our knowledge, DUALQUANT is the first dual-scaling method to formally support MX and NV block formats. Across Llama and Qwen models evaluated on WikiText-2, C4, and zero-shot downstream tasks, DUALQUANT matches or exceeds the accuracy of calibration-based SmoothQuant across most models and formats, and substantially outperforms calibration-free baselines, with large gains for MX and NV block formats. We validate these gains with a real packed W4A4 inference path for Llama-3.1-8B on an RTX 4090. The resulting implementation closely matches the accuracy of fake quantization while delivering up to 2.61x memory reduction and up to 2.46x throughput speedup over BF16.
Zhentao Zhang, Yanzhao Zheng, Yanlin Liu et al. · cs.LG
Compression 96%
Abstract
On-Policy Distillation (OPD) has emerged as a powerful paradigm for transferring reasoning capabilities from frontier models to smaller models, while extending its success to multi-turn agentic environments remains challenging. We study agentic OPD through interaction states, distinguishing the objective task and executed action-observation record from the internal semantic trajectory of student reasoning. Through this lens, we identify two roles of student reasoning in multi-turn OPD: prior reasoning distorts teacher supervision, while current reasoning shapes the action alternatives sampled at the same state. Specifically, at matched interaction states, removing prior student reasoning improves teacher continuation success and changes the teacher's token-level targets, with larger continuation gains at later turns. Meanwhile, resampling complete reasoning-action responses yields greater action diversity at the same state than resampling actions under a fixed reasoning trace. We introduce State-Grounded On-Policy Distillation, a framework that grounds teacher supervision in interaction states and reallocates the sampling budget to generate multiple complete reasoning-action responses at states with high response entropy, while retaining full-history student generation. Comprehensive evaluations across ALFWorld, ScienceWorld and WebShop demonstrate that State-OPD achieves the highest macro-average success rate among the evaluated methods for both Qwen3-1.7B and Qwen3-4B students. Notably, for Qwen3-4B, State-OPD raises macro-average success from 26.07\% with the strongest baseline to 31.74\% and surpasses its Qwen3-32B teacher on ALFWorld Seen. Further analyses reveal that State-OPD achieves higher task success with fewer supervised turn responses and substantially reduces persistent failed-action repetition; qualitative traces illustrate decision revision based on interaction feedback.
Efficient autoregressive generation is increasingly limited by the memory footprint of the key-value (KV) cache, which grows linearly with sequence length, batch size, and model depth. Existing approaches such as Multi-Query Attention (MQA), Grouped-Query Attention (GQA), Multi-Head Latent Attention (MLA), and Tensor Product Attention (TPA) reduce this cost through different trade-offs, but often sacrifice expressivity, require additional positional-encoding machinery, or impose restrictive low-rank structure. We propose Kronecker Product Attention (KPA), which represents key and value activations as sums of Kronecker products of compact, token-dependent matrix factors. KPA strictly generalizes TPA: rank-one vector-outer-product factorization arises as a degenerate Kronecker split, while non-degenerate matrix factors can represent richer block-structured patterns. We further show that a structured rotary embedding operator matched to the Kronecker representation preserves the relative-position property of RoPE while rotating only the compact factors, removing the need for a decoupled RoPE branch as in MLA. Empirically, KPA matches or exceeds TPA, MHA, MQA, GQA, and MLA across four model scales and nine downstream benchmarks under identical training conditions, while achieving up to 10x KV cache compression. The Kronecker head split is a tuneable hyperparameter that allows practitioners to navigate the compression-accuracy trade-off at design time without architectural changes.
State-space models (SSMs) maintain a fixed-size recurrent state, but reading and writing it at every decoding step can bottleneck hybrid models. This eager state management causes large state I/O with low arithmetic intensity, leaves speculative decoding no cheap rollback, and feeds state quantization error into subsequent recurrent updates. We introduce ReplaySSM, which unrolls the recurrence to represent the state as a checkpoint and a buffer of recent SSM inputs. This separates output computation from checkpoint updates. By updating the checkpoint periodically, ReplaySSM avoids state writes on most steps and roughly halves the dominant state memory traffic. Keeping recent inputs also makes speculative rollback a buffer-pointer operation and enables output-only verification without materializing intermediate states. The same separation allows a low-precision checkpoint to serve output computation on most steps while a higher-precision checkpoint serves recurrent updates, so rounding error in the low-precision copy never enters later updates. We implement ReplaySSM in vLLM for Mamba-2 and Gated DeltaNet. At serving batch size 256, it accelerates standard decoding by $1.20$--$1.47\times$ over a tuned vLLM baseline across four hybrid models, and delivers about $1.8\times$ the throughput of vLLM's speculative decoding on the two 120B-class models. Separate state precisions also reduce output KL divergence and downstream accuracy losses compared with a shared low-precision checkpoint.
Yuhong Chou, Zehao Liu, Yuqi Pan et al. · cs.LG, cs.AI
Inference speed 94%
Abstract
Efficient attention mechanisms, such as linear and sliding window attention, aim to reduce the computational complexity of Transformers from quadratic to linear. However, their degraded in-context recall often necessitates interleaving full attention layers, leaving the quadratic bottleneck unresolved in long-context processing. To address this, we introduce Orthros attention, which shifts the efficiency paradigm from structural allocation to phase-aware decoupling. As a unified module, it dynamically switches its computation pattern across different phases of sequence processing, enabling linear complexity prefilling and full attention decoding. Our design grants Orthros attention a lower asymptotic complexity compared to vanilla Transformers in prevalent long-context scenarios characterized by heavy prefilling and light decoding. Beyond theoretical formulation and complexity analysis, extensive empirical evaluations demonstrate Orthros's superiority across multiple dimensions. First, we validate its fundamental language modeling capabilities, demonstrating scalability across varying model sizes and strong competitiveness in comprehensive architectural comparisons. Second, Orthros exhibits exceptional conversion feasibility via a data-efficient Transformer-to-Orthros conversion, yielding performance that even surpasses the original Transformer. Finally, Orthros achieves a $21.9\times$ prefilling speedup over Transformers with end-to-end inference gains, while preserving long-context capabilities, as demonstrated by near-perfect 128K needle-in-a-haystack results and Transformer-level performance on real-world in-context recall tasks. Overall, these advantages in high performance, flexibility, and efficiency position Orthros as a promising candidate to succeed the Transformer in industrial applications.
Jingru Yang, Jinhang Wu, Minjing Shi et al. · cs.LG, cs.CL
Compression 93%
Abstract
Cross-tokenizer on-policy distillation (OPD) requires aligning teacher and student sequences before their scores can be compared. Decoded-text alignment provides such correspondence at the chunk level, but leaves a distinct problem unresolved: credit alignment. A chunk-level signal specifies how much a span should change, but not how that signal should be distributed across student tokens, while tokens without teacher correspondence remain unconstrained. We propose CAST (Chunk-aware credit Assignment with Selective Token Anchoring), a shift-or-stay framework that addresses both issues. For aligned chunks, CAST redistributes each chunk's advantage across student tokens using predictive uncertainty and local teacher-signal smoothness, emphasizing learning opportunities supported by locally consistent teacher feedback while preserving the chunk's total advantage. The resulting weights exactly solve a KL-regularized credit-allocation problem. For unaligned positions, CAST selectively anchors the student to its initial policy through a KL penalty, limiting drift where teacher supervision is unavailable. Because credit assignment is separated from chunk construction, CAST also applies to punctuation-defined or fixed-length chunks under shared tokenization. Across six mathematics, science, and code benchmarks, CAST improves the average over the strongest evaluated baseline by $2.45$ and $1.32$ points for cross-tokenizer transfer from Qwen3-8B to Mistral-7B and Llama-3.1-8B, respectively, and by $2.10$ points for shared-tokenizer transfer to Qwen3-1.7B.
Mohammad Siavashi, Pascal Schwinte, Ali Banaei et al. · cs.DC, cs.CL, cs.LG
Inference speed 92%Serving systems 55%
Abstract
Agentic workloads resend long contexts that barely change between calls. Prefix caching now skips most of their prefill, yet every request must still be tokenized in full before the cache can even be consulted. Tokenization, a CPU stage long treated as negligible, becomes the largest part of time-to-first-token (TTFT), and a frontend that tokenizes for a whole GPU fleet limits how many GPUs it can feed. We present GraphTok, a GPU tokenizer that runs the entire encoding pipeline, from pre-tokenization through byte-pair encoding (BPE) merging to token-ID assembly, as one replayable CUDA graph. GraphTok reproduces the reference token IDs bit-for-bit on GPT-2, Llama 3, Qwen 2.5, DeepSeek-V3, and Gemma 3, hands short inputs to a CPU engine where the GPU cannot win, and plugs into unmodified vLLM and NVIDIA Dynamo. GraphTok encodes one million GPT-2 tokens in 1.7-3.2 ms including data transfers, and on Llama 3 it is 4.7-6.0x faster than gigatoken, the strongest CPU tokenizer. On TokTier's own single-request benchmark, it is 2.6-3.9x faster than gigatoken and up to 18.1x faster than TokTier. In vLLM with a cached prefix, it lowers mean TTFT by 2.1-2.4x. The largest gain is capacity. Within a 2 s P99 target, one Dynamo frontend with GraphTok sustains 518 requests/s of 121k-token traffic, the busiest-hour load of 56 GPUs, against 296 with gigatoken and 18.5 with the default tokenizer. Across 121k to one million tokens, this is 1.75-2.5x the load gigatoken sustains, with a P99 up to 64x lower where gigatoken's tail breaks.
Anthony Kostal-Vazquez, Vassil Dimitrov, Branden Butler · cs.LG
Compression 90%Inference speed 75%
Abstract
In this article, we propose algebraic integer number formats (AINF) as a framework for highly configurable and accurate compressed computation. A key property of AINF is its straightforward reduction to purely integer arithmetic, enabling error-free multiply-accumulate operations. We develop a matrix multiplication engine that exploits this property, obtaining performance within a predictable factor of the most efficient numerical counterparts on modern hardware accelerators. Another key advantage of AIQ is the freedom to select algebraic integers tailored to the numerical characteristics of a given use case, offering optimization opportunities that are fundamentally unavailable in floating-point or fixed-point formats. Motivated by these properties, we introduce algebraic integer quantization (AIQ), a lossy compression format for large language model (LLM) inference, and provide results demonstrating that AIQ achieves competitive accuracy compared to the industry-standard MX format across a diverse set of models and benchmarks at equivalent bit-per-parameter budgets. The experiments presented in this article employ algebraic integers of degree two (quadratic integers); generalizations to higher-order and, in particular, multidimensional algebraic integers promise even greater computational advantages and remain a promising direction for future work.