We introduce ExTernD (Expanded-rank Ternary Decomposition), a post-training factorization of each LLM weight matrix $A \in \mathbb{R}^{m \times n}$ into $A \approx B \mathrm{diag}(D) C$ with ternary factors $B \in \{-1,0,+1\}^{m \times k}$, $C \in \{-1,0,+1\}^{k \times n}$ and a real scale vector $D \in \mathbb{R}^k$. The inner rank $k = μ\min(m,n)$ is deliberately expanded beyond full rank ($μ> 1$), so that components past full rank correct the quantization error of earlier ones.
arXiv:2607. 13511v1 Announce Type: cross Abstract: We introduce ExTernD (Expanded-rank Ternary Decomposition), a post-training factorization of each LLM weight matrix $A \in \mathbb{R}^{m \times n}$ into $A \approx B \mathrm{diag}(D) C$ with ternary factors $B \in \{-1,0,+1\}^{m \times k}$, $C \in \{-1,0,+1\}^{k \times n}$ and a real scale vector $D \in \mathbb{R}^k$.
ShamAN-Q is a sub‑1‑bit post‑training quantization technique that builds on NanoQuant by replacing its diagonal reconstruction geometry with a dense curvature metric inspired by the Shampoo optimizer. For each linear weight, it fits a Kronecker product to the empirical Fisher information matrix of a small calibration set via Kullback–Leibler minimization, yielding a Mahalanobis reconstruction loss. The method updates continuous ADMM steps to Sylvester equations while keeping the discrete projection and deployment format unchanged, and it redistributes uniform rank across layers, achieving lower perplexity on Qwen3‑Base at roughly 1 bpw and matching or improving zero‑shot accuracy on the Eleuther LM Evaluation Harness.
arXiv:2609.00224v1 Announce Type: cross Abstract: Weight-only post-training quantization (PTQ) can alleviate the computational burden of serving large language models (LLMs) at scale. However, existi...
arXiv:2605.11222v2 Announce Type: replace Abstract: Quantization is an effective strategy to reduce the storage and computation footprint of large language models (LLMs). Post-training quantization (...
The paper reports a large‑scale post‑training ternarisation of the Qwen3 language model, extending a conversion pipeline from the 4B to the 8B variant. Using KOTMS rotation, E2M‑ATQ adaptive ternarisation, and GPTQ‑style error compensation, the authors achieve a 1.361× perplexity ratio across three corpora and retain 78.5% of the FP16 accuracy on zero‑shot tasks, with the 8B model outperforming the 4B by 8.9 percentage points. The study also demonstrates lossless lattice‑aware packing, producing an 8.24 GiB checkpoint that preserves perplexity, and shows that direct packed execution can reach 15.52 tokens/s in 7.35 GiB, though packed GEMV remains slower than FP16 cuBLAS.
arXiv:2608.28150v2 Announce Type: replace Abstract: How much matrix rank is required to preserve every bounded value output of normalized softmax attention? We study the unrestricted maximum-row-\(\e...
The paper investigates whether sink-aware attention head selection remains valid after 4‑bit NF4 weight‑only post‑training quantization. Using Sink Topology Consistency metrics, it finds that global rank preservation stays high across Qwen2.5 and Llama‑3.2 models, yet top‑k head overlap drops to 61–79% and layer‑specific sink‑mass shifts can be substantial. The study also shows that cross‑domain calibration degrades more than within‑domain precision and that recalibration with a small number of samples can recover most of the stability, though full‑map stability may require updating more layers.
The paper reports a post‑training ternarization of the 4‑billion‑parameter Qwen model, achieving an effective 1.641‑bit representation for 81.62 % of its weights while keeping activations at 16‑bit precision. Accuracy drops from 64.5 % to 54.7 % across ten capability tests, with uneven degradation (e.g., BoolQ 84.6 % of teacher performance, ARC‑Challenge 43.8 %). After packing the ternary planes, the model size shrinks from 8.29 GiB to 3.96 GiB with negligible change in perplexity, though inference speed is not improved.
arXiv:2512. 00956v3 Announce Type: replace Abstract: Quantizing LLM weights and activations is a standard approach for efficient deployment, but a few extreme outliers can stretch the dynamic range and amplify low-bit quantization errors.
The paper investigates post‑training quantization of transformer attention blocks by optimizing a joint loss over the Q, K, V projections rather than individual weight matrices. Using this joint attention‑based objective (JAB), the authors achieve significant compression on Mistral‑7B, recovering 77‑90% of the performance gap at 3 bits, but the method fails when MLP layers are included. A role‑aware offset rule that ignores sensitivity estimates outperforms JAB on GPT‑2 and full Mistral‑7B, demonstrating that the matrix a weight belongs to is more critical than sensitivity metrics.
The paper introduces a ternary multiplicative adaptation technique that enables fine‑tuning of ternary transformers without dequantization. By representing discrete weight updates as a low‑rank Kronecker factorization of two small ternary matrices applied element‑wise, the method preserves the ternary domain and allows direct merging of adaptation weights. Experiments on six language and vision models, including ternarized LLaMA‑3 and ViT‑B/16, show that the approach recovers most of the performance lost to quantization and outperforms existing low‑bit and ternary baselines.
arXiv:2607. 04302v1 Announce Type: cross Abstract: We present HiFA4, a post-training operator-level design that executes both QK^T and PV in FlashAttention as 4-bit HIF4 Cube GEMMs for LLM inference on Ascend NPUs, while maintaining the online softmax state in FP16.