BellowsKV introduces a training‑free approach that maps the hidden states of multiple adjacent layers in a large language model to a single shared latent representation per token, allowing each layer to retrieve its own keys and values via a layer‑specific projection. The method retains the latent at full dimension and offers two compression strategies: BellowsKV‑Quant, which stores the latent at low precision, and BellowsKV‑Evict, which makes a shared eviction decision per block. Experiments on Llama‑3.1‑8B, Mistral‑7B, and Qwen2.5‑7B show that BellowsKV‑Quant preserves about 99% of full‑cache accuracy at 3.6× compression and 97‑98% at 13‑14×, outperforming existing cross‑layer methods that drop to at most 84% accuracy at 16× compression.
By Jiamu Zhang, Hanjie Chen
arXiv:2607. 14568v1 Announce Type: cross Abstract: A companion study ran a 35B mixture-of-experts model on a 2011 NVIDIA Tesla C2075 (Fermi, sm_20, 6GB) as a GPU-prefill/CPU-decode hybrid, because the 4-bit model did not fit in device memory (arXiv:2606.
By A. C. Opus, J. Q. Lu
DanLing NestedTensor is a PyTorch tensor abstraction that embeds multi‑ragged structure directly into the tensor, allowing packed values to carry partition information and logical dimension order. This design enables broadcasting, feature transformations, and reductions to automatically respect ragged axes while preserving the same representation through autograd and both eager and compiled execution. Benchmarks on an A100 show significant speedups—up to 3.39× over padding for BERT models and 2.40–4.32× for a Pairformer‑style workload—while dramatically reducing peak memory usage.
By Zhiyuan Chen
The paper introduces a system that enables first‑order LoRA training on phone NPUs by running all frozen‑weight multiplications on the NPU while keeping adapters and the optimizer on the CPU. It reuses weight tiles for input‑gradient computation, selects precision based on gradient sensitivity, and fuses remaining CPU operations to overlap with NPU execution. The approach is evaluated on Qualcomm Snapdragon devices with Qwen and Llama LLMs, demonstrating background training for nightly personalization and benchmark workloads.
By Xu Zheng, Wei Cheng, Zhuomin Chen, Dongkuan Xu, Mo Sha, Jingchao Ni, Dongsheng Luo
arXiv:2605. 15250v3 Announce Type: replace-cross Abstract: Multi-head Latent Attention (MLA), the attention used in DeepSeek-V2/V3, jointly compresses keys and values into a low-rank latent and matches the H100 roofline almost perfectly.
By Fanxu Meng
The paper investigates whether training Mixture-of-Experts (MoE) routers can improve memory‑bandwidth locality on consumer GPUs. Using a new zero‑surgery telemetry tool, the authors measure that a large Qwen3‑235B model is bottlenecked by disk‑based expert access, and that an LRU cache can serve a majority of requests. They pre‑register experiments training 137 M‑parameter MoE models with locality‑aware losses, finding that while cache misses can drop up to 60 % (99 % static‑pin hit rate), every configuration fails to meet a strict 1 % perplexity threshold, indicating a tight coupling between cache efficiency and model quality.
By Shriniwas Ramesh Suram
arXiv:2610.02598v1 Announce Type: new
Abstract: Deploying LLMs on consumer-grade GPUs with insufficient memory to hold their weights can result in prohibitively slow inference, because decoding repea...
By JuneHyung Kim, Sankeerth Durvasula, Nandita Vijaykumar
BlockLock is a method for reproducible large‑language‑model inference that preserves the output of a target request while allowing other requests to share batched prefill computation. It offers two modes: BlockLock‑Rows, which fixes row widths and keeps batched decode for all requests, and BlockLock‑Pack, which uses terminal‑anchored packing with an isolated dummy segment to isolate the target’s KV state for decode. Experiments on five dense models across multiple GPUs show that both modes maintain target output bitwise, reduce latency by over 50 % and increase throughput by more than 100 % compared to vLLM’s batch‑invariant approach, without requiring custom kernel implementations.
By Xiuyuan Qi, Pedro Reviriego, Javier Conde, Shanshan Liu
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.
By Anirudh Malik, M Sparsh Mehra, Poojith Devan
STAPLE is a method that integrates static sparse memory into pretrained language models by adding token‑indexed embeddings as a residual pathway while preserving the original feed‑forward network. On the OLMo‑2 1B and 7B models, STAPLE outperforms dense training and other static sparse baselines, improving benchmark scores by up to 4.27 points and achieving faster decoding with less GPU memory and energy usage. The approach demonstrates that large pretrained models can be expanded efficiently without costly sparse pretraining from scratch.
By Ranajoy Sadhukhan, Euijin Hong, Sheng Cao, Zechun Liu, Shuowei Jin, Sha Li, Changlong Yu, Xin Liu, Qin Lu, Zhenyu Shi, Priyanka Nigam, Beidi Chen
arXiv:2609. 16391v1 Announce Type: cross Abstract: Weight-only post-training quantization is the cheapest way to shrink a retrieval embedder, and the received advice for applying it -- protect the embedding table, allocate bits by module sensitivity, prefer a ranking-aware objective over weight reconstruction -- was carried into LLM quantization largely intact.
By Hyojung Han
The paper compares two strategies for handling memory limits in large language model (LLM) serving: tensor parallelism, which distributes weights and KV cache across multiple GPUs, and KV compression, which reduces cache size via quantisation and eviction on a single GPU. Using a cost‑normalised simulator calibrated on A100, A40, and H100 hardware, the authors find that across two models (Llama‑2 7B and 70B) and various GPU configurations, compression consistently outperforms tensor parallelism in cost per million tokens, offering 1.20× to 2.00× savings. The study identifies a model‑size threshold (~36B parameters on an 80 GB card) where compression dominates, while tensor parallelism becomes necessary only for larger models where weights alone exceed a single GPU’s capacity.
By Srikanta Datta Tumkur, Mehar Simhadri, Anshu Bansal, Jay Iyer, Sai Pavan Kumar, Sai Kapil Kumar, Ramesh Nampelly, Raj Dandekar